reinit

.gitattributes

@@ -0,0 +1,32 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.gif filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text

LICENSE

@@ -0,0 +1,162 @@
+## creative commons
+
+# Attribution-NonCommercial 4.0 International
+
+Creative Commons Corporation (“Creative Commons”) is not a law firm and does not provide legal services or legal advice. Distribution of Creative Commons public licenses does not create a lawyer-client or other relationship. Creative Commons makes its licenses and related information available on an “as-is” basis. Creative Commons gives no warranties regarding its licenses, any material licensed under their terms and conditions, or any related information. Creative Commons disclaims all liability for damages resulting from their use to the fullest extent possible.
+
+### Using Creative Commons Public Licenses
+
+Creative Commons public licenses provide a standard set of terms and conditions that creators and other rights holders may use to share original works of authorship and other material subject to copyright and certain other rights specified in the public license below. The following considerations are for informational purposes only, are not exhaustive, and do not form part of our licenses.
+
+* __Considerations for licensors:__ Our public licenses are intended for use by those authorized to give the public permission to use material in ways otherwise restricted by copyright and certain other rights. Our licenses are irrevocable. Licensors should read and understand the terms and conditions of the license they choose before applying it. Licensors should also secure all rights necessary before applying our licenses so that the public can reuse the material as expected. Licensors should clearly mark any material not subject to the license. This includes other CC-licensed material, or material used under an exception or limitation to copyright. [More considerations for licensors](http://wiki.creativecommons.org/Considerations_for_licensors_and_licensees#Considerations_for_licensors).
+
+* __Considerations for the public:__ By using one of our public licenses, a licensor grants the public permission to use the licensed material under specified terms and conditions. If the licensor’s permission is not necessary for any reason–for example, because of any applicable exception or limitation to copyright–then that use is not regulated by the license. Our licenses grant only permissions under copyright and certain other rights that a licensor has authority to grant. Use of the licensed material may still be restricted for other reasons, including because others have copyright or other rights in the material. A licensor may make special requests, such as asking that all changes be marked or described. Although not required by our licenses, you are encouraged to respect those requests where reasonable. [More considerations for the public](http://wiki.creativecommons.org/Considerations_for_licensors_and_licensees#Considerations_for_licensees).
+
+## Creative Commons Attribution-NonCommercial 4.0 International Public License
+
+By exercising the Licensed Rights (defined below), You accept and agree to be bound by the terms and conditions of this Creative Commons Attribution-NonCommercial 4.0 International Public License ("Public License"). To the extent this Public License may be interpreted as a contract, You are granted the Licensed Rights in consideration of Your acceptance of these terms and conditions, and the Licensor grants You such rights in consideration of benefits the Licensor receives from making the Licensed Material available under these terms and conditions.
+
+### Section 1 – Definitions.
+
+a. __Adapted Material__ means material subject to Copyright and Similar Rights that is derived from or based upon the Licensed Material and in which the Licensed Material is translated, altered, arranged, transformed, or otherwise modified in a manner requiring permission under the Copyright and Similar Rights held by the Licensor. For purposes of this Public License, where the Licensed Material is a musical work, performance, or sound recording, Adapted Material is always produced where the Licensed Material is synched in timed relation with a moving image.
+
+b. __Adapter's License__ means the license You apply to Your Copyright and Similar Rights in Your contributions to Adapted Material in accordance with the terms and conditions of this Public License.
+
+c. __Copyright and Similar Rights__ means copyright and/or similar rights closely related to copyright including, without limitation, performance, broadcast, sound recording, and Sui Generis Database Rights, without regard to how the rights are labeled or categorized. For purposes of this Public License, the rights specified in Section 2(b)(1)-(2) are not Copyright and Similar Rights.
+
+d. __Effective Technological Measures__ means those measures that, in the absence of proper authority, may not be circumvented under laws fulfilling obligations under Article 11 of the WIPO Copyright Treaty adopted on December 20, 1996, and/or similar international agreements.
+
+e. __Exceptions and Limitations__ means fair use, fair dealing, and/or any other exception or limitation to Copyright and Similar Rights that applies to Your use of the Licensed Material.
+
+f. __Licensed Material__ means the artistic or literary work, database, or other material to which the Licensor applied this Public License.
+
+g. __Licensed Rights__ means the rights granted to You subject to the terms and conditions of this Public License, which are limited to all Copyright and Similar Rights that apply to Your use of the Licensed Material and that the Licensor has authority to license.
+
+h. __Licensor__ means the individual(s) or entity(ies) granting rights under this Public License.
+
+i. __NonCommercial__ means not primarily intended for or directed towards commercial advantage or monetary compensation. For purposes of this Public License, the exchange of the Licensed Material for other material subject to Copyright and Similar Rights by digital file-sharing or similar means is NonCommercial provided there is no payment of monetary compensation in connection with the exchange.
+
+j. __Share__ means to provide material to the public by any means or process that requires permission under the Licensed Rights, such as reproduction, public display, public performance, distribution, dissemination, communication, or importation, and to make material available to the public including in ways that members of the public may access the material from a place and at a time individually chosen by them.
+
+k. __Sui Generis Database Rights__ means rights other than copyright resulting from Directive 96/9/EC of the European Parliament and of the Council of 11 March 1996 on the legal protection of databases, as amended and/or succeeded, as well as other essentially equivalent rights anywhere in the world.
+
+l. __You__ means the individual or entity exercising the Licensed Rights under this Public License. Your has a corresponding meaning.
+
+### Section 2 – Scope.
+
+a. ___License grant.___
+
+   1. Subject to the terms and conditions of this Public License, the Licensor hereby grants You a worldwide, royalty-free, non-sublicensable, non-exclusive, irrevocable license to exercise the Licensed Rights in the Licensed Material to:
+
+       A. reproduce and Share the Licensed Material, in whole or in part, for NonCommercial purposes only; and
+
+       B. produce, reproduce, and Share Adapted Material for NonCommercial purposes only.
+
+   2. __Exceptions and Limitations.__ For the avoidance of doubt, where Exceptions and Limitations apply to Your use, this Public License does not apply, and You do not need to comply with its terms and conditions.
+
+   3. __Term.__ The term of this Public License is specified in Section 6(a).
+
+   4. __Media and formats; technical modifications allowed.__ The Licensor authorizes You to exercise the Licensed Rights in all media and formats whether now known or hereafter created, and to make technical modifications necessary to do so. The Licensor waives and/or agrees not to assert any right or authority to forbid You from making technical modifications necessary to exercise the Licensed Rights, including technical modifications necessary to circumvent Effective Technological Measures. For purposes of this Public License, simply making modifications authorized by this Section 2(a)(4) never produces Adapted Material.
+
+    5. __Downstream recipients.__
+
+        A. __Offer from the Licensor – Licensed Material.__ Every recipient of the Licensed Material automatically receives an offer from the Licensor to exercise the Licensed Rights under the terms and conditions of this Public License.
+
+        B. __No downstream restrictions.__ You may not offer or impose any additional or different terms or conditions on, or apply any Effective Technological Measures to, the Licensed Material if doing so restricts exercise of the Licensed Rights by any recipient of the Licensed Material.
+
+    6. __No endorsement.__ Nothing in this Public License constitutes or may be construed as permission to assert or imply that You are, or that Your use of the Licensed Material is, connected with, or sponsored, endorsed, or granted official status by, the Licensor or others designated to receive attribution as provided in Section 3(a)(1)(A)(i).
+
+b. ___Other rights.___
+
+   1. Moral rights, such as the right of integrity, are not licensed under this Public License, nor are publicity, privacy, and/or other similar personality rights; however, to the extent possible, the Licensor waives and/or agrees not to assert any such rights held by the Licensor to the limited extent necessary to allow You to exercise the Licensed Rights, but not otherwise.
+
+   2. Patent and trademark rights are not licensed under this Public License.
+
+   3. To the extent possible, the Licensor waives any right to collect royalties from You for the exercise of the Licensed Rights, whether directly or through a collecting society under any voluntary or waivable statutory or compulsory licensing scheme. In all other cases the Licensor expressly reserves any right to collect such royalties, including when the Licensed Material is used other than for NonCommercial purposes.
+
+### Section 3 – License Conditions.
+
+Your exercise of the Licensed Rights is expressly made subject to the following conditions.
+
+a. ___Attribution.___
+
+   1. If You Share the Licensed Material (including in modified form), You must:
+
+       A. retain the following if it is supplied by the Licensor with the Licensed Material:
+
+         i. identification of the creator(s) of the Licensed Material and any others designated to receive attribution, in any reasonable manner requested by the Licensor (including by pseudonym if designated);
+
+         ii. a copyright notice;
+
+         iii. a notice that refers to this Public License;
+
+         iv. a notice that refers to the disclaimer of warranties;
+
+         v. a URI or hyperlink to the Licensed Material to the extent reasonably practicable;
+
+       B. indicate if You modified the Licensed Material and retain an indication of any previous modifications; and
+
+       C. indicate the Licensed Material is licensed under this Public License, and include the text of, or the URI or hyperlink to, this Public License.
+
+   2. You may satisfy the conditions in Section 3(a)(1) in any reasonable manner based on the medium, means, and context in which You Share the Licensed Material. For example, it may be reasonable to satisfy the conditions by providing a URI or hyperlink to a resource that includes the required information.
+
+   3. If requested by the Licensor, You must remove any of the information required by Section 3(a)(1)(A) to the extent reasonably practicable.
+
+   4. If You Share Adapted Material You produce, the Adapter's License You apply must not prevent recipients of the Adapted Material from complying with this Public License.
+
+### Section 4 – Sui Generis Database Rights.
+
+Where the Licensed Rights include Sui Generis Database Rights that apply to Your use of the Licensed Material:
+
+a. for the avoidance of doubt, Section 2(a)(1) grants You the right to extract, reuse, reproduce, and Share all or a substantial portion of the contents of the database for NonCommercial purposes only;
+
+b. if You include all or a substantial portion of the database contents in a database in which You have Sui Generis Database Rights, then the database in which You have Sui Generis Database Rights (but not its individual contents) is Adapted Material; and
+
+c. You must comply with the conditions in Section 3(a) if You Share all or a substantial portion of the contents of the database.
+
+For the avoidance of doubt, this Section 4 supplements and does not replace Your obligations under this Public License where the Licensed Rights include other Copyright and Similar Rights.
+
+### Section 5 – Disclaimer of Warranties and Limitation of Liability.
+
+a. __Unless otherwise separately undertaken by the Licensor, to the extent possible, the Licensor offers the Licensed Material as-is and as-available, and makes no representations or warranties of any kind concerning the Licensed Material, whether express, implied, statutory, or other. This includes, without limitation, warranties of title, merchantability, fitness for a particular purpose, non-infringement, absence of latent or other defects, accuracy, or the presence or absence of errors, whether or not known or discoverable. Where disclaimers of warranties are not allowed in full or in part, this disclaimer may not apply to You.__
+
+b. __To the extent possible, in no event will the Licensor be liable to You on any legal theory (including, without limitation, negligence) or otherwise for any direct, special, indirect, incidental, consequential, punitive, exemplary, or other losses, costs, expenses, or damages arising out of this Public License or use of the Licensed Material, even if the Licensor has been advised of the possibility of such losses, costs, expenses, or damages. Where a limitation of liability is not allowed in full or in part, this limitation may not apply to You.__
+
+c. The disclaimer of warranties and limitation of liability provided above shall be interpreted in a manner that, to the extent possible, most closely approximates an absolute disclaimer and waiver of all liability.
+
+### Section 6 – Term and Termination.
+
+a. This Public License applies for the term of the Copyright and Similar Rights licensed here. However, if You fail to comply with this Public License, then Your rights under this Public License terminate automatically.
+
+b. Where Your right to use the Licensed Material has terminated under Section 6(a), it reinstates:
+
+   1. automatically as of the date the violation is cured, provided it is cured within 30 days of Your discovery of the violation; or
+
+   2. upon express reinstatement by the Licensor.
+
+   For the avoidance of doubt, this Section 6(b) does not affect any right the Licensor may have to seek remedies for Your violations of this Public License.
+
+c. For the avoidance of doubt, the Licensor may also offer the Licensed Material under separate terms or conditions or stop distributing the Licensed Material at any time; however, doing so will not terminate this Public License.
+
+d. Sections 1, 5, 6, 7, and 8 survive termination of this Public License.
+
+### Section 7 – Other Terms and Conditions.
+
+a. The Licensor shall not be bound by any additional or different terms or conditions communicated by You unless expressly agreed.
+
+b. Any arrangements, understandings, or agreements regarding the Licensed Material not stated herein are separate from and independent of the terms and conditions of this Public License.
+
+### Section 8 – Interpretation.
+
+a. For the avoidance of doubt, this Public License does not, and shall not be interpreted to, reduce, limit, restrict, or impose conditions on any use of the Licensed Material that could lawfully be made without permission under this Public License.
+
+b. To the extent possible, if any provision of this Public License is deemed unenforceable, it shall be automatically reformed to the minimum extent necessary to make it enforceable. If the provision cannot be reformed, it shall be severed from this Public License without affecting the enforceability of the remaining terms and conditions.
+
+c. No term or condition of this Public License will be waived and no failure to comply consented to unless expressly agreed to by the Licensor.
+
+d. Nothing in this Public License constitutes or may be interpreted as a limitation upon, or waiver of, any privileges and immunities that apply to the Licensor or You, including from the legal processes of any jurisdiction or authority.
+
+> Creative Commons is not a party to its public licenses. Notwithstanding, Creative Commons may elect to apply one of its public licenses to material it publishes and in those instances will be considered the “Licensor.” Except for the limited purpose of indicating that material is shared under a Creative Commons public license or as otherwise permitted by the Creative Commons policies published at [creativecommons.org/policies](http://creativecommons.org/policies), Creative Commons does not authorize the use of the trademark “Creative Commons” or any other trademark or logo of Creative Commons without its prior written consent including, without limitation, in connection with any unauthorized modifications to any of its public licenses or any other arrangements, understandings, or agreements concerning use of licensed material. For the avoidance of doubt, this paragraph does not form part of the public licenses.
+>
+> Creative Commons may be contacted at creativecommons.org
+

README.md

@@ -0,0 +1,13 @@
+---
+title: Stable Diffusion Mat Outpainting Primer
+emoji: 🐢
+colorFrom: red
+colorTo: purple
+sdk: gradio
+sdk_version: 3.4
+app_file: app.py
+pinned: false
+license: cc-by-nc-4.0
+---
+
+Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

app.py

@@ -0,0 +1,327 @@
+# %%
+
+# Copyright (c) 2021, NVIDIA CORPORATION.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+from networks.mat import Generator
+import gradio as gr
+import gradio.components as gc
+import base64
+import glob
+import os
+import random
+import re
+from http import HTTPStatus
+from io import BytesIO
+from typing import Dict, List, NamedTuple, Optional, Tuple
+
+import click
+import cv2
+import numpy as np
+import PIL.Image
+import torch
+import torch.nn.functional as F
+from PIL import Image, ImageDraw, ImageOps
+from pydantic import BaseModel
+
+import dnnlib
+import legacy
+
+
+pyspng = None
+
+
+def num_range(s: str) -> List[int]:
+    '''Accept either a comma separated list of numbers 'a,b,c' or a range 'a-c' and return as a list of ints.'''
+
+    range_re = re.compile(r'^(\d+)-(\d+)$')
+    m = range_re.match(s)
+    if m:
+        return list(range(int(m.group(1)), int(m.group(2))+1))
+    vals = s.split(',')
+    return [int(x) for x in vals]
+
+
+def copy_params_and_buffers(src_module, dst_module, require_all=False):
+    assert isinstance(src_module, torch.nn.Module)
+    assert isinstance(dst_module, torch.nn.Module)
+    src_tensors = {name: tensor for name,
+                   tensor in named_params_and_buffers(src_module)}
+    for name, tensor in named_params_and_buffers(dst_module):
+        assert (name in src_tensors) or (not require_all)
+        if name in src_tensors:
+            tensor.copy_(src_tensors[name].detach()).requires_grad_(
+                tensor.requires_grad)
+
+
+def params_and_buffers(module):
+    assert isinstance(module, torch.nn.Module)
+    return list(module.parameters()) + list(module.buffers())
+
+
+def named_params_and_buffers(module):
+    assert isinstance(module, torch.nn.Module)
+    return list(module.named_parameters()) + list(module.named_buffers())
+
+
+class Inpainter:
+    def __init__(self,
+                 network_pkl,
+                 resolution=512,
+                 truncation_psi=1,
+                 noise_mode='const',
+                 sdevice='cpu'
+                 ):
+        self.resolution = resolution
+        self.truncation_psi = truncation_psi
+        self.noise_mode = noise_mode
+        print(f'Loading networks from: {network_pkl}')
+        self.device = torch.device(sdevice)
+        with dnnlib.util.open_url(network_pkl) as f:
+            G_saved = (
+                legacy.load_network_pkl(f)
+                ['G_ema']
+                .to(self.device)
+                .eval()
+                .requires_grad_(False))  # type: ignore
+        net_res = 512 if resolution > 512 else resolution
+        self.G = (
+            Generator(
+                z_dim=512,
+                c_dim=0,
+                w_dim=512,
+                img_resolution=net_res,
+                img_channels=3
+            )
+            .to(self.device)
+            .eval()
+            .requires_grad_(False)
+        )
+        copy_params_and_buffers(G_saved,  self.G, require_all=True)
+
+    def generate_images2(
+        self,
+        dpath: List[PIL.Image.Image],
+        mpath: List[Optional[PIL.Image.Image]],
+        seed: int = 42,
+    ):
+        """
+        Generate images using pretrained network pickle.
+        """
+        resolution = self.resolution
+        truncation_psi = self.truncation_psi
+        noise_mode = self.noise_mode
+        # seed = 240  # pick up a random number
+
+        def seed_all(seed):
+            random.seed(seed)
+            np.random.seed(seed)
+            torch.manual_seed(seed)
+            torch.cuda.manual_seed(seed)
+        if seed is not None:
+            seed_all(seed)
+
+        # no Labels.
+        label = torch.zeros([1,  self.G.c_dim], device=self.device)
+
+        def read_image(image):
+            image = np.array(image)
+            if image.ndim == 2:
+                image = image[:, :, np.newaxis]  # HW => HWC
+                image = np.repeat(image, 3, axis=2)
+            image = image.transpose(2, 0, 1)  # HWC => CHW
+            image = image[:3]
+            return image
+        if resolution != 512:
+            noise_mode = 'random'
+        results = []
+        with torch.no_grad():
+            for i, (ipath, m) in enumerate(zip(dpath, mpath)):
+                if seed is None:
+                    seed_all(i)
+
+                image = read_image(ipath)
+                image = (torch.from_numpy(image).float().to(
+                    self. device) / 127.5 - 1).unsqueeze(0)
+
+                mask = np.array(m).astype(np.float32) / 255.0
+                mask = torch.from_numpy(mask).float().to(
+                    self. device).unsqueeze(0).unsqueeze(0)
+
+                z = torch.from_numpy(np.random.randn(
+                    1,  self.G.z_dim)).to(self.device)
+                output = self.G(image, mask, z, label,
+                                truncation_psi=truncation_psi, noise_mode=noise_mode)
+                output = (output.permute(0, 2, 3, 1) * 127.5 +
+                          127.5).round().clamp(0, 255).to(torch.uint8)
+                output = output[0].cpu().numpy()
+                results.append(PIL.Image.fromarray(output, 'RGB'))
+
+        return results
+
+
+# if __name__ == "__main__":
+#     generate_images()  # pylint: disable=no-value-for-parameter
+
+# ----------------------------------------------------------------------------
+def mask_to_alpha(img, mask):
+    img = img.copy()
+    img.putalpha(mask)
+    return img
+
+
+def blend(src, target, mask):
+    mask = np.expand_dims(mask, axis=-1)
+    result = (1-mask) * src + mask * target
+    return Image.fromarray(result.astype(np.uint8))
+
+
+def pad(img, size=(128, 128), tosize=(512, 512), border=1):
+    if isinstance(size, float):
+        size = (int(img.size[0] * size), int(img.size[1] * size))
+    # remove border
+    w, h = tosize
+
+    new_img = Image.new('RGBA', (w, h))
+
+    rimg = img.resize(size, resample=Image.Resampling.NEAREST)
+    rimg = ImageOps.crop(rimg, border=border)
+    tw, th = size
+    tw, th = tw - border*2, th - border*2
+    tc = ((w-tw)//2, (h-th)//2)
+
+    new_img.paste(rimg, tc)
+    mask = Image.new('L', (w, h))
+    white = Image.new('L', (tw, th), 255)
+    mask.paste(white, tc)
+
+    if 'A' in rimg.getbands():
+        mask.paste(img.getchannel('A'), tc)
+    return new_img, mask
+
+
+def b64_to_img(b64):
+    return Image.open(BytesIO(base64.b64decode(b64)))
+
+
+def img_to_b64(img):
+    with BytesIO() as f:
+        img.save(f, format='PNG')
+        return base64.b64encode(f.getvalue()).decode('utf-8')
+
+
+class Predictor:
+    def __init__(self):
+        """Load the model into memory to make running multiple predictions efficient"""
+        self.models = {
+            "places2": Inpainter(
+                network_pkl='models/Places_512_FullData.pkl',
+                resolution=512,
+                truncation_psi=1.,
+                noise_mode='const',
+            ),
+            "places2+laion300k": Inpainter(
+                network_pkl='models/Places_512_FullData+LAION300k.pkl',
+                resolution=512,
+                truncation_psi=1.,
+                noise_mode='const',
+            ),
+        }
+
+    # The arguments and types the model takes as input
+
+    def predict(
+        self,
+        img: Image.Image,
+        tosize=(512, 512),
+        border=5,
+        seed=42,
+        size=0.5,
+        model='places2',
+    ) -> Image:
+        i, m = pad(
+            img,
+            size=size,  # (328, 328),
+            tosize=tosize,
+            border=border
+        )
+        """Run a single prediction on the model"""
+        imgs = self.models[model].generate_images2(
+            dpath=[i.resize((512, 512), resample=Image.Resampling.NEAREST)],
+            mpath=[m.resize((512, 512), resample=Image.Resampling.NEAREST)],
+            seed=seed,
+        )
+        img_op_raw = imgs[0].convert('RGBA')
+        img_op_raw = img_op_raw.resize(
+            tosize, resample=Image.Resampling.NEAREST)
+        inpainted = img_op_raw.copy()
+
+        # paste original image to remove inpainting/scaling artifacts
+        inpainted = blend(
+            i,
+            inpainted,
+            1-(np.array(m) / 255)
+        )
+        minpainted = mask_to_alpha(inpainted, m)
+        return minpainted, inpainted, ImageOps.invert(m)
+
+
+predictor = Predictor()
+
+# %%
+
+
+def _outpaint(img, tosize, border, seed, size, model):
+    img_op = predictor.predict(
+        img,
+        border=border,
+        seed=seed,
+        tosize=(tosize, tosize),
+        size=float(size),
+        model=model,
+    )
+    return img_op
+# %%
+
+
+searchimage = gc.Image(shape=(224, 224), label="image", type='pil')
+to_size = gc.Slider(1, 1920, 512, step=1, label='output size')
+border = gc.Slider(
+    1, 50, 0, step=1, label='border to crop from the image before outpainting')
+seed = gc.Slider(1, 65536, 10, step=1, label='seed')
+size = gc.Slider(0, 1, .5, step=0.01,
+                 label='scale of the image before outpainting')
+
+out = gc.Image(label="primed image with alpha channel", type='pil')
+outwithoutalpha = gc.Image(
+    label="primed image without alpha channel", type='pil')
+mask = gc.Image(label="outpainting mask", type='pil')
+
+model = gc.Dropdown(
+    choices=['places2', 'places2+laion300k'],
+    value='places2',
+    label='model',
+)
+
+
+maturl = 'https://github.com/fenglinglwb/MAT'
+gr.Interface(
+    _outpaint,
+    [searchimage, to_size, border, seed, size, model],
+    [out, outwithoutalpha, mask],
+    title=f"MAT Primer for Stable Diffusion\n\nbased on MAT: Mask-Aware Transformer for Large Hole Image Inpainting\n\n{maturl}",
+    description=f"""<html>
+    create an primer for use in stable diffusion outpainting<br>
+    example with strength 0.5
+    <img src='file/op.gif' />
+    </html>""",
+    analytics_enabled=False,
+    allow_flagging='never',
+
+
+).launch()

dataset_tool.py

@@ -0,0 +1,444 @@
+# Copyright (c) 2021, NVIDIA CORPORATION.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+import functools
+import io
+import json
+import os
+import pickle
+import sys
+import tarfile
+import gzip
+import zipfile
+from pathlib import Path
+from typing import Callable, Optional, Tuple, Union
+
+import click
+import numpy as np
+import PIL.Image
+from tqdm import tqdm
+
+#----------------------------------------------------------------------------
+
+def error(msg):
+    print('Error: ' + msg)
+    sys.exit(1)
+
+#----------------------------------------------------------------------------
+
+def maybe_min(a: int, b: Optional[int]) -> int:
+    if b is not None:
+        return min(a, b)
+    return a
+
+#----------------------------------------------------------------------------
+
+def file_ext(name: Union[str, Path]) -> str:
+    return str(name).split('.')[-1]
+
+#----------------------------------------------------------------------------
+
+def is_image_ext(fname: Union[str, Path]) -> bool:
+    ext = file_ext(fname).lower()
+    return f'.{ext}' in PIL.Image.EXTENSION # type: ignore
+
+#----------------------------------------------------------------------------
+
+def open_image_folder(source_dir, *, max_images: Optional[int]):
+    input_images = [str(f) for f in sorted(Path(source_dir).rglob('*')) if is_image_ext(f) and os.path.isfile(f)]
+
+    # Load labels.
+    labels = {}
+    meta_fname = os.path.join(source_dir, 'dataset.json')
+    if os.path.isfile(meta_fname):
+        with open(meta_fname, 'r') as file:
+            labels = json.load(file)['labels']
+            if labels is not None:
+                labels = { x[0]: x[1] for x in labels }
+            else:
+                labels = {}
+
+    max_idx = maybe_min(len(input_images), max_images)
+
+    def iterate_images():
+        for idx, fname in enumerate(input_images):
+            arch_fname = os.path.relpath(fname, source_dir)
+            arch_fname = arch_fname.replace('\\', '/')
+            img = np.array(PIL.Image.open(fname))
+            yield dict(img=img, label=labels.get(arch_fname))
+            if idx >= max_idx-1:
+                break
+    return max_idx, iterate_images()
+
+#----------------------------------------------------------------------------
+
+def open_image_zip(source, *, max_images: Optional[int]):
+    with zipfile.ZipFile(source, mode='r') as z:
+        input_images = [str(f) for f in sorted(z.namelist()) if is_image_ext(f)]
+
+        # Load labels.
+        labels = {}
+        if 'dataset.json' in z.namelist():
+            with z.open('dataset.json', 'r') as file:
+                labels = json.load(file)['labels']
+                if labels is not None:
+                    labels = { x[0]: x[1] for x in labels }
+                else:
+                    labels = {}
+
+    max_idx = maybe_min(len(input_images), max_images)
+
+    def iterate_images():
+        with zipfile.ZipFile(source, mode='r') as z:
+            for idx, fname in enumerate(input_images):
+                with z.open(fname, 'r') as file:
+                    img = PIL.Image.open(file) # type: ignore
+                    img = np.array(img)
+                yield dict(img=img, label=labels.get(fname))
+                if idx >= max_idx-1:
+                    break
+    return max_idx, iterate_images()
+
+#----------------------------------------------------------------------------
+
+def open_lmdb(lmdb_dir: str, *, max_images: Optional[int]):
+    import cv2  # pip install opencv-python
+    import lmdb  # pip install lmdb # pylint: disable=import-error
+
+    with lmdb.open(lmdb_dir, readonly=True, lock=False).begin(write=False) as txn:
+        max_idx = maybe_min(txn.stat()['entries'], max_images)
+
+    def iterate_images():
+        with lmdb.open(lmdb_dir, readonly=True, lock=False).begin(write=False) as txn:
+            for idx, (_key, value) in enumerate(txn.cursor()):
+                try:
+                    try:
+                        img = cv2.imdecode(np.frombuffer(value, dtype=np.uint8), 1)
+                        if img is None:
+                            raise IOError('cv2.imdecode failed')
+                        img = img[:, :, ::-1] # BGR => RGB
+                    except IOError:
+                        img = np.array(PIL.Image.open(io.BytesIO(value)))
+                    yield dict(img=img, label=None)
+                    if idx >= max_idx-1:
+                        break
+                except:
+                    print(sys.exc_info()[1])
+
+    return max_idx, iterate_images()
+
+#----------------------------------------------------------------------------
+
+def open_cifar10(tarball: str, *, max_images: Optional[int]):
+    images = []
+    labels = []
+
+    with tarfile.open(tarball, 'r:gz') as tar:
+        for batch in range(1, 6):
+            member = tar.getmember(f'cifar-10-batches-py/data_batch_{batch}')
+            with tar.extractfile(member) as file:
+                data = pickle.load(file, encoding='latin1')
+            images.append(data['data'].reshape(-1, 3, 32, 32))
+            labels.append(data['labels'])
+
+    images = np.concatenate(images)
+    labels = np.concatenate(labels)
+    images = images.transpose([0, 2, 3, 1]) # NCHW -> NHWC
+    assert images.shape == (50000, 32, 32, 3) and images.dtype == np.uint8
+    assert labels.shape == (50000,) and labels.dtype in [np.int32, np.int64]
+    assert np.min(images) == 0 and np.max(images) == 255
+    assert np.min(labels) == 0 and np.max(labels) == 9
+
+    max_idx = maybe_min(len(images), max_images)
+
+    def iterate_images():
+        for idx, img in enumerate(images):
+            yield dict(img=img, label=int(labels[idx]))
+            if idx >= max_idx-1:
+                break
+
+    return max_idx, iterate_images()
+
+#----------------------------------------------------------------------------
+
+def open_mnist(images_gz: str, *, max_images: Optional[int]):
+    labels_gz = images_gz.replace('-images-idx3-ubyte.gz', '-labels-idx1-ubyte.gz')
+    assert labels_gz != images_gz
+    images = []
+    labels = []
+
+    with gzip.open(images_gz, 'rb') as f:
+        images = np.frombuffer(f.read(), np.uint8, offset=16)
+    with gzip.open(labels_gz, 'rb') as f:
+        labels = np.frombuffer(f.read(), np.uint8, offset=8)
+
+    images = images.reshape(-1, 28, 28)
+    images = np.pad(images, [(0,0), (2,2), (2,2)], 'constant', constant_values=0)
+    assert images.shape == (60000, 32, 32) and images.dtype == np.uint8
+    assert labels.shape == (60000,) and labels.dtype == np.uint8
+    assert np.min(images) == 0 and np.max(images) == 255
+    assert np.min(labels) == 0 and np.max(labels) == 9
+
+    max_idx = maybe_min(len(images), max_images)
+
+    def iterate_images():
+        for idx, img in enumerate(images):
+            yield dict(img=img, label=int(labels[idx]))
+            if idx >= max_idx-1:
+                break
+
+    return max_idx, iterate_images()
+
+#----------------------------------------------------------------------------
+
+def make_transform(
+    transform: Optional[str],
+    output_width: Optional[int],
+    output_height: Optional[int],
+    resize_filter: str
+) -> Callable[[np.ndarray], Optional[np.ndarray]]:
+    resample = { 'box': PIL.Image.BOX, 'lanczos': PIL.Image.LANCZOS }[resize_filter]
+    def scale(width, height, img):
+        w = img.shape[1]
+        h = img.shape[0]
+        if width == w and height == h:
+            return img
+        img = PIL.Image.fromarray(img)
+        ww = width if width is not None else w
+        hh = height if height is not None else h
+        img = img.resize((ww, hh), resample)
+        return np.array(img)
+
+    def center_crop(width, height, img):
+        crop = np.min(img.shape[:2])
+        img = img[(img.shape[0] - crop) // 2 : (img.shape[0] + crop) // 2, (img.shape[1] - crop) // 2 : (img.shape[1] + crop) // 2]
+        img = PIL.Image.fromarray(img, 'RGB')
+        img = img.resize((width, height), resample)
+        return np.array(img)
+
+    def center_crop_wide(width, height, img):
+        ch = int(np.round(width * img.shape[0] / img.shape[1]))
+        if img.shape[1] < width or ch < height:
+            return None
+
+        img = img[(img.shape[0] - ch) // 2 : (img.shape[0] + ch) // 2]
+        img = PIL.Image.fromarray(img, 'RGB')
+        img = img.resize((width, height), resample)
+        img = np.array(img)
+
+        canvas = np.zeros([width, width, 3], dtype=np.uint8)
+        canvas[(width - height) // 2 : (width + height) // 2, :] = img
+        return canvas
+
+    if transform is None:
+        return functools.partial(scale, output_width, output_height)
+    if transform == 'center-crop':
+        if (output_width is None) or (output_height is None):
+            error ('must specify --width and --height when using ' + transform + 'transform')
+        return functools.partial(center_crop, output_width, output_height)
+    if transform == 'center-crop-wide':
+        if (output_width is None) or (output_height is None):
+            error ('must specify --width and --height when using ' + transform + ' transform')
+        return functools.partial(center_crop_wide, output_width, output_height)
+    assert False, 'unknown transform'
+
+#----------------------------------------------------------------------------
+
+def open_dataset(source, *, max_images: Optional[int]):
+    if os.path.isdir(source):
+        if source.rstrip('/').endswith('_lmdb'):
+            return open_lmdb(source, max_images=max_images)
+        else:
+            return open_image_folder(source, max_images=max_images)
+    elif os.path.isfile(source):
+        if os.path.basename(source) == 'cifar-10-python.tar.gz':
+            return open_cifar10(source, max_images=max_images)
+        elif os.path.basename(source) == 'train-images-idx3-ubyte.gz':
+            return open_mnist(source, max_images=max_images)
+        elif file_ext(source) == 'zip':
+            return open_image_zip(source, max_images=max_images)
+        else:
+            assert False, 'unknown archive type'
+    else:
+        error(f'Missing input file or directory: {source}')
+
+#----------------------------------------------------------------------------
+
+def open_dest(dest: str) -> Tuple[str, Callable[[str, Union[bytes, str]], None], Callable[[], None]]:
+    dest_ext = file_ext(dest)
+
+    if dest_ext == 'zip':
+        if os.path.dirname(dest) != '':
+            os.makedirs(os.path.dirname(dest), exist_ok=True)
+        zf = zipfile.ZipFile(file=dest, mode='w', compression=zipfile.ZIP_STORED)
+        def zip_write_bytes(fname: str, data: Union[bytes, str]):
+            zf.writestr(fname, data)
+        return '', zip_write_bytes, zf.close
+    else:
+        # If the output folder already exists, check that is is
+        # empty.
+        #
+        # Note: creating the output directory is not strictly
+        # necessary as folder_write_bytes() also mkdirs, but it's better
+        # to give an error message earlier in case the dest folder
+        # somehow cannot be created.
+        if os.path.isdir(dest) and len(os.listdir(dest)) != 0:
+            error('--dest folder must be empty')
+        os.makedirs(dest, exist_ok=True)
+
+        def folder_write_bytes(fname: str, data: Union[bytes, str]):
+            os.makedirs(os.path.dirname(fname), exist_ok=True)
+            with open(fname, 'wb') as fout:
+                if isinstance(data, str):
+                    data = data.encode('utf8')
+                fout.write(data)
+        return dest, folder_write_bytes, lambda: None
+
+#----------------------------------------------------------------------------
+
+@click.command()
+@click.pass_context
+@click.option('--source', help='Directory or archive name for input dataset', required=True, metavar='PATH')
+@click.option('--dest', help='Output directory or archive name for output dataset', required=True, metavar='PATH')
+@click.option('--max-images', help='Output only up to `max-images` images', type=int, default=None)
+@click.option('--resize-filter', help='Filter to use when resizing images for output resolution', type=click.Choice(['box', 'lanczos']), default='lanczos', show_default=True)
+@click.option('--transform', help='Input crop/resize mode', type=click.Choice(['center-crop', 'center-crop-wide']))
+@click.option('--width', help='Output width', type=int)
+@click.option('--height', help='Output height', type=int)
+def convert_dataset(
+    ctx: click.Context,
+    source: str,
+    dest: str,
+    max_images: Optional[int],
+    transform: Optional[str],
+    resize_filter: str,
+    width: Optional[int],
+    height: Optional[int]
+):
+    """Convert an image dataset into a dataset archive usable with StyleGAN2 ADA PyTorch.
+
+    The input dataset format is guessed from the --source argument:
+
+    \b
+    --source *_lmdb/                    Load LSUN dataset
+    --source cifar-10-python.tar.gz     Load CIFAR-10 dataset
+    --source train-images-idx3-ubyte.gz Load MNIST dataset
+    --source path/                      Recursively load all images from path/
+    --source dataset.zip                Recursively load all images from dataset.zip
+
+    Specifying the output format and path:
+
+    \b
+    --dest /path/to/dir                 Save output files under /path/to/dir
+    --dest /path/to/dataset.zip         Save output files into /path/to/dataset.zip
+
+    The output dataset format can be either an image folder or an uncompressed zip archive.
+    Zip archives makes it easier to move datasets around file servers and clusters, and may
+    offer better training performance on network file systems.
+
+    Images within the dataset archive will be stored as uncompressed PNG.
+    Uncompresed PNGs can be efficiently decoded in the training loop.
+
+    Class labels are stored in a file called 'dataset.json' that is stored at the
+    dataset root folder.  This file has the following structure:
+
+    \b
+    {
+        "labels": [
+            ["00000/img00000000.png",6],
+            ["00000/img00000001.png",9],
+            ... repeated for every image in the datase
+            ["00049/img00049999.png",1]
+        ]
+    }
+
+    If the 'dataset.json' file cannot be found, the dataset is interpreted as
+    not containing class labels.
+
+    Image scale/crop and resolution requirements:
+
+    Output images must be square-shaped and they must all have the same power-of-two
+    dimensions.
+
+    To scale arbitrary input image size to a specific width and height, use the
+    --width and --height options.  Output resolution will be either the original
+    input resolution (if --width/--height was not specified) or the one specified with
+    --width/height.
+
+    Use the --transform=center-crop or --transform=center-crop-wide options to apply a
+    center crop transform on the input image.  These options should be used with the
+    --width and --height options.  For example:
+
+    \b
+    python dataset_tool.py --source LSUN/raw/cat_lmdb --dest /tmp/lsun_cat \\
+        --transform=center-crop-wide --width 512 --height=384
+    """
+
+    PIL.Image.init() # type: ignore
+
+    if dest == '':
+        ctx.fail('--dest output filename or directory must not be an empty string')
+
+    num_files, input_iter = open_dataset(source, max_images=max_images)
+    archive_root_dir, save_bytes, close_dest = open_dest(dest)
+
+    transform_image = make_transform(transform, width, height, resize_filter)
+
+    dataset_attrs = None
+
+    labels = []
+    for idx, image in tqdm(enumerate(input_iter), total=num_files):
+        idx_str = f'{idx:08d}'
+        archive_fname = f'{idx_str[:5]}/img{idx_str}.png'
+
+        # Apply crop and resize.
+        img = transform_image(image['img'])
+
+        # Transform may drop images.
+        if img is None:
+            continue
+
+        # Error check to require uniform image attributes across
+        # the whole dataset.
+        channels = img.shape[2] if img.ndim == 3 else 1
+        cur_image_attrs = {
+            'width': img.shape[1],
+            'height': img.shape[0],
+            'channels': channels
+        }
+        if dataset_attrs is None:
+            dataset_attrs = cur_image_attrs
+            width = dataset_attrs['width']
+            height = dataset_attrs['height']
+            if width != height:
+                error(f'Image dimensions after scale and crop are required to be square.  Got {width}x{height}')
+            if dataset_attrs['channels'] not in [1, 3]:
+                error('Input images must be stored as RGB or grayscale')
+            if width != 2 ** int(np.floor(np.log2(width))):
+                error('Image width/height after scale and crop are required to be power-of-two')
+        elif dataset_attrs != cur_image_attrs:
+            err = [f'  dataset {k}/cur image {k}: {dataset_attrs[k]}/{cur_image_attrs[k]}' for k in dataset_attrs.keys()]
+            error(f'Image {archive_fname} attributes must be equal across all images of the dataset.  Got:\n' + '\n'.join(err))
+
+        # Save the image as an uncompressed PNG.
+        img = PIL.Image.fromarray(img, { 1: 'L', 3: 'RGB' }[channels])
+        image_bits = io.BytesIO()
+        img.save(image_bits, format='png', compress_level=0, optimize=False)
+        save_bytes(os.path.join(archive_root_dir, archive_fname), image_bits.getbuffer())
+        labels.append([archive_fname, image['label']] if image['label'] is not None else None)
+
+    metadata = {
+        'labels': labels if all(x is not None for x in labels) else None
+    }
+    save_bytes(os.path.join(archive_root_dir, 'dataset.json'), json.dumps(metadata))
+    close_dest()
+
+#----------------------------------------------------------------------------
+
+if __name__ == "__main__":
+    convert_dataset() # pylint: disable=no-value-for-parameter

datasets/dataset_256.py

@@ -0,0 +1,286 @@
+# Copyright (c) 2021, NVIDIA CORPORATION.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+import cv2
+import os
+import numpy as np
+import zipfile
+import PIL.Image
+import json
+import torch
+import dnnlib
+import random
+
+try:
+    import pyspng
+except ImportError:
+    pyspng = None
+
+from datasets.mask_generator_256 import RandomMask
+
+#----------------------------------------------------------------------------
+
+class Dataset(torch.utils.data.Dataset):
+    def __init__(self,
+        name,                   # Name of the dataset.
+        raw_shape,              # Shape of the raw image data (NCHW).
+        max_size    = None,     # Artificially limit the size of the dataset. None = no limit. Applied before xflip.
+        use_labels  = False,    # Enable conditioning labels? False = label dimension is zero.
+        xflip       = False,    # Artificially double the size of the dataset via x-flips. Applied after max_size.
+        random_seed = 0,        # Random seed to use when applying max_size.
+    ):
+        self._name = name
+        self._raw_shape = list(raw_shape)
+        self._use_labels = use_labels
+        self._raw_labels = None
+        self._label_shape = None
+
+        # Apply max_size.
+        self._raw_idx = np.arange(self._raw_shape[0], dtype=np.int64)
+        if (max_size is not None) and (self._raw_idx.size > max_size):
+            np.random.RandomState(random_seed).shuffle(self._raw_idx)
+            self._raw_idx = np.sort(self._raw_idx[:max_size])
+
+        # Apply xflip.
+        self._xflip = np.zeros(self._raw_idx.size, dtype=np.uint8)
+        if xflip:
+            self._raw_idx = np.tile(self._raw_idx, 2)
+            self._xflip = np.concatenate([self._xflip, np.ones_like(self._xflip)])
+
+    def _get_raw_labels(self):
+        if self._raw_labels is None:
+            self._raw_labels = self._load_raw_labels() if self._use_labels else None
+            if self._raw_labels is None:
+                self._raw_labels = np.zeros([self._raw_shape[0], 0], dtype=np.float32)
+            assert isinstance(self._raw_labels, np.ndarray)
+            assert self._raw_labels.shape[0] == self._raw_shape[0]
+            assert self._raw_labels.dtype in [np.float32, np.int64]
+            if self._raw_labels.dtype == np.int64:
+                assert self._raw_labels.ndim == 1
+                assert np.all(self._raw_labels >= 0)
+        return self._raw_labels
+
+    def close(self): # to be overridden by subclass
+        pass
+
+    def _load_raw_image(self, raw_idx): # to be overridden by subclass
+        raise NotImplementedError
+
+    def _load_raw_labels(self): # to be overridden by subclass
+        raise NotImplementedError
+
+    def __getstate__(self):
+        return dict(self.__dict__, _raw_labels=None)
+
+    def __del__(self):
+        try:
+            self.close()
+        except:
+            pass
+
+    def __len__(self):
+        return self._raw_idx.size
+
+    def __getitem__(self, idx):
+        image = self._load_raw_image(self._raw_idx[idx])
+        assert isinstance(image, np.ndarray)
+        assert list(image.shape) == self.image_shape
+        assert image.dtype == np.uint8
+        if self._xflip[idx]:
+            assert image.ndim == 3 # CHW
+            image = image[:, :, ::-1]
+        return image.copy(), self.get_label(idx)
+
+    def get_label(self, idx):
+        label = self._get_raw_labels()[self._raw_idx[idx]]
+        if label.dtype == np.int64:
+            onehot = np.zeros(self.label_shape, dtype=np.float32)
+            onehot[label] = 1
+            label = onehot
+        return label.copy()
+
+    def get_details(self, idx):
+        d = dnnlib.EasyDict()
+        d.raw_idx = int(self._raw_idx[idx])
+        d.xflip = (int(self._xflip[idx]) != 0)
+        d.raw_label = self._get_raw_labels()[d.raw_idx].copy()
+        return d
+
+    @property
+    def name(self):
+        return self._name
+
+    @property
+    def image_shape(self):
+        return list(self._raw_shape[1:])
+
+    @property
+    def num_channels(self):
+        assert len(self.image_shape) == 3 # CHW
+        return self.image_shape[0]
+
+    @property
+    def resolution(self):
+        assert len(self.image_shape) == 3 # CHW
+        assert self.image_shape[1] == self.image_shape[2]
+        return self.image_shape[1]
+
+    @property
+    def label_shape(self):
+        if self._label_shape is None:
+            raw_labels = self._get_raw_labels()
+            if raw_labels.dtype == np.int64:
+                self._label_shape = [int(np.max(raw_labels)) + 1]
+            else:
+                self._label_shape = raw_labels.shape[1:]
+        return list(self._label_shape)
+
+    @property
+    def label_dim(self):
+        assert len(self.label_shape) == 1
+        return self.label_shape[0]
+
+    @property
+    def has_labels(self):
+        return any(x != 0 for x in self.label_shape)
+
+    @property
+    def has_onehot_labels(self):
+        return self._get_raw_labels().dtype == np.int64
+
+
+#----------------------------------------------------------------------------
+
+
+class ImageFolderMaskDataset(Dataset):
+    def __init__(self,
+        path,                   # Path to directory or zip.
+        resolution      = None, # Ensure specific resolution, None = highest available.
+        hole_range=[0,1],
+        **super_kwargs,         # Additional arguments for the Dataset base class.
+    ):
+        self._path = path
+        self._zipfile = None
+        self._hole_range = hole_range
+
+        if os.path.isdir(self._path):
+            self._type = 'dir'
+            self._all_fnames = {os.path.relpath(os.path.join(root, fname), start=self._path) for root, _dirs, files in os.walk(self._path) for fname in files}
+        elif self._file_ext(self._path) == '.zip':
+            self._type = 'zip'
+            self._all_fnames = set(self._get_zipfile().namelist())
+        else:
+            raise IOError('Path must point to a directory or zip')
+
+        PIL.Image.init()
+        self._image_fnames = sorted(fname for fname in self._all_fnames if self._file_ext(fname) in PIL.Image.EXTENSION)
+        if len(self._image_fnames) == 0:
+            raise IOError('No image files found in the specified path')
+
+        name = os.path.splitext(os.path.basename(self._path))[0]
+        raw_shape = [len(self._image_fnames)] + list(self._load_raw_image(0).shape)
+        if resolution is not None and (raw_shape[2] != resolution or raw_shape[3] != resolution):
+            raise IOError('Image files do not match the specified resolution')
+        super().__init__(name=name, raw_shape=raw_shape, **super_kwargs)
+
+    @staticmethod
+    def _file_ext(fname):
+        return os.path.splitext(fname)[1].lower()
+
+    def _get_zipfile(self):
+        assert self._type == 'zip'
+        if self._zipfile is None:
+            self._zipfile = zipfile.ZipFile(self._path)
+        return self._zipfile
+
+    def _open_file(self, fname):
+        if self._type == 'dir':
+            return open(os.path.join(self._path, fname), 'rb')
+        if self._type == 'zip':
+            return self._get_zipfile().open(fname, 'r')
+        return None
+
+    def close(self):
+        try:
+            if self._zipfile is not None:
+                self._zipfile.close()
+        finally:
+            self._zipfile = None
+
+    def __getstate__(self):
+        return dict(super().__getstate__(), _zipfile=None)
+
+    def _load_raw_image(self, raw_idx):
+        fname = self._image_fnames[raw_idx]
+        with self._open_file(fname) as f:
+            if pyspng is not None and self._file_ext(fname) == '.png':
+                image = pyspng.load(f.read())
+            else:
+                image = np.array(PIL.Image.open(f))
+        if image.ndim == 2:
+            image = image[:, :, np.newaxis] # HW => HWC
+
+        # for grayscale image
+        if image.shape[2] == 1:
+            image = np.repeat(image, 3, axis=2)
+
+        # restricted to 256x256
+        res = 256
+        H, W, C = image.shape
+        if H < res or W < res:
+            top = 0
+            bottom = max(0, res - H)
+            left = 0
+            right = max(0, res - W)
+            image = cv2.copyMakeBorder(image, top, bottom, left, right, cv2.BORDER_REFLECT)
+        H, W, C = image.shape
+        h = random.randint(0, H - res)
+        w = random.randint(0, W - res)
+        image = image[h:h+res, w:w+res, :]
+
+        image = np.ascontiguousarray(image.transpose(2, 0, 1)) # HWC => CHW
+
+        return image
+
+    def _load_raw_labels(self):
+        fname = 'labels.json'
+        if fname not in self._all_fnames:
+            return None
+        with self._open_file(fname) as f:
+            labels = json.load(f)['labels']
+        if labels is None:
+            return None
+        labels = dict(labels)
+        labels = [labels[fname.replace('\\', '/')] for fname in self._image_fnames]
+        labels = np.array(labels)
+        labels = labels.astype({1: np.int64, 2: np.float32}[labels.ndim])
+        return labels
+
+    def __getitem__(self, idx):
+        image = self._load_raw_image(self._raw_idx[idx])
+
+        assert isinstance(image, np.ndarray)
+        assert list(image.shape) == self.image_shape
+        assert image.dtype == np.uint8
+        if self._xflip[idx]:
+            assert image.ndim == 3 # CHW
+            image = image[:, :, ::-1]
+        mask = RandomMask(image.shape[-1], hole_range=self._hole_range)  # hole as 0, reserved as 1
+        return image.copy(), mask, self.get_label(idx)
+
+
+if __name__ == '__main__':
+    res = 256
+    dpath = '/data/liwenbo/datasets/Places365/standard/val_256'
+    D = ImageFolderMaskDataset(path=dpath)
+    print(D.__len__())
+    for i in range(D.__len__()):
+        print(i)
+        a, b, c = D.__getitem__(i)
+        if a.shape != (3, 256, 256):
+            print(i, a.shape)

datasets/dataset_256_val.py

@@ -0,0 +1,282 @@
+# Copyright (c) 2021, NVIDIA CORPORATION.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+import os
+import numpy as np
+import zipfile
+import PIL.Image
+import cv2
+import json
+import torch
+import dnnlib
+import glob
+
+try:
+    import pyspng
+except ImportError:
+    pyspng = None
+
+from datasets.mask_generator_256 import RandomMask
+
+#----------------------------------------------------------------------------
+
+class Dataset(torch.utils.data.Dataset):
+    def __init__(self,
+        name,                   # Name of the dataset.
+        raw_shape,              # Shape of the raw image data (NCHW).
+        max_size    = None,     # Artificially limit the size of the dataset. None = no limit. Applied before xflip.
+        use_labels  = False,    # Enable conditioning labels? False = label dimension is zero.
+        xflip       = False,    # Artificially double the size of the dataset via x-flips. Applied after max_size.
+        random_seed = 0,        # Random seed to use when applying max_size.
+    ):
+        self._name = name
+        self._raw_shape = list(raw_shape)
+        self._use_labels = use_labels
+        self._raw_labels = None
+        self._label_shape = None
+
+        # Apply max_size.
+        self._raw_idx = np.arange(self._raw_shape[0], dtype=np.int64)
+        if (max_size is not None) and (self._raw_idx.size > max_size):
+            np.random.RandomState(random_seed).shuffle(self._raw_idx)
+            self._raw_idx = np.sort(self._raw_idx[:max_size])
+
+        # Apply xflip.
+        self._xflip = np.zeros(self._raw_idx.size, dtype=np.uint8)
+        if xflip:
+            self._raw_idx = np.tile(self._raw_idx, 2)
+            self._xflip = np.concatenate([self._xflip, np.ones_like(self._xflip)])
+
+    def _get_raw_labels(self):
+        if self._raw_labels is None:
+            self._raw_labels = self._load_raw_labels() if self._use_labels else None
+            if self._raw_labels is None:
+                self._raw_labels = np.zeros([self._raw_shape[0], 0], dtype=np.float32)
+            assert isinstance(self._raw_labels, np.ndarray)
+            assert self._raw_labels.shape[0] == self._raw_shape[0]
+            assert self._raw_labels.dtype in [np.float32, np.int64]
+            if self._raw_labels.dtype == np.int64:
+                assert self._raw_labels.ndim == 1
+                assert np.all(self._raw_labels >= 0)
+        return self._raw_labels
+
+    def close(self): # to be overridden by subclass
+        pass
+
+    def _load_raw_image(self, raw_idx): # to be overridden by subclass
+        raise NotImplementedError
+
+    def _load_raw_labels(self): # to be overridden by subclass
+        raise NotImplementedError
+
+    def __getstate__(self):
+        return dict(self.__dict__, _raw_labels=None)
+
+    def __del__(self):
+        try:
+            self.close()
+        except:
+            pass
+
+    def __len__(self):
+        return self._raw_idx.size
+
+    def __getitem__(self, idx):
+        image = self._load_raw_image(self._raw_idx[idx])
+        assert isinstance(image, np.ndarray)
+        assert list(image.shape) == self.image_shape
+        assert image.dtype == np.uint8
+        if self._xflip[idx]:
+            assert image.ndim == 3 # CHW
+            image = image[:, :, ::-1]
+        return image.copy(), self.get_label(idx)
+
+    def get_label(self, idx):
+        label = self._get_raw_labels()[self._raw_idx[idx]]
+        if label.dtype == np.int64:
+            onehot = np.zeros(self.label_shape, dtype=np.float32)
+            onehot[label] = 1
+            label = onehot
+        return label.copy()
+
+    def get_details(self, idx):
+        d = dnnlib.EasyDict()
+        d.raw_idx = int(self._raw_idx[idx])
+        d.xflip = (int(self._xflip[idx]) != 0)
+        d.raw_label = self._get_raw_labels()[d.raw_idx].copy()
+        return d
+
+    @property
+    def name(self):
+        return self._name
+
+    @property
+    def image_shape(self):
+        return list(self._raw_shape[1:])
+
+    @property
+    def num_channels(self):
+        assert len(self.image_shape) == 3 # CHW
+        return self.image_shape[0]
+
+    @property
+    def resolution(self):
+        assert len(self.image_shape) == 3 # CHW
+        assert self.image_shape[1] == self.image_shape[2]
+        return self.image_shape[1]
+
+    @property
+    def label_shape(self):
+        if self._label_shape is None:
+            raw_labels = self._get_raw_labels()
+            if raw_labels.dtype == np.int64:
+                self._label_shape = [int(np.max(raw_labels)) + 1]
+            else:
+                self._label_shape = raw_labels.shape[1:]
+        return list(self._label_shape)
+
+    @property
+    def label_dim(self):
+        assert len(self.label_shape) == 1
+        return self.label_shape[0]
+
+    @property
+    def has_labels(self):
+        return any(x != 0 for x in self.label_shape)
+
+    @property
+    def has_onehot_labels(self):
+        return self._get_raw_labels().dtype == np.int64
+
+
+#----------------------------------------------------------------------------
+
+
+class ImageFolderMaskDataset(Dataset):
+    def __init__(self,
+        path,                   # Path to directory or zip.
+        resolution      = None, # Ensure specific resolution, None = highest available.
+        hole_range=[0,1],
+        **super_kwargs,         # Additional arguments for the Dataset base class.
+    ):
+        self._path = path
+        self._zipfile = None
+        self._hole_range = hole_range
+
+        if os.path.isdir(self._path):
+            self._type = 'dir'
+            self._all_fnames = {os.path.relpath(os.path.join(root, fname), start=self._path) for root, _dirs, files in os.walk(self._path) for fname in files}
+        elif self._file_ext(self._path) == '.zip':
+            self._type = 'zip'
+            self._all_fnames = set(self._get_zipfile().namelist())
+        else:
+            raise IOError('Path must point to a directory or zip')
+
+        PIL.Image.init()
+        self._image_fnames = sorted(fname for fname in self._all_fnames if self._file_ext(fname) in PIL.Image.EXTENSION)
+        if len(self._image_fnames) == 0:
+            raise IOError('No image files found in the specified path')
+
+        name = os.path.splitext(os.path.basename(self._path))[0]
+        raw_shape = [len(self._image_fnames)] + list(self._load_raw_image(0).shape)
+        if resolution is not None and (raw_shape[2] != resolution or raw_shape[3] != resolution):
+            raise IOError('Image files do not match the specified resolution')
+        self._load_mask()
+        super().__init__(name=name, raw_shape=raw_shape, **super_kwargs)
+
+    def _load_mask(self, mpath='/data/liwenbo/datasets/Places365/standard/masks_val_256_eval'):
+        self.masks = sorted(glob.glob(mpath + '/*.png'))
+
+    @staticmethod
+    def _file_ext(fname):
+        return os.path.splitext(fname)[1].lower()
+
+    def _get_zipfile(self):
+        assert self._type == 'zip'
+        if self._zipfile is None:
+            self._zipfile = zipfile.ZipFile(self._path)
+        return self._zipfile
+
+    def _open_file(self, fname):
+        if self._type == 'dir':
+            return open(os.path.join(self._path, fname), 'rb')
+        if self._type == 'zip':
+            return self._get_zipfile().open(fname, 'r')
+        return None
+
+    def close(self):
+        try:
+            if self._zipfile is not None:
+                self._zipfile.close()
+        finally:
+            self._zipfile = None
+
+    def __getstate__(self):
+        return dict(super().__getstate__(), _zipfile=None)
+
+    def _load_raw_image(self, raw_idx):
+        fname = self._image_fnames[raw_idx]
+        with self._open_file(fname) as f:
+            if pyspng is not None and self._file_ext(fname) == '.png':
+                image = pyspng.load(f.read())
+            else:
+                image = np.array(PIL.Image.open(f))
+        if image.ndim == 2:
+            image = image[:, :, np.newaxis] # HW => HWC
+
+        # for grayscale image
+        if image.shape[2] == 1:
+            image = np.repeat(image, 3, axis=2)
+
+        # restricted to 256x256
+        res = 256
+        H, W, C = image.shape
+        if H < res or W < res:
+            top = 0
+            bottom = max(0, res - H)
+            left = 0
+            right = max(0, res - W)
+            image = cv2.copyMakeBorder(image, top, bottom, left, right, cv2.BORDER_REFLECT)
+        H, W, C = image.shape
+        h = (H - res) // 2
+        w = (W - res) // 2
+        image = image[h:h+res, w:w+res, :]
+
+        image = np.ascontiguousarray(image.transpose(2, 0, 1)) # HWC => CHW
+        return image
+
+    def _load_raw_labels(self):
+        fname = 'labels.json'
+        if fname not in self._all_fnames:
+            return None
+        with self._open_file(fname) as f:
+            labels = json.load(f)['labels']
+        if labels is None:
+            return None
+        labels = dict(labels)
+        labels = [labels[fname.replace('\\', '/')] for fname in self._image_fnames]
+        labels = np.array(labels)
+        labels = labels.astype({1: np.int64, 2: np.float32}[labels.ndim])
+        return labels
+
+    def __getitem__(self, idx):
+        image = self._load_raw_image(self._raw_idx[idx])
+
+        # for grayscale image
+        if image.shape[0] == 1:
+            image = np.repeat(image, 3, axis=0)
+
+        assert isinstance(image, np.ndarray)
+        assert list(image.shape) == self.image_shape
+        assert image.dtype == np.uint8
+        if self._xflip[idx]:
+            assert image.ndim == 3 # CHW
+            image = image[:, :, ::-1]
+        # mask = RandomMask(image.shape[-1], hole_range=self._hole_range)  # hole as 0, reserved as 1
+        mask = cv2.imread(self.masks[idx], cv2.IMREAD_GRAYSCALE).astype(np.float32)[np.newaxis, :, :] / 255.0
+        return image.copy(), mask, self.get_label(idx)

datasets/dataset_512.py

@@ -0,0 +1,286 @@
+# Copyright (c) 2021, NVIDIA CORPORATION.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+import cv2
+import os
+import numpy as np
+import zipfile
+import PIL.Image
+import json
+import torch
+import dnnlib
+import random
+
+try:
+    import pyspng
+except ImportError:
+    pyspng = None
+
+from datasets.mask_generator_512 import RandomMask
+
+#----------------------------------------------------------------------------
+
+class Dataset(torch.utils.data.Dataset):
+    def __init__(self,
+        name,                   # Name of the dataset.
+        raw_shape,              # Shape of the raw image data (NCHW).
+        max_size    = None,     # Artificially limit the size of the dataset. None = no limit. Applied before xflip.
+        use_labels  = False,    # Enable conditioning labels? False = label dimension is zero.
+        xflip       = False,    # Artificially double the size of the dataset via x-flips. Applied after max_size.
+        random_seed = 0,        # Random seed to use when applying max_size.
+    ):
+        self._name = name
+        self._raw_shape = list(raw_shape)
+        self._use_labels = use_labels
+        self._raw_labels = None
+        self._label_shape = None
+
+        # Apply max_size.
+        self._raw_idx = np.arange(self._raw_shape[0], dtype=np.int64)
+        if (max_size is not None) and (self._raw_idx.size > max_size):
+            np.random.RandomState(random_seed).shuffle(self._raw_idx)
+            self._raw_idx = np.sort(self._raw_idx[:max_size])
+
+        # Apply xflip.
+        self._xflip = np.zeros(self._raw_idx.size, dtype=np.uint8)
+        if xflip:
+            self._raw_idx = np.tile(self._raw_idx, 2)
+            self._xflip = np.concatenate([self._xflip, np.ones_like(self._xflip)])
+
+    def _get_raw_labels(self):
+        if self._raw_labels is None:
+            self._raw_labels = self._load_raw_labels() if self._use_labels else None
+            if self._raw_labels is None:
+                self._raw_labels = np.zeros([self._raw_shape[0], 0], dtype=np.float32)
+            assert isinstance(self._raw_labels, np.ndarray)
+            assert self._raw_labels.shape[0] == self._raw_shape[0]
+            assert self._raw_labels.dtype in [np.float32, np.int64]
+            if self._raw_labels.dtype == np.int64:
+                assert self._raw_labels.ndim == 1
+                assert np.all(self._raw_labels >= 0)
+        return self._raw_labels
+
+    def close(self): # to be overridden by subclass
+        pass
+
+    def _load_raw_image(self, raw_idx): # to be overridden by subclass
+        raise NotImplementedError
+
+    def _load_raw_labels(self): # to be overridden by subclass
+        raise NotImplementedError
+
+    def __getstate__(self):
+        return dict(self.__dict__, _raw_labels=None)
+
+    def __del__(self):
+        try:
+            self.close()
+        except:
+            pass
+
+    def __len__(self):
+        return self._raw_idx.size
+
+    def __getitem__(self, idx):
+        image = self._load_raw_image(self._raw_idx[idx])
+        assert isinstance(image, np.ndarray)
+        assert list(image.shape) == self.image_shape
+        assert image.dtype == np.uint8
+        if self._xflip[idx]:
+            assert image.ndim == 3 # CHW
+            image = image[:, :, ::-1]
+        return image.copy(), self.get_label(idx)
+
+    def get_label(self, idx):
+        label = self._get_raw_labels()[self._raw_idx[idx]]
+        if label.dtype == np.int64:
+            onehot = np.zeros(self.label_shape, dtype=np.float32)
+            onehot[label] = 1
+            label = onehot
+        return label.copy()
+
+    def get_details(self, idx):
+        d = dnnlib.EasyDict()
+        d.raw_idx = int(self._raw_idx[idx])
+        d.xflip = (int(self._xflip[idx]) != 0)
+        d.raw_label = self._get_raw_labels()[d.raw_idx].copy()
+        return d
+
+    @property
+    def name(self):
+        return self._name
+
+    @property
+    def image_shape(self):
+        return list(self._raw_shape[1:])
+
+    @property
+    def num_channels(self):
+        assert len(self.image_shape) == 3 # CHW
+        return self.image_shape[0]
+
+    @property
+    def resolution(self):
+        assert len(self.image_shape) == 3 # CHW
+        assert self.image_shape[1] == self.image_shape[2]
+        return self.image_shape[1]
+
+    @property
+    def label_shape(self):
+        if self._label_shape is None:
+            raw_labels = self._get_raw_labels()
+            if raw_labels.dtype == np.int64:
+                self._label_shape = [int(np.max(raw_labels)) + 1]
+            else:
+                self._label_shape = raw_labels.shape[1:]
+        return list(self._label_shape)
+
+    @property
+    def label_dim(self):
+        assert len(self.label_shape) == 1
+        return self.label_shape[0]
+
+    @property
+    def has_labels(self):
+        return any(x != 0 for x in self.label_shape)
+
+    @property
+    def has_onehot_labels(self):
+        return self._get_raw_labels().dtype == np.int64
+
+
+#----------------------------------------------------------------------------
+
+
+class ImageFolderMaskDataset(Dataset):
+    def __init__(self,
+        path,                   # Path to directory or zip.
+        resolution      = None, # Ensure specific resolution, None = highest available.
+        hole_range=[0,1],
+        **super_kwargs,         # Additional arguments for the Dataset base class.
+    ):
+        self._path = path
+        self._zipfile = None
+        self._hole_range = hole_range
+
+        if os.path.isdir(self._path):
+            self._type = 'dir'
+            self._all_fnames = {os.path.relpath(os.path.join(root, fname), start=self._path) for root, _dirs, files in os.walk(self._path) for fname in files}
+        elif self._file_ext(self._path) == '.zip':
+            self._type = 'zip'
+            self._all_fnames = set(self._get_zipfile().namelist())
+        else:
+            raise IOError('Path must point to a directory or zip')
+
+        PIL.Image.init()
+        self._image_fnames = sorted(fname for fname in self._all_fnames if self._file_ext(fname) in PIL.Image.EXTENSION)
+        if len(self._image_fnames) == 0:
+            raise IOError('No image files found in the specified path')
+
+        name = os.path.splitext(os.path.basename(self._path))[0]
+        raw_shape = [len(self._image_fnames)] + list(self._load_raw_image(0).shape)
+        if resolution is not None and (raw_shape[2] != resolution or raw_shape[3] != resolution):
+            raise IOError('Image files do not match the specified resolution')
+        super().__init__(name=name, raw_shape=raw_shape, **super_kwargs)
+
+    @staticmethod
+    def _file_ext(fname):
+        return os.path.splitext(fname)[1].lower()
+
+    def _get_zipfile(self):
+        assert self._type == 'zip'
+        if self._zipfile is None:
+            self._zipfile = zipfile.ZipFile(self._path)
+        return self._zipfile
+
+    def _open_file(self, fname):
+        if self._type == 'dir':
+            return open(os.path.join(self._path, fname), 'rb')
+        if self._type == 'zip':
+            return self._get_zipfile().open(fname, 'r')
+        return None
+
+    def close(self):
+        try:
+            if self._zipfile is not None:
+                self._zipfile.close()
+        finally:
+            self._zipfile = None
+
+    def __getstate__(self):
+        return dict(super().__getstate__(), _zipfile=None)
+
+    def _load_raw_image(self, raw_idx):
+        fname = self._image_fnames[raw_idx]
+        with self._open_file(fname) as f:
+            if pyspng is not None and self._file_ext(fname) == '.png':
+                image = pyspng.load(f.read())
+            else:
+                image = np.array(PIL.Image.open(f))
+        if image.ndim == 2:
+            image = image[:, :, np.newaxis] # HW => HWC
+
+        # for grayscale image
+        if image.shape[2] == 1:
+            image = np.repeat(image, 3, axis=2)
+
+        # restricted to 512x512
+        res = 512
+        H, W, C = image.shape
+        if H < res or W < res:
+            top = 0
+            bottom = max(0, res - H)
+            left = 0
+            right = max(0, res - W)
+            image = cv2.copyMakeBorder(image, top, bottom, left, right, cv2.BORDER_REFLECT)
+        H, W, C = image.shape
+        h = random.randint(0, H - res)
+        w = random.randint(0, W - res)
+        image = image[h:h+res, w:w+res, :]
+
+        image = np.ascontiguousarray(image.transpose(2, 0, 1)) # HWC => CHW
+
+        return image
+
+    def _load_raw_labels(self):
+        fname = 'labels.json'
+        if fname not in self._all_fnames:
+            return None
+        with self._open_file(fname) as f:
+            labels = json.load(f)['labels']
+        if labels is None:
+            return None
+        labels = dict(labels)
+        labels = [labels[fname.replace('\\', '/')] for fname in self._image_fnames]
+        labels = np.array(labels)
+        labels = labels.astype({1: np.int64, 2: np.float32}[labels.ndim])
+        return labels
+
+    def __getitem__(self, idx):
+        image = self._load_raw_image(self._raw_idx[idx])
+
+        assert isinstance(image, np.ndarray)
+        assert list(image.shape) == self.image_shape
+        assert image.dtype == np.uint8
+        if self._xflip[idx]:
+            assert image.ndim == 3 # CHW
+            image = image[:, :, ::-1]
+        mask = RandomMask(image.shape[-1], hole_range=self._hole_range)  # hole as 0, reserved as 1
+        return image.copy(), mask, self.get_label(idx)
+
+
+if __name__ == '__main__':
+    res = 512
+    dpath = '/data/liwenbo/datasets/Places365/standard/val_large'
+    D = ImageFolderMaskDataset(path=dpath)
+    print(D.__len__())
+    for i in range(D.__len__()):
+        print(i)
+        a, b, c = D.__getitem__(i)
+        if a.shape != (3, 512, 512):
+            print(i, a.shape)

datasets/dataset_512_val.py

@@ -0,0 +1,282 @@
+# Copyright (c) 2021, NVIDIA CORPORATION.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+import os
+import numpy as np
+import zipfile
+import PIL.Image
+import cv2
+import json
+import torch
+import dnnlib
+import glob
+
+try:
+    import pyspng
+except ImportError:
+    pyspng = None
+
+from datasets.mask_generator_512 import RandomMask
+
+#----------------------------------------------------------------------------
+
+class Dataset(torch.utils.data.Dataset):
+    def __init__(self,
+        name,                   # Name of the dataset.
+        raw_shape,              # Shape of the raw image data (NCHW).
+        max_size    = None,     # Artificially limit the size of the dataset. None = no limit. Applied before xflip.
+        use_labels  = False,    # Enable conditioning labels? False = label dimension is zero.
+        xflip       = False,    # Artificially double the size of the dataset via x-flips. Applied after max_size.
+        random_seed = 0,        # Random seed to use when applying max_size.
+    ):
+        self._name = name
+        self._raw_shape = list(raw_shape)
+        self._use_labels = use_labels
+        self._raw_labels = None
+        self._label_shape = None
+
+        # Apply max_size.
+        self._raw_idx = np.arange(self._raw_shape[0], dtype=np.int64)
+        if (max_size is not None) and (self._raw_idx.size > max_size):
+            np.random.RandomState(random_seed).shuffle(self._raw_idx)
+            self._raw_idx = np.sort(self._raw_idx[:max_size])
+
+        # Apply xflip.
+        self._xflip = np.zeros(self._raw_idx.size, dtype=np.uint8)
+        if xflip:
+            self._raw_idx = np.tile(self._raw_idx, 2)
+            self._xflip = np.concatenate([self._xflip, np.ones_like(self._xflip)])
+
+    def _get_raw_labels(self):
+        if self._raw_labels is None:
+            self._raw_labels = self._load_raw_labels() if self._use_labels else None
+            if self._raw_labels is None:
+                self._raw_labels = np.zeros([self._raw_shape[0], 0], dtype=np.float32)
+            assert isinstance(self._raw_labels, np.ndarray)
+            assert self._raw_labels.shape[0] == self._raw_shape[0]
+            assert self._raw_labels.dtype in [np.float32, np.int64]
+            if self._raw_labels.dtype == np.int64:
+                assert self._raw_labels.ndim == 1
+                assert np.all(self._raw_labels >= 0)
+        return self._raw_labels
+
+    def close(self): # to be overridden by subclass
+        pass
+
+    def _load_raw_image(self, raw_idx): # to be overridden by subclass
+        raise NotImplementedError
+
+    def _load_raw_labels(self): # to be overridden by subclass
+        raise NotImplementedError
+
+    def __getstate__(self):
+        return dict(self.__dict__, _raw_labels=None)
+
+    def __del__(self):
+        try:
+            self.close()
+        except:
+            pass
+
+    def __len__(self):
+        return self._raw_idx.size
+
+    def __getitem__(self, idx):
+        image = self._load_raw_image(self._raw_idx[idx])
+        assert isinstance(image, np.ndarray)
+        assert list(image.shape) == self.image_shape
+        assert image.dtype == np.uint8
+        if self._xflip[idx]:
+            assert image.ndim == 3 # CHW
+            image = image[:, :, ::-1]
+        return image.copy(), self.get_label(idx)
+
+    def get_label(self, idx):
+        label = self._get_raw_labels()[self._raw_idx[idx]]
+        if label.dtype == np.int64:
+            onehot = np.zeros(self.label_shape, dtype=np.float32)
+            onehot[label] = 1
+            label = onehot
+        return label.copy()
+
+    def get_details(self, idx):
+        d = dnnlib.EasyDict()
+        d.raw_idx = int(self._raw_idx[idx])
+        d.xflip = (int(self._xflip[idx]) != 0)
+        d.raw_label = self._get_raw_labels()[d.raw_idx].copy()
+        return d
+
+    @property
+    def name(self):
+        return self._name
+
+    @property
+    def image_shape(self):
+        return list(self._raw_shape[1:])
+
+    @property
+    def num_channels(self):
+        assert len(self.image_shape) == 3 # CHW
+        return self.image_shape[0]
+
+    @property
+    def resolution(self):
+        assert len(self.image_shape) == 3 # CHW
+        assert self.image_shape[1] == self.image_shape[2]
+        return self.image_shape[1]
+
+    @property
+    def label_shape(self):
+        if self._label_shape is None:
+            raw_labels = self._get_raw_labels()
+            if raw_labels.dtype == np.int64:
+                self._label_shape = [int(np.max(raw_labels)) + 1]
+            else:
+                self._label_shape = raw_labels.shape[1:]
+        return list(self._label_shape)
+
+    @property
+    def label_dim(self):
+        assert len(self.label_shape) == 1
+        return self.label_shape[0]
+
+    @property
+    def has_labels(self):
+        return any(x != 0 for x in self.label_shape)
+
+    @property
+    def has_onehot_labels(self):
+        return self._get_raw_labels().dtype == np.int64
+
+
+#----------------------------------------------------------------------------
+
+
+class ImageFolderMaskDataset(Dataset):
+    def __init__(self,
+        path,                   # Path to directory or zip.
+        resolution      = None, # Ensure specific resolution, None = highest available.
+        hole_range=[0,1],
+        **super_kwargs,         # Additional arguments for the Dataset base class.
+    ):
+        self._path = path
+        self._zipfile = None
+        self._hole_range = hole_range
+
+        if os.path.isdir(self._path):
+            self._type = 'dir'
+            self._all_fnames = {os.path.relpath(os.path.join(root, fname), start=self._path) for root, _dirs, files in os.walk(self._path) for fname in files}
+        elif self._file_ext(self._path) == '.zip':
+            self._type = 'zip'
+            self._all_fnames = set(self._get_zipfile().namelist())
+        else:
+            raise IOError('Path must point to a directory or zip')
+
+        PIL.Image.init()
+        self._image_fnames = sorted(fname for fname in self._all_fnames if self._file_ext(fname) in PIL.Image.EXTENSION)
+        if len(self._image_fnames) == 0:
+            raise IOError('No image files found in the specified path')
+
+        name = os.path.splitext(os.path.basename(self._path))[0]
+        raw_shape = [len(self._image_fnames)] + list(self._load_raw_image(0).shape)
+        if resolution is not None and (raw_shape[2] != resolution or raw_shape[3] != resolution):
+            raise IOError('Image files do not match the specified resolution')
+        self._load_mask()
+        super().__init__(name=name, raw_shape=raw_shape, **super_kwargs)
+
+    def _load_mask(self, mpath='/data/liwenbo/datasets/Places365/standard/masks_val_512_eval'):
+        self.masks = sorted(glob.glob(mpath + '/*.png'))
+
+    @staticmethod
+    def _file_ext(fname):
+        return os.path.splitext(fname)[1].lower()
+
+    def _get_zipfile(self):
+        assert self._type == 'zip'
+        if self._zipfile is None:
+            self._zipfile = zipfile.ZipFile(self._path)
+        return self._zipfile
+
+    def _open_file(self, fname):
+        if self._type == 'dir':
+            return open(os.path.join(self._path, fname), 'rb')
+        if self._type == 'zip':
+            return self._get_zipfile().open(fname, 'r')
+        return None
+
+    def close(self):
+        try:
+            if self._zipfile is not None:
+                self._zipfile.close()
+        finally:
+            self._zipfile = None
+
+    def __getstate__(self):
+        return dict(super().__getstate__(), _zipfile=None)
+
+    def _load_raw_image(self, raw_idx):
+        fname = self._image_fnames[raw_idx]
+        with self._open_file(fname) as f:
+            if pyspng is not None and self._file_ext(fname) == '.png':
+                image = pyspng.load(f.read())
+            else:
+                image = np.array(PIL.Image.open(f))
+        if image.ndim == 2:
+            image = image[:, :, np.newaxis] # HW => HWC
+
+        # for grayscale image
+        if image.shape[2] == 1:
+            image = np.repeat(image, 3, axis=2)
+
+        # restricted to 512x512
+        res = 512
+        H, W, C = image.shape
+        if H < res or W < res:
+            top = 0
+            bottom = max(0, res - H)
+            left = 0
+            right = max(0, res - W)
+            image = cv2.copyMakeBorder(image, top, bottom, left, right, cv2.BORDER_REFLECT)
+        H, W, C = image.shape
+        h = (H - res) // 2
+        w = (W - res) // 2
+        image = image[h:h+res, w:w+res, :]
+
+        image = np.ascontiguousarray(image.transpose(2, 0, 1)) # HWC => CHW
+        return image
+
+    def _load_raw_labels(self):
+        fname = 'labels.json'
+        if fname not in self._all_fnames:
+            return None
+        with self._open_file(fname) as f:
+            labels = json.load(f)['labels']
+        if labels is None:
+            return None
+        labels = dict(labels)
+        labels = [labels[fname.replace('\\', '/')] for fname in self._image_fnames]
+        labels = np.array(labels)
+        labels = labels.astype({1: np.int64, 2: np.float32}[labels.ndim])
+        return labels
+
+    def __getitem__(self, idx):
+        image = self._load_raw_image(self._raw_idx[idx])
+
+        # for grayscale image
+        if image.shape[0] == 1:
+            image = np.repeat(image, 3, axis=0)
+
+        assert isinstance(image, np.ndarray)
+        assert list(image.shape) == self.image_shape
+        assert image.dtype == np.uint8
+        if self._xflip[idx]:
+            assert image.ndim == 3 # CHW
+            image = image[:, :, ::-1]
+        # mask = RandomMask(image.shape[-1], hole_range=self._hole_range)  # hole as 0, reserved as 1
+        mask = cv2.imread(self.masks[idx], cv2.IMREAD_GRAYSCALE).astype(np.float32)[np.newaxis, :, :] / 255.0
+        return image.copy(), mask, self.get_label(idx)

datasets/mask_generator_256.py

@@ -0,0 +1,93 @@
+import numpy as np
+from PIL import Image, ImageDraw
+import math
+import random
+
+
+def RandomBrush(
+    max_tries,
+    s,
+    min_num_vertex = 4,
+    max_num_vertex = 18,
+    mean_angle = 2*math.pi / 5,
+    angle_range = 2*math.pi / 15,
+    min_width = 12,
+    max_width = 48):
+    H, W = s, s
+    average_radius = math.sqrt(H*H+W*W) / 8
+    mask = Image.new('L', (W, H), 0)
+    for _ in range(np.random.randint(max_tries)):
+        num_vertex = np.random.randint(min_num_vertex, max_num_vertex)
+        angle_min = mean_angle - np.random.uniform(0, angle_range)
+        angle_max = mean_angle + np.random.uniform(0, angle_range)
+        angles = []
+        vertex = []
+        for i in range(num_vertex):
+            if i % 2 == 0:
+                angles.append(2*math.pi - np.random.uniform(angle_min, angle_max))
+            else:
+                angles.append(np.random.uniform(angle_min, angle_max))
+
+        h, w = mask.size
+        vertex.append((int(np.random.randint(0, w)), int(np.random.randint(0, h))))
+        for i in range(num_vertex):
+            r = np.clip(
+                np.random.normal(loc=average_radius, scale=average_radius//2),
+                0, 2*average_radius)
+            new_x = np.clip(vertex[-1][0] + r * math.cos(angles[i]), 0, w)
+            new_y = np.clip(vertex[-1][1] + r * math.sin(angles[i]), 0, h)
+            vertex.append((int(new_x), int(new_y)))
+
+        draw = ImageDraw.Draw(mask)
+        width = int(np.random.uniform(min_width, max_width))
+        draw.line(vertex, fill=1, width=width)
+        for v in vertex:
+            draw.ellipse((v[0] - width//2,
+                          v[1] - width//2,
+                          v[0] + width//2,
+                          v[1] + width//2),
+                         fill=1)
+        if np.random.random() > 0.5:
+            mask.transpose(Image.FLIP_LEFT_RIGHT)
+        if np.random.random() > 0.5:
+            mask.transpose(Image.FLIP_TOP_BOTTOM)
+    mask = np.asarray(mask, np.uint8)
+    if np.random.random() > 0.5:
+        mask = np.flip(mask, 0)
+    if np.random.random() > 0.5:
+        mask = np.flip(mask, 1)
+    return mask
+
+def RandomMask(s, hole_range=[0,1]):
+    coef = min(hole_range[0] + hole_range[1], 1.0)
+    while True:
+        mask = np.ones((s, s), np.uint8)
+        def Fill(max_size):
+            w, h = np.random.randint(max_size), np.random.randint(max_size)
+            ww, hh = w // 2, h // 2
+            x, y = np.random.randint(-ww, s - w + ww), np.random.randint(-hh, s - h + hh)
+            mask[max(y, 0): min(y + h, s), max(x, 0): min(x + w, s)] = 0
+        def MultiFill(max_tries, max_size):
+            for _ in range(np.random.randint(max_tries)):
+                Fill(max_size)
+        MultiFill(int(4 * coef), s // 2)
+        MultiFill(int(2 * coef), s)
+        mask = np.logical_and(mask, 1 - RandomBrush(int(8 * coef), s))  # hole denoted as 0, reserved as 1
+        hole_ratio = 1 - np.mean(mask)
+        if hole_range is not None and (hole_ratio <= hole_range[0] or hole_ratio >= hole_range[1]):
+            continue
+        return mask[np.newaxis, ...].astype(np.float32)
+
+def BatchRandomMask(batch_size, s, hole_range=[0, 1]):
+    return np.stack([RandomMask(s, hole_range=hole_range) for _ in range(batch_size)], axis=0)
+
+
+if __name__ == '__main__':
+    # res = 512
+    res = 256
+    cnt = 2000
+    tot = 0
+    for i in range(cnt):
+        mask = RandomMask(s=res)
+        tot += mask.mean()
+    print(tot / cnt)

datasets/mask_generator_256_small.py

@@ -0,0 +1,93 @@
+import numpy as np
+from PIL import Image, ImageDraw
+import math
+import random
+
+
+def RandomBrush(
+    max_tries,
+    s,
+    min_num_vertex = 4,
+    max_num_vertex = 18,
+    mean_angle = 2*math.pi / 5,
+    angle_range = 2*math.pi / 15,
+    min_width = 12,
+    max_width = 48):
+    H, W = s, s
+    average_radius = math.sqrt(H*H+W*W) / 8
+    mask = Image.new('L', (W, H), 0)
+    for _ in range(np.random.randint(max_tries)):
+        num_vertex = np.random.randint(min_num_vertex, max_num_vertex)
+        angle_min = mean_angle - np.random.uniform(0, angle_range)
+        angle_max = mean_angle + np.random.uniform(0, angle_range)
+        angles = []
+        vertex = []
+        for i in range(num_vertex):
+            if i % 2 == 0:
+                angles.append(2*math.pi - np.random.uniform(angle_min, angle_max))
+            else:
+                angles.append(np.random.uniform(angle_min, angle_max))
+
+        h, w = mask.size
+        vertex.append((int(np.random.randint(0, w)), int(np.random.randint(0, h))))
+        for i in range(num_vertex):
+            r = np.clip(
+                np.random.normal(loc=average_radius, scale=average_radius//2),
+                0, 2*average_radius)
+            new_x = np.clip(vertex[-1][0] + r * math.cos(angles[i]), 0, w)
+            new_y = np.clip(vertex[-1][1] + r * math.sin(angles[i]), 0, h)
+            vertex.append((int(new_x), int(new_y)))
+
+        draw = ImageDraw.Draw(mask)
+        width = int(np.random.uniform(min_width, max_width))
+        draw.line(vertex, fill=1, width=width)
+        for v in vertex:
+            draw.ellipse((v[0] - width//2,
+                          v[1] - width//2,
+                          v[0] + width//2,
+                          v[1] + width//2),
+                         fill=1)
+        if np.random.random() > 0.5:
+            mask.transpose(Image.FLIP_LEFT_RIGHT)
+        if np.random.random() > 0.5:
+            mask.transpose(Image.FLIP_TOP_BOTTOM)
+    mask = np.asarray(mask, np.uint8)
+    if np.random.random() > 0.5:
+        mask = np.flip(mask, 0)
+    if np.random.random() > 0.5:
+        mask = np.flip(mask, 1)
+    return mask
+
+def RandomMask(s, hole_range=[0,1]):
+    coef = min(hole_range[0] + hole_range[1], 1.0)
+    while True:
+        mask = np.ones((s, s), np.uint8)
+        def Fill(max_size):
+            w, h = np.random.randint(max_size), np.random.randint(max_size)
+            ww, hh = w // 2, h // 2
+            x, y = np.random.randint(-ww, s - w + ww), np.random.randint(-hh, s - h + hh)
+            mask[max(y, 0): min(y + h, s), max(x, 0): min(x + w, s)] = 0
+        def MultiFill(max_tries, max_size):
+            for _ in range(np.random.randint(max_tries)):
+                Fill(max_size)
+        MultiFill(int(2 * coef), s // 2)
+        MultiFill(int(2 * coef), s)
+        mask = np.logical_and(mask, 1 - RandomBrush(int(3 * coef), s))  # hole denoted as 0, reserved as 1
+        hole_ratio = 1 - np.mean(mask)
+        if hole_range is not None and (hole_ratio <= hole_range[0] or hole_ratio >= hole_range[1]):
+            continue
+        return mask[np.newaxis, ...].astype(np.float32)
+
+def BatchRandomMask(batch_size, s, hole_range=[0, 1]):
+    return np.stack([RandomMask(s, hole_range=hole_range) for _ in range(batch_size)], axis=0)
+
+
+if __name__ == '__main__':
+    # res = 512
+    res = 256
+    cnt = 2000
+    tot = 0
+    for i in range(cnt):
+        mask = RandomMask(s=res)
+        tot += mask.mean()
+    print(tot / cnt)

datasets/mask_generator_512.py

@@ -0,0 +1,93 @@
+import numpy as np
+from PIL import Image, ImageDraw
+import math
+import random
+
+
+def RandomBrush(
+    max_tries,
+    s,
+    min_num_vertex = 4,
+    max_num_vertex = 18,
+    mean_angle = 2*math.pi / 5,
+    angle_range = 2*math.pi / 15,
+    min_width = 12,
+    max_width = 48):
+    H, W = s, s
+    average_radius = math.sqrt(H*H+W*W) / 8
+    mask = Image.new('L', (W, H), 0)
+    for _ in range(np.random.randint(max_tries)):
+        num_vertex = np.random.randint(min_num_vertex, max_num_vertex)
+        angle_min = mean_angle - np.random.uniform(0, angle_range)
+        angle_max = mean_angle + np.random.uniform(0, angle_range)
+        angles = []
+        vertex = []
+        for i in range(num_vertex):
+            if i % 2 == 0:
+                angles.append(2*math.pi - np.random.uniform(angle_min, angle_max))
+            else:
+                angles.append(np.random.uniform(angle_min, angle_max))
+
+        h, w = mask.size
+        vertex.append((int(np.random.randint(0, w)), int(np.random.randint(0, h))))
+        for i in range(num_vertex):
+            r = np.clip(
+                np.random.normal(loc=average_radius, scale=average_radius//2),
+                0, 2*average_radius)
+            new_x = np.clip(vertex[-1][0] + r * math.cos(angles[i]), 0, w)
+            new_y = np.clip(vertex[-1][1] + r * math.sin(angles[i]), 0, h)
+            vertex.append((int(new_x), int(new_y)))
+
+        draw = ImageDraw.Draw(mask)
+        width = int(np.random.uniform(min_width, max_width))
+        draw.line(vertex, fill=1, width=width)
+        for v in vertex:
+            draw.ellipse((v[0] - width//2,
+                          v[1] - width//2,
+                          v[0] + width//2,
+                          v[1] + width//2),
+                         fill=1)
+        if np.random.random() > 0.5:
+            mask.transpose(Image.FLIP_LEFT_RIGHT)
+        if np.random.random() > 0.5:
+            mask.transpose(Image.FLIP_TOP_BOTTOM)
+    mask = np.asarray(mask, np.uint8)
+    if np.random.random() > 0.5:
+        mask = np.flip(mask, 0)
+    if np.random.random() > 0.5:
+        mask = np.flip(mask, 1)
+    return mask
+
+def RandomMask(s, hole_range=[0,1]):
+    coef = min(hole_range[0] + hole_range[1], 1.0)
+    while True:
+        mask = np.ones((s, s), np.uint8)
+        def Fill(max_size):
+            w, h = np.random.randint(max_size), np.random.randint(max_size)
+            ww, hh = w // 2, h // 2
+            x, y = np.random.randint(-ww, s - w + ww), np.random.randint(-hh, s - h + hh)
+            mask[max(y, 0): min(y + h, s), max(x, 0): min(x + w, s)] = 0
+        def MultiFill(max_tries, max_size):
+            for _ in range(np.random.randint(max_tries)):
+                Fill(max_size)
+        MultiFill(int(5 * coef), s // 2)
+        MultiFill(int(3 * coef), s)
+        mask = np.logical_and(mask, 1 - RandomBrush(int(9 * coef), s))  # hole denoted as 0, reserved as 1
+        hole_ratio = 1 - np.mean(mask)
+        if hole_range is not None and (hole_ratio <= hole_range[0] or hole_ratio >= hole_range[1]):
+            continue
+        return mask[np.newaxis, ...].astype(np.float32)
+
+def BatchRandomMask(batch_size, s, hole_range=[0, 1]):
+    return np.stack([RandomMask(s, hole_range=hole_range) for _ in range(batch_size)], axis=0)
+
+
+if __name__ == '__main__':
+    res = 512
+    # res = 256
+    cnt = 2000
+    tot = 0
+    for i in range(cnt):
+        mask = RandomMask(s=res)
+        tot += mask.mean()
+    print(tot / cnt)

datasets/mask_generator_512_small.py

@@ -0,0 +1,93 @@
+import numpy as np
+from PIL import Image, ImageDraw
+import math
+import random
+
+
+def RandomBrush(
+    max_tries,
+    s,
+    min_num_vertex = 4,
+    max_num_vertex = 18,
+    mean_angle = 2*math.pi / 5,
+    angle_range = 2*math.pi / 15,
+    min_width = 12,
+    max_width = 48):
+    H, W = s, s
+    average_radius = math.sqrt(H*H+W*W) / 8
+    mask = Image.new('L', (W, H), 0)
+    for _ in range(np.random.randint(max_tries)):
+        num_vertex = np.random.randint(min_num_vertex, max_num_vertex)
+        angle_min = mean_angle - np.random.uniform(0, angle_range)
+        angle_max = mean_angle + np.random.uniform(0, angle_range)
+        angles = []
+        vertex = []
+        for i in range(num_vertex):
+            if i % 2 == 0:
+                angles.append(2*math.pi - np.random.uniform(angle_min, angle_max))
+            else:
+                angles.append(np.random.uniform(angle_min, angle_max))
+
+        h, w = mask.size
+        vertex.append((int(np.random.randint(0, w)), int(np.random.randint(0, h))))
+        for i in range(num_vertex):
+            r = np.clip(
+                np.random.normal(loc=average_radius, scale=average_radius//2),
+                0, 2*average_radius)
+            new_x = np.clip(vertex[-1][0] + r * math.cos(angles[i]), 0, w)
+            new_y = np.clip(vertex[-1][1] + r * math.sin(angles[i]), 0, h)
+            vertex.append((int(new_x), int(new_y)))
+
+        draw = ImageDraw.Draw(mask)
+        width = int(np.random.uniform(min_width, max_width))
+        draw.line(vertex, fill=1, width=width)
+        for v in vertex:
+            draw.ellipse((v[0] - width//2,
+                          v[1] - width//2,
+                          v[0] + width//2,
+                          v[1] + width//2),
+                         fill=1)
+        if np.random.random() > 0.5:
+            mask.transpose(Image.FLIP_LEFT_RIGHT)
+        if np.random.random() > 0.5:
+            mask.transpose(Image.FLIP_TOP_BOTTOM)
+    mask = np.asarray(mask, np.uint8)
+    if np.random.random() > 0.5:
+        mask = np.flip(mask, 0)
+    if np.random.random() > 0.5:
+        mask = np.flip(mask, 1)
+    return mask
+
+def RandomMask(s, hole_range=[0,1]):
+    coef = min(hole_range[0] + hole_range[1], 1.0)
+    while True:
+        mask = np.ones((s, s), np.uint8)
+        def Fill(max_size):
+            w, h = np.random.randint(max_size), np.random.randint(max_size)
+            ww, hh = w // 2, h // 2
+            x, y = np.random.randint(-ww, s - w + ww), np.random.randint(-hh, s - h + hh)
+            mask[max(y, 0): min(y + h, s), max(x, 0): min(x + w, s)] = 0
+        def MultiFill(max_tries, max_size):
+            for _ in range(np.random.randint(max_tries)):
+                Fill(max_size)
+        MultiFill(int(3 * coef), s // 2)
+        MultiFill(int(2 * coef), s)
+        mask = np.logical_and(mask, 1 - RandomBrush(int(4 * coef), s))  # hole denoted as 0, reserved as 1
+        hole_ratio = 1 - np.mean(mask)
+        if hole_range is not None and (hole_ratio <= hole_range[0] or hole_ratio >= hole_range[1]):
+            continue
+        return mask[np.newaxis, ...].astype(np.float32)
+
+def BatchRandomMask(batch_size, s, hole_range=[0, 1]):
+    return np.stack([RandomMask(s, hole_range=hole_range) for _ in range(batch_size)], axis=0)
+
+
+if __name__ == '__main__':
+    res = 512
+    # res = 256
+    cnt = 2000
+    tot = 0
+    for i in range(cnt):
+        mask = RandomMask(s=res)
+        tot += mask.mean()
+    print(tot / cnt)

dnnlib/__init__.py

@@ -0,0 +1,9 @@
+# Copyright (c) 2021, NVIDIA CORPORATION.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+from .util import EasyDict, make_cache_dir_path

dnnlib/util.py

@@ -0,0 +1,477 @@
+# Copyright (c) 2021, NVIDIA CORPORATION.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+"""Miscellaneous utility classes and functions."""
+
+import ctypes
+import fnmatch
+import importlib
+import inspect
+import numpy as np
+import os
+import shutil
+import sys
+import types
+import io
+import pickle
+import re
+import requests
+import html
+import hashlib
+import glob
+import tempfile
+import urllib
+import urllib.request
+import uuid
+
+from distutils.util import strtobool
+from typing import Any, List, Tuple, Union
+
+
+# Util classes
+# ------------------------------------------------------------------------------------------
+
+
+class EasyDict(dict):
+    """Convenience class that behaves like a dict but allows access with the attribute syntax."""
+
+    def __getattr__(self, name: str) -> Any:
+        try:
+            return self[name]
+        except KeyError:
+            raise AttributeError(name)
+
+    def __setattr__(self, name: str, value: Any) -> None:
+        self[name] = value
+
+    def __delattr__(self, name: str) -> None:
+        del self[name]
+
+
+class Logger(object):
+    """Redirect stderr to stdout, optionally print stdout to a file, and optionally force flushing on both stdout and the file."""
+
+    def __init__(self, file_name: str = None, file_mode: str = "w", should_flush: bool = True):
+        self.file = None
+
+        if file_name is not None:
+            self.file = open(file_name, file_mode)
+
+        self.should_flush = should_flush
+        self.stdout = sys.stdout
+        self.stderr = sys.stderr
+
+        sys.stdout = self
+        sys.stderr = self
+
+    def __enter__(self) -> "Logger":
+        return self
+
+    def __exit__(self, exc_type: Any, exc_value: Any, traceback: Any) -> None:
+        self.close()
+
+    def write(self, text: Union[str, bytes]) -> None:
+        """Write text to stdout (and a file) and optionally flush."""
+        if isinstance(text, bytes):
+            text = text.decode()
+        if len(text) == 0: # workaround for a bug in VSCode debugger: sys.stdout.write(''); sys.stdout.flush() => crash
+            return
+
+        if self.file is not None:
+            self.file.write(text)
+
+        self.stdout.write(text)
+
+        if self.should_flush:
+            self.flush()
+
+    def flush(self) -> None:
+        """Flush written text to both stdout and a file, if open."""
+        if self.file is not None:
+            self.file.flush()
+
+        self.stdout.flush()
+
+    def close(self) -> None:
+        """Flush, close possible files, and remove stdout/stderr mirroring."""
+        self.flush()
+
+        # if using multiple loggers, prevent closing in wrong order
+        if sys.stdout is self:
+            sys.stdout = self.stdout
+        if sys.stderr is self:
+            sys.stderr = self.stderr
+
+        if self.file is not None:
+            self.file.close()
+            self.file = None
+
+
+# Cache directories
+# ------------------------------------------------------------------------------------------
+
+_dnnlib_cache_dir = None
+
+def set_cache_dir(path: str) -> None:
+    global _dnnlib_cache_dir
+    _dnnlib_cache_dir = path
+
+def make_cache_dir_path(*paths: str) -> str:
+    if _dnnlib_cache_dir is not None:
+        return os.path.join(_dnnlib_cache_dir, *paths)
+    if 'DNNLIB_CACHE_DIR' in os.environ:
+        return os.path.join(os.environ['DNNLIB_CACHE_DIR'], *paths)
+    if 'HOME' in os.environ:
+        return os.path.join(os.environ['HOME'], '.cache', 'dnnlib', *paths)
+    if 'USERPROFILE' in os.environ:
+        return os.path.join(os.environ['USERPROFILE'], '.cache', 'dnnlib', *paths)
+    return os.path.join(tempfile.gettempdir(), '.cache', 'dnnlib', *paths)
+
+# Small util functions
+# ------------------------------------------------------------------------------------------
+
+
+def format_time(seconds: Union[int, float]) -> str:
+    """Convert the seconds to human readable string with days, hours, minutes and seconds."""
+    s = int(np.rint(seconds))
+
+    if s < 60:
+        return "{0}s".format(s)
+    elif s < 60 * 60:
+        return "{0}m {1:02}s".format(s // 60, s % 60)
+    elif s < 24 * 60 * 60:
+        return "{0}h {1:02}m {2:02}s".format(s // (60 * 60), (s // 60) % 60, s % 60)
+    else:
+        return "{0}d {1:02}h {2:02}m".format(s // (24 * 60 * 60), (s // (60 * 60)) % 24, (s // 60) % 60)
+
+
+def ask_yes_no(question: str) -> bool:
+    """Ask the user the question until the user inputs a valid answer."""
+    while True:
+        try:
+            print("{0} [y/n]".format(question))
+            return strtobool(input().lower())
+        except ValueError:
+            pass
+
+
+def tuple_product(t: Tuple) -> Any:
+    """Calculate the product of the tuple elements."""
+    result = 1
+
+    for v in t:
+        result *= v
+
+    return result
+
+
+_str_to_ctype = {
+    "uint8": ctypes.c_ubyte,
+    "uint16": ctypes.c_uint16,
+    "uint32": ctypes.c_uint32,
+    "uint64": ctypes.c_uint64,
+    "int8": ctypes.c_byte,
+    "int16": ctypes.c_int16,
+    "int32": ctypes.c_int32,
+    "int64": ctypes.c_int64,
+    "float32": ctypes.c_float,
+    "float64": ctypes.c_double
+}
+
+
+def get_dtype_and_ctype(type_obj: Any) -> Tuple[np.dtype, Any]:
+    """Given a type name string (or an object having a __name__ attribute), return matching Numpy and ctypes types that have the same size in bytes."""
+    type_str = None
+
+    if isinstance(type_obj, str):
+        type_str = type_obj
+    elif hasattr(type_obj, "__name__"):
+        type_str = type_obj.__name__
+    elif hasattr(type_obj, "name"):
+        type_str = type_obj.name
+    else:
+        raise RuntimeError("Cannot infer type name from input")
+
+    assert type_str in _str_to_ctype.keys()
+
+    my_dtype = np.dtype(type_str)
+    my_ctype = _str_to_ctype[type_str]
+
+    assert my_dtype.itemsize == ctypes.sizeof(my_ctype)
+
+    return my_dtype, my_ctype
+
+
+def is_pickleable(obj: Any) -> bool:
+    try:
+        with io.BytesIO() as stream:
+            pickle.dump(obj, stream)
+        return True
+    except:
+        return False
+
+
+# Functionality to import modules/objects by name, and call functions by name
+# ------------------------------------------------------------------------------------------
+
+def get_module_from_obj_name(obj_name: str) -> Tuple[types.ModuleType, str]:
+    """Searches for the underlying module behind the name to some python object.
+    Returns the module and the object name (original name with module part removed)."""
+
+    # allow convenience shorthands, substitute them by full names
+    obj_name = re.sub("^np.", "numpy.", obj_name)
+    obj_name = re.sub("^tf.", "tensorflow.", obj_name)
+
+    # list alternatives for (module_name, local_obj_name)
+    parts = obj_name.split(".")
+    name_pairs = [(".".join(parts[:i]), ".".join(parts[i:])) for i in range(len(parts), 0, -1)]
+
+    # try each alternative in turn
+    for module_name, local_obj_name in name_pairs:
+        try:
+            module = importlib.import_module(module_name) # may raise ImportError
+            get_obj_from_module(module, local_obj_name) # may raise AttributeError
+            return module, local_obj_name
+        except:
+            pass
+
+    # maybe some of the modules themselves contain errors?
+    for module_name, _local_obj_name in name_pairs:
+        try:
+            importlib.import_module(module_name) # may raise ImportError
+        except ImportError:
+            if not str(sys.exc_info()[1]).startswith("No module named '" + module_name + "'"):
+                raise
+
+    # maybe the requested attribute is missing?
+    for module_name, local_obj_name in name_pairs:
+        try:
+            module = importlib.import_module(module_name) # may raise ImportError
+            get_obj_from_module(module, local_obj_name) # may raise AttributeError
+        except ImportError:
+            pass
+
+    # we are out of luck, but we have no idea why
+    raise ImportError(obj_name)
+
+
+def get_obj_from_module(module: types.ModuleType, obj_name: str) -> Any:
+    """Traverses the object name and returns the last (rightmost) python object."""
+    if obj_name == '':
+        return module
+    obj = module
+    for part in obj_name.split("."):
+        obj = getattr(obj, part)
+    return obj
+
+
+def get_obj_by_name(name: str) -> Any:
+    """Finds the python object with the given name."""
+    module, obj_name = get_module_from_obj_name(name)
+    return get_obj_from_module(module, obj_name)
+
+
+def call_func_by_name(*args, func_name: str = None, **kwargs) -> Any:
+    """Finds the python object with the given name and calls it as a function."""
+    assert func_name is not None
+    func_obj = get_obj_by_name(func_name)
+    assert callable(func_obj)
+    return func_obj(*args, **kwargs)
+
+
+def construct_class_by_name(*args, class_name: str = None, **kwargs) -> Any:
+    """Finds the python class with the given name and constructs it with the given arguments."""
+    return call_func_by_name(*args, func_name=class_name, **kwargs)
+
+
+def get_module_dir_by_obj_name(obj_name: str) -> str:
+    """Get the directory path of the module containing the given object name."""
+    module, _ = get_module_from_obj_name(obj_name)
+    return os.path.dirname(inspect.getfile(module))
+
+
+def is_top_level_function(obj: Any) -> bool:
+    """Determine whether the given object is a top-level function, i.e., defined at module scope using 'def'."""
+    return callable(obj) and obj.__name__ in sys.modules[obj.__module__].__dict__
+
+
+def get_top_level_function_name(obj: Any) -> str:
+    """Return the fully-qualified name of a top-level function."""
+    assert is_top_level_function(obj)
+    module = obj.__module__
+    if module == '__main__':
+        module = os.path.splitext(os.path.basename(sys.modules[module].__file__))[0]
+    return module + "." + obj.__name__
+
+
+# File system helpers
+# ------------------------------------------------------------------------------------------
+
+def list_dir_recursively_with_ignore(dir_path: str, ignores: List[str] = None, add_base_to_relative: bool = False) -> List[Tuple[str, str]]:
+    """List all files recursively in a given directory while ignoring given file and directory names.
+    Returns list of tuples containing both absolute and relative paths."""
+    assert os.path.isdir(dir_path)
+    base_name = os.path.basename(os.path.normpath(dir_path))
+
+    if ignores is None:
+        ignores = []
+
+    result = []
+
+    for root, dirs, files in os.walk(dir_path, topdown=True):
+        for ignore_ in ignores:
+            dirs_to_remove = [d for d in dirs if fnmatch.fnmatch(d, ignore_)]
+
+            # dirs need to be edited in-place
+            for d in dirs_to_remove:
+                dirs.remove(d)
+
+            files = [f for f in files if not fnmatch.fnmatch(f, ignore_)]
+
+        absolute_paths = [os.path.join(root, f) for f in files]
+        relative_paths = [os.path.relpath(p, dir_path) for p in absolute_paths]
+
+        if add_base_to_relative:
+            relative_paths = [os.path.join(base_name, p) for p in relative_paths]
+
+        assert len(absolute_paths) == len(relative_paths)
+        result += zip(absolute_paths, relative_paths)
+
+    return result
+
+
+def copy_files_and_create_dirs(files: List[Tuple[str, str]]) -> None:
+    """Takes in a list of tuples of (src, dst) paths and copies files.
+    Will create all necessary directories."""
+    for file in files:
+        target_dir_name = os.path.dirname(file[1])
+
+        # will create all intermediate-level directories
+        if not os.path.exists(target_dir_name):
+            os.makedirs(target_dir_name)
+
+        shutil.copyfile(file[0], file[1])
+
+
+# URL helpers
+# ------------------------------------------------------------------------------------------
+
+def is_url(obj: Any, allow_file_urls: bool = False) -> bool:
+    """Determine whether the given object is a valid URL string."""
+    if not isinstance(obj, str) or not "://" in obj:
+        return False
+    if allow_file_urls and obj.startswith('file://'):
+        return True
+    try:
+        res = requests.compat.urlparse(obj)
+        if not res.scheme or not res.netloc or not "." in res.netloc:
+            return False
+        res = requests.compat.urlparse(requests.compat.urljoin(obj, "/"))
+        if not res.scheme or not res.netloc or not "." in res.netloc:
+            return False
+    except:
+        return False
+    return True
+
+
+def open_url(url: str, cache_dir: str = None, num_attempts: int = 10, verbose: bool = True, return_filename: bool = False, cache: bool = True) -> Any:
+    """Download the given URL and return a binary-mode file object to access the data."""
+    assert num_attempts >= 1
+    assert not (return_filename and (not cache))
+
+    # Doesn't look like an URL scheme so interpret it as a local filename.
+    if not re.match('^[a-z]+://', url):
+        return url if return_filename else open(url, "rb")
+
+    # Handle file URLs.  This code handles unusual file:// patterns that
+    # arise on Windows:
+    #
+    # file:///c:/foo.txt
+    #
+    # which would translate to a local '/c:/foo.txt' filename that's
+    # invalid.  Drop the forward slash for such pathnames.
+    #
+    # If you touch this code path, you should test it on both Linux and
+    # Windows.
+    #
+    # Some internet resources suggest using urllib.request.url2pathname() but
+    # but that converts forward slashes to backslashes and this causes
+    # its own set of problems.
+    if url.startswith('file://'):
+        filename = urllib.parse.urlparse(url).path
+        if re.match(r'^/[a-zA-Z]:', filename):
+            filename = filename[1:]
+        return filename if return_filename else open(filename, "rb")
+
+    assert is_url(url)
+
+    # Lookup from cache.
+    if cache_dir is None:
+        cache_dir = make_cache_dir_path('downloads')
+
+    url_md5 = hashlib.md5(url.encode("utf-8")).hexdigest()
+    if cache:
+        cache_files = glob.glob(os.path.join(cache_dir, url_md5 + "_*"))
+        if len(cache_files) == 1:
+            filename = cache_files[0]
+            return filename if return_filename else open(filename, "rb")
+
+    # Download.
+    url_name = None
+    url_data = None
+    with requests.Session() as session:
+        if verbose:
+            print("Downloading %s ..." % url, end="", flush=True)
+        for attempts_left in reversed(range(num_attempts)):
+            try:
+                with session.get(url) as res:
+                    res.raise_for_status()
+                    if len(res.content) == 0:
+                        raise IOError("No data received")
+
+                    if len(res.content) < 8192:
+                        content_str = res.content.decode("utf-8")
+                        if "download_warning" in res.headers.get("Set-Cookie", ""):
+                            links = [html.unescape(link) for link in content_str.split('"') if "export=download" in link]
+                            if len(links) == 1:
+                                url = requests.compat.urljoin(url, links[0])
+                                raise IOError("Google Drive virus checker nag")
+                        if "Google Drive - Quota exceeded" in content_str:
+                            raise IOError("Google Drive download quota exceeded -- please try again later")
+
+                    match = re.search(r'filename="([^"]*)"', res.headers.get("Content-Disposition", ""))
+                    url_name = match[1] if match else url
+                    url_data = res.content
+                    if verbose:
+                        print(" done")
+                    break
+            except KeyboardInterrupt:
+                raise
+            except:
+                if not attempts_left:
+                    if verbose:
+                        print(" failed")
+                    raise
+                if verbose:
+                    print(".", end="", flush=True)
+
+    # Save to cache.
+    if cache:
+        safe_name = re.sub(r"[^0-9a-zA-Z-._]", "_", url_name)
+        cache_file = os.path.join(cache_dir, url_md5 + "_" + safe_name)
+        temp_file = os.path.join(cache_dir, "tmp_" + uuid.uuid4().hex + "_" + url_md5 + "_" + safe_name)
+        os.makedirs(cache_dir, exist_ok=True)
+        with open(temp_file, "wb") as f:
+            f.write(url_data)
+        os.replace(temp_file, cache_file) # atomic
+        if return_filename:
+            return cache_file
+
+    # Return data as file object.
+    assert not return_filename
+    return io.BytesIO(url_data)

evaluatoin/cal_fid_pids_uids.py

@@ -0,0 +1,193 @@
+import cv2
+import os
+import sys
+sys.path.insert(0, '../')
+import numpy as np
+import math
+import glob
+import pyspng
+import PIL.Image
+import torch
+import dnnlib
+import scipy.linalg
+import sklearn.svm
+
+
+_feature_detector_cache = dict()
+
+def get_feature_detector(url, device=torch.device('cpu'), num_gpus=1, rank=0, verbose=False):
+    assert 0 <= rank < num_gpus
+    key = (url, device)
+    if key not in _feature_detector_cache:
+        is_leader = (rank == 0)
+        if not is_leader and num_gpus > 1:
+            torch.distributed.barrier() # leader goes first
+        with dnnlib.util.open_url(url, verbose=(verbose and is_leader)) as f:
+            _feature_detector_cache[key] = torch.jit.load(f).eval().to(device)
+        if is_leader and num_gpus > 1:
+            torch.distributed.barrier() # others follow
+    return _feature_detector_cache[key]
+
+
+def read_image(image_path):
+    with open(image_path, 'rb') as f:
+        if pyspng is not None and image_path.endswith('.png'):
+            image = pyspng.load(f.read())
+        else:
+            image = np.array(PIL.Image.open(f))
+    if image.ndim == 2:
+        image = image[:, :, np.newaxis] # HW => HWC
+    if image.shape[2] == 1:
+        image = np.repeat(image, 3, axis=2)
+    image = image.transpose(2, 0, 1) # HWC => CHW
+    image = torch.from_numpy(image).unsqueeze(0).to(torch.uint8)
+
+    return image
+
+
+class FeatureStats:
+    def __init__(self, capture_all=False, capture_mean_cov=False, max_items=None):
+        self.capture_all = capture_all
+        self.capture_mean_cov = capture_mean_cov
+        self.max_items = max_items
+        self.num_items = 0
+        self.num_features = None
+        self.all_features = None
+        self.raw_mean = None
+        self.raw_cov = None
+
+    def set_num_features(self, num_features):
+        if self.num_features is not None:
+            assert num_features == self.num_features
+        else:
+            self.num_features = num_features
+            self.all_features = []
+            self.raw_mean = np.zeros([num_features], dtype=np.float64)
+            self.raw_cov = np.zeros([num_features, num_features], dtype=np.float64)
+
+    def is_full(self):
+        return (self.max_items is not None) and (self.num_items >= self.max_items)
+
+    def append(self, x):
+        x = np.asarray(x, dtype=np.float32)
+        assert x.ndim == 2
+        if (self.max_items is not None) and (self.num_items + x.shape[0] > self.max_items):
+            if self.num_items >= self.max_items:
+                return
+            x = x[:self.max_items - self.num_items]
+
+        self.set_num_features(x.shape[1])
+        self.num_items += x.shape[0]
+        if self.capture_all:
+            self.all_features.append(x)
+        if self.capture_mean_cov:
+            x64 = x.astype(np.float64)
+            self.raw_mean += x64.sum(axis=0)
+            self.raw_cov += x64.T @ x64
+
+    def append_torch(self, x, num_gpus=1, rank=0):
+        assert isinstance(x, torch.Tensor) and x.ndim == 2
+        assert 0 <= rank < num_gpus
+        if num_gpus > 1:
+            ys = []
+            for src in range(num_gpus):
+                y = x.clone()
+                torch.distributed.broadcast(y, src=src)
+                ys.append(y)
+            x = torch.stack(ys, dim=1).flatten(0, 1) # interleave samples
+        self.append(x.cpu().numpy())
+
+    def get_all(self):
+        assert self.capture_all
+        return np.concatenate(self.all_features, axis=0)
+
+    def get_all_torch(self):
+        return torch.from_numpy(self.get_all())
+
+    def get_mean_cov(self):
+        assert self.capture_mean_cov
+        mean = self.raw_mean / self.num_items
+        cov = self.raw_cov / self.num_items
+        cov = cov - np.outer(mean, mean)
+        return mean, cov
+
+    def save(self, pkl_file):
+        with open(pkl_file, 'wb') as f:
+            pickle.dump(self.__dict__, f)
+
+    @staticmethod
+    def load(pkl_file):
+        with open(pkl_file, 'rb') as f:
+            s = dnnlib.EasyDict(pickle.load(f))
+        obj = FeatureStats(capture_all=s.capture_all, max_items=s.max_items)
+        obj.__dict__.update(s)
+        return obj
+
+
+def calculate_metrics(folder1, folder2):
+    l1 = sorted(glob.glob(folder1 + '/*.png') + glob.glob(folder1 + '/*.jpg'))
+    l2 = sorted(glob.glob(folder2 + '/*.png') + glob.glob(folder2 + '/*.jpg'))
+    assert(len(l1) == len(l2))
+    print('length:', len(l1))
+
+    # l1 = l1[:3]; l2 = l2[:3];
+
+    # build detector
+    detector_url = 'https://nvlabs-fi-cdn.nvidia.com/stylegan2-ada-pytorch/pretrained/metrics/inception-2015-12-05.pt'
+    detector_kwargs = dict(return_features=True) # Return raw features before the softmax layer.
+    device = torch.device('cuda:0')
+    detector = get_feature_detector(url=detector_url, device=device, num_gpus=1, rank=0, verbose=False)
+    detector.eval()
+
+    stat1 = FeatureStats(capture_all=True, capture_mean_cov=True, max_items=len(l1))
+    stat2 = FeatureStats(capture_all=True, capture_mean_cov=True, max_items=len(l1))
+
+    with torch.no_grad():
+        for i, (fpath1, fpath2) in enumerate(zip(l1, l2)):
+            print(i)
+            _, name1 = os.path.split(fpath1)
+            _, name2 = os.path.split(fpath2)
+            name1 = name1.split('.')[0]
+            name2 = name2.split('.')[0]
+            assert name1 == name2, 'Illegal mapping: %s, %s' % (name1, name2)
+
+            img1 = read_image(fpath1).to(device)
+            img2 = read_image(fpath2).to(device)
+            assert img1.shape == img2.shape, 'Illegal shape'
+            fea1 = detector(img1, **detector_kwargs)
+            stat1.append_torch(fea1, num_gpus=1, rank=0)
+            fea2 = detector(img2, **detector_kwargs)
+            stat2.append_torch(fea2, num_gpus=1, rank=0)
+
+    # calculate fid
+    mu1, sigma1 = stat1.get_mean_cov()
+    mu2, sigma2 = stat2.get_mean_cov()
+    m = np.square(mu1 - mu2).sum()
+    s, _ = scipy.linalg.sqrtm(np.dot(sigma1, sigma2), disp=False) # pylint: disable=no-member
+    fid = np.real(m + np.trace(sigma1 + sigma2 - s * 2))
+
+    # calculate pids and uids
+    fake_activations = stat1.get_all()
+    real_activations = stat2.get_all()
+    svm = sklearn.svm.LinearSVC(dual=False)
+    svm_inputs = np.concatenate([real_activations, fake_activations])
+    svm_targets = np.array([1] * real_activations.shape[0] + [0] * fake_activations.shape[0])
+    print('SVM fitting ...')
+    svm.fit(svm_inputs, svm_targets)
+    uids = 1 - svm.score(svm_inputs, svm_targets)
+    real_outputs = svm.decision_function(real_activations)
+    fake_outputs = svm.decision_function(fake_activations)
+    pids = np.mean(fake_outputs > real_outputs)
+
+    return fid, pids, uids
+
+
+if __name__ == '__main__':
+    folder1 = 'path to the inpainted result'
+    folder2 = 'path to the gt'
+
+    fid, pids, uids = calculate_metrics(folder1, folder2)
+    print('fid: %.4f, pids: %.4f, uids: %.4f' % (fid, pids, uids))
+    with open('fid_pids_uids.txt', 'w') as f:
+        f.write('fid: %.4f, pids: %.4f, uids: %.4f' % (fid, pids, uids))
+

evaluatoin/cal_lpips.py

@@ -0,0 +1,71 @@
+import cv2
+import os
+import sys
+import numpy as np
+import math
+import glob
+import pyspng
+import PIL.Image
+
+import torch
+import lpips
+
+
+def read_image(image_path):
+    with open(image_path, 'rb') as f:
+        if pyspng is not None and image_path.endswith('.png'):
+            image = pyspng.load(f.read())
+        else:
+            image = np.array(PIL.Image.open(f))
+    if image.ndim == 2:
+        image = image[:, :, np.newaxis] # HW => HWC
+    if image.shape[2] == 1:
+        image = np.repeat(image, 3, axis=2)
+    image = image.transpose(2, 0, 1) # HWC => CHW
+    image = torch.from_numpy(image).float().unsqueeze(0)
+    image = image / 127.5 - 1
+
+    return image
+
+
+def calculate_metrics(folder1, folder2):
+    l1 = sorted(glob.glob(folder1 + '/*.png') + glob.glob(folder1 + '/*.jpg'))
+    l2 = sorted(glob.glob(folder2 + '/*.png') + glob.glob(folder2 + '/*.jpg'))
+    assert(len(l1) == len(l2))
+    print('length:', len(l1))
+
+    # l1 = l1[:3]; l2 = l2[:3];
+
+    device = torch.device('cuda:0')
+    loss_fn = lpips.LPIPS(net='alex').to(device)
+    loss_fn.eval()
+    # loss_fn = lpips.LPIPS(net='vgg').to(device)
+
+    lpips_l = []
+    with torch.no_grad():
+        for i, (fpath1, fpath2) in enumerate(zip(l1, l2)):
+            print(i)
+            _, name1 = os.path.split(fpath1)
+            _, name2 = os.path.split(fpath2)
+            name1 = name1.split('.')[0]
+            name2 = name2.split('.')[0]
+            assert name1 == name2, 'Illegal mapping: %s, %s' % (name1, name2)
+
+            img1 = read_image(fpath1).to(device)
+            img2 = read_image(fpath2).to(device)
+            assert img1.shape == img2.shape, 'Illegal shape'
+            lpips_l.append(loss_fn(img1, img2).mean().cpu().numpy())
+
+    res = sum(lpips_l) / len(lpips_l)
+
+    return res
+
+
+if __name__ == '__main__':
+    folder1 = 'path to the inpainted result'
+    folder2 = 'path to the gt'
+
+    res = calculate_metrics(folder1, folder2)
+    print('lpips: %.4f' % res)
+    with open('lpips.txt', 'w') as f:
+        f.write('lpips: %.4f' % res)

evaluatoin/cal_psnr_ssim_l1.py

@@ -0,0 +1,107 @@
+import cv2
+import os
+import sys
+import numpy as np
+import math
+import glob
+import pyspng
+import PIL.Image
+
+
+def calculate_psnr(img1, img2):
+    # img1 and img2 have range [0, 255]
+    img1 = img1.astype(np.float64)
+    img2 = img2.astype(np.float64)
+    mse = np.mean((img1 - img2) ** 2)
+    if mse == 0:
+        return float('inf')
+
+    return 20 * math.log10(255.0 / math.sqrt(mse))
+
+
+def calculate_ssim(img1, img2):
+    C1 = (0.01 * 255) ** 2
+    C2 = (0.03 * 255) ** 2
+
+    img1 = img1.astype(np.float64)
+    img2 = img2.astype(np.float64)
+    kernel = cv2.getGaussianKernel(11, 1.5)
+    window = np.outer(kernel, kernel.transpose())
+
+    mu1 = cv2.filter2D(img1, -1, window)[5:-5, 5:-5]
+    mu2 = cv2.filter2D(img2, -1, window)[5:-5, 5:-5]
+    mu1_sq = mu1 ** 2
+    mu2_sq = mu2 ** 2
+    mu1_mu2 = mu1 * mu2
+    sigma1_sq = cv2.filter2D(img1 ** 2, -1, window)[5:-5, 5:-5] - mu1_sq
+    sigma2_sq = cv2.filter2D(img2 ** 2, -1, window)[5:-5, 5:-5] - mu2_sq
+    sigma12 = cv2.filter2D(img1 * img2, -1, window)[5:-5, 5:-5] - mu1_mu2
+
+    ssim_map = ((2 * mu1_mu2 + C1) * (2 * sigma12 + C2)) / ((mu1_sq + mu2_sq + C1) * (sigma1_sq + sigma2_sq + C2))
+
+    return ssim_map.mean()
+
+
+def calculate_l1(img1, img2):
+    img1 = img1.astype(np.float64) / 255.0
+    img2 = img2.astype(np.float64) / 255.0
+    l1 = np.mean(np.abs(img1 - img2))
+
+    return l1
+
+
+def read_image(image_path):
+    with open(image_path, 'rb') as f:
+        if pyspng is not None and image_path.endswith('.png'):
+            image = pyspng.load(f.read())
+        else:
+            image = np.array(PIL.Image.open(f))
+    if image.ndim == 2:
+        image = image[:, :, np.newaxis] # HW => HWC
+    if image.shape[2] == 1:
+        image = np.repeat(image, 3, axis=2)
+    # image = image.transpose(2, 0, 1) # HWC => CHW
+
+    return image
+
+
+def calculate_metrics(folder1, folder2):
+    l1 = sorted(glob.glob(folder1 + '/*.png') + glob.glob(folder1 + '/*.jpg'))
+    l2 = sorted(glob.glob(folder2 + '/*.png') + glob.glob(folder2 + '/*.jpg'))
+    assert(len(l1) == len(l2))
+    print('length:', len(l1))
+
+    # l1 = l1[:3]; l2 = l2[:3];
+
+    psnr_l, ssim_l, dl1_l = [], [], []
+    for i, (fpath1, fpath2) in enumerate(zip(l1, l2)):
+        print(i)
+        _, name1 = os.path.split(fpath1)
+        _, name2 = os.path.split(fpath2)
+        name1 = name1.split('.')[0]
+        name2 = name2.split('.')[0]
+        assert name1 == name2, 'Illegal mapping: %s, %s' % (name1, name2)
+
+        img1 = read_image(fpath1).astype(np.float64)
+        img2 = read_image(fpath2).astype(np.float64)
+        assert img1.shape == img2.shape, 'Illegal shape'
+        psnr_l.append(calculate_psnr(img1, img2))
+        ssim_l.append(calculate_ssim(img1, img2))
+        dl1_l.append(calculate_l1(img1, img2))
+
+    psnr = sum(psnr_l) / len(psnr_l)
+    ssim = sum(ssim_l) / len(ssim_l)
+    dl1 = sum(dl1_l) / len(dl1_l)
+
+    return psnr, ssim, dl1
+
+
+if __name__ == '__main__':
+    folder1 = 'path to the inpainted result'
+    folder2 = 'path to the gt'
+
+    psnr, ssim, dl1 = calculate_metrics(folder1, folder2)
+    print('psnr: %.4f, ssim: %.4f, l1: %.4f' % (psnr, ssim, dl1))
+    with open('psnr_ssim_l1.txt', 'w') as f:
+        f.write('psnr: %.4f, ssim: %.4f, l1: %.4f' % (psnr, ssim, dl1))
+

legacy.py

@@ -0,0 +1,323 @@
+# Copyright (c) 2021, NVIDIA CORPORATION.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+import click
+import pickle
+import re
+import copy
+import numpy as np
+import torch
+import dnnlib
+from torch_utils import misc
+
+#----------------------------------------------------------------------------
+
+def load_network_pkl(f, force_fp16=False):
+    data = _LegacyUnpickler(f).load()
+
+    # Legacy TensorFlow pickle => convert.
+    if isinstance(data, tuple) and len(data) == 3 and all(isinstance(net, _TFNetworkStub) for net in data):
+        tf_G, tf_D, tf_Gs = data
+        G = convert_tf_generator(tf_G)
+        D = convert_tf_discriminator(tf_D)
+        G_ema = convert_tf_generator(tf_Gs)
+        data = dict(G=G, D=D, G_ema=G_ema)
+
+    # Add missing fields.
+    if 'training_set_kwargs' not in data:
+        data['training_set_kwargs'] = None
+    if 'augment_pipe' not in data:
+        data['augment_pipe'] = None
+
+    # Validate contents.
+    assert isinstance(data['G'], torch.nn.Module)
+    assert isinstance(data['D'], torch.nn.Module)
+    assert isinstance(data['G_ema'], torch.nn.Module)
+    assert isinstance(data['training_set_kwargs'], (dict, type(None)))
+    assert isinstance(data['augment_pipe'], (torch.nn.Module, type(None)))
+
+    # Force FP16.
+    if force_fp16:
+        for key in ['G', 'D', 'G_ema']:
+            old = data[key]
+            kwargs = copy.deepcopy(old.init_kwargs)
+            if key.startswith('G'):
+                kwargs.synthesis_kwargs = dnnlib.EasyDict(kwargs.get('synthesis_kwargs', {}))
+                kwargs.synthesis_kwargs.num_fp16_res = 4
+                kwargs.synthesis_kwargs.conv_clamp = 256
+            if key.startswith('D'):
+                kwargs.num_fp16_res = 4
+                kwargs.conv_clamp = 256
+            if kwargs != old.init_kwargs:
+                new = type(old)(**kwargs).eval().requires_grad_(False)
+                misc.copy_params_and_buffers(old, new, require_all=True)
+                data[key] = new
+    return data
+
+#----------------------------------------------------------------------------
+
+class _TFNetworkStub(dnnlib.EasyDict):
+    pass
+
+class _LegacyUnpickler(pickle.Unpickler):
+    def find_class(self, module, name):
+        if module == 'torch.storage' and name == '_load_from_bytes':
+            import io
+            return lambda b: torch.load(io.BytesIO(b), map_location='cpu')
+        if module == 'dnnlib.tflib.network' and name == 'Network':
+            return _TFNetworkStub
+        return super().find_class(module, name)
+
+#----------------------------------------------------------------------------
+
+def _collect_tf_params(tf_net):
+    # pylint: disable=protected-access
+    tf_params = dict()
+    def recurse(prefix, tf_net):
+        for name, value in tf_net.variables:
+            tf_params[prefix + name] = value
+        for name, comp in tf_net.components.items():
+            recurse(prefix + name + '/', comp)
+    recurse('', tf_net)
+    return tf_params
+
+#----------------------------------------------------------------------------
+
+def _populate_module_params(module, *patterns):
+    for name, tensor in misc.named_params_and_buffers(module):
+        found = False
+        value = None
+        for pattern, value_fn in zip(patterns[0::2], patterns[1::2]):
+            match = re.fullmatch(pattern, name)
+            if match:
+                found = True
+                if value_fn is not None:
+                    value = value_fn(*match.groups())
+                break
+        try:
+            assert found
+            if value is not None:
+                tensor.copy_(torch.from_numpy(np.array(value)))
+        except:
+            print(name, list(tensor.shape))
+            raise
+
+#----------------------------------------------------------------------------
+
+def convert_tf_generator(tf_G):
+    if tf_G.version < 4:
+        raise ValueError('TensorFlow pickle version too low')
+
+    # Collect kwargs.
+    tf_kwargs = tf_G.static_kwargs
+    known_kwargs = set()
+    def kwarg(tf_name, default=None, none=None):
+        known_kwargs.add(tf_name)
+        val = tf_kwargs.get(tf_name, default)
+        return val if val is not None else none
+
+    # Convert kwargs.
+    kwargs = dnnlib.EasyDict(
+        z_dim                   = kwarg('latent_size',          512),
+        c_dim                   = kwarg('label_size',           0),
+        w_dim                   = kwarg('dlatent_size',         512),
+        img_resolution          = kwarg('resolution',           1024),
+        img_channels            = kwarg('num_channels',         3),
+        mapping_kwargs = dnnlib.EasyDict(
+            num_layers          = kwarg('mapping_layers',       8),
+            embed_features      = kwarg('label_fmaps',          None),
+            layer_features      = kwarg('mapping_fmaps',        None),
+            activation          = kwarg('mapping_nonlinearity', 'lrelu'),
+            lr_multiplier       = kwarg('mapping_lrmul',        0.01),
+            w_avg_beta          = kwarg('w_avg_beta',           0.995,  none=1),
+        ),
+        synthesis_kwargs = dnnlib.EasyDict(
+            channel_base        = kwarg('fmap_base',            16384) * 2,
+            channel_max         = kwarg('fmap_max',             512),
+            num_fp16_res        = kwarg('num_fp16_res',         0),
+            conv_clamp          = kwarg('conv_clamp',           None),
+            architecture        = kwarg('architecture',         'skip'),
+            resample_filter     = kwarg('resample_kernel',      [1,3,3,1]),
+            use_noise           = kwarg('use_noise',            True),
+            activation          = kwarg('nonlinearity',         'lrelu'),
+        ),
+    )
+
+    # Check for unknown kwargs.
+    kwarg('truncation_psi')
+    kwarg('truncation_cutoff')
+    kwarg('style_mixing_prob')
+    kwarg('structure')
+    unknown_kwargs = list(set(tf_kwargs.keys()) - known_kwargs)
+    if len(unknown_kwargs) > 0:
+        raise ValueError('Unknown TensorFlow kwarg', unknown_kwargs[0])
+
+    # Collect params.
+    tf_params = _collect_tf_params(tf_G)
+    for name, value in list(tf_params.items()):
+        match = re.fullmatch(r'ToRGB_lod(\d+)/(.*)', name)
+        if match:
+            r = kwargs.img_resolution // (2 ** int(match.group(1)))
+            tf_params[f'{r}x{r}/ToRGB/{match.group(2)}'] = value
+            kwargs.synthesis.kwargs.architecture = 'orig'
+    #for name, value in tf_params.items(): print(f'{name:<50s}{list(value.shape)}')
+
+    # Convert params.
+    from training import networks
+    G = networks.Generator(**kwargs).eval().requires_grad_(False)
+    # pylint: disable=unnecessary-lambda
+    _populate_module_params(G,
+        r'mapping\.w_avg',                                  lambda:     tf_params[f'dlatent_avg'],
+        r'mapping\.embed\.weight',                          lambda:     tf_params[f'mapping/LabelEmbed/weight'].transpose(),
+        r'mapping\.embed\.bias',                            lambda:     tf_params[f'mapping/LabelEmbed/bias'],
+        r'mapping\.fc(\d+)\.weight',                        lambda i:   tf_params[f'mapping/Dense{i}/weight'].transpose(),
+        r'mapping\.fc(\d+)\.bias',                          lambda i:   tf_params[f'mapping/Dense{i}/bias'],
+        r'synthesis\.b4\.const',                            lambda:     tf_params[f'synthesis/4x4/Const/const'][0],
+        r'synthesis\.b4\.conv1\.weight',                    lambda:     tf_params[f'synthesis/4x4/Conv/weight'].transpose(3, 2, 0, 1),
+        r'synthesis\.b4\.conv1\.bias',                      lambda:     tf_params[f'synthesis/4x4/Conv/bias'],
+        r'synthesis\.b4\.conv1\.noise_const',               lambda:     tf_params[f'synthesis/noise0'][0, 0],
+        r'synthesis\.b4\.conv1\.noise_strength',            lambda:     tf_params[f'synthesis/4x4/Conv/noise_strength'],
+        r'synthesis\.b4\.conv1\.affine\.weight',            lambda:     tf_params[f'synthesis/4x4/Conv/mod_weight'].transpose(),
+        r'synthesis\.b4\.conv1\.affine\.bias',              lambda:     tf_params[f'synthesis/4x4/Conv/mod_bias'] + 1,
+        r'synthesis\.b(\d+)\.conv0\.weight',                lambda r:   tf_params[f'synthesis/{r}x{r}/Conv0_up/weight'][::-1, ::-1].transpose(3, 2, 0, 1),
+        r'synthesis\.b(\d+)\.conv0\.bias',                  lambda r:   tf_params[f'synthesis/{r}x{r}/Conv0_up/bias'],
+        r'synthesis\.b(\d+)\.conv0\.noise_const',           lambda r:   tf_params[f'synthesis/noise{int(np.log2(int(r)))*2-5}'][0, 0],
+        r'synthesis\.b(\d+)\.conv0\.noise_strength',        lambda r:   tf_params[f'synthesis/{r}x{r}/Conv0_up/noise_strength'],
+        r'synthesis\.b(\d+)\.conv0\.affine\.weight',        lambda r:   tf_params[f'synthesis/{r}x{r}/Conv0_up/mod_weight'].transpose(),
+        r'synthesis\.b(\d+)\.conv0\.affine\.bias',          lambda r:   tf_params[f'synthesis/{r}x{r}/Conv0_up/mod_bias'] + 1,
+        r'synthesis\.b(\d+)\.conv1\.weight',                lambda r:   tf_params[f'synthesis/{r}x{r}/Conv1/weight'].transpose(3, 2, 0, 1),
+        r'synthesis\.b(\d+)\.conv1\.bias',                  lambda r:   tf_params[f'synthesis/{r}x{r}/Conv1/bias'],
+        r'synthesis\.b(\d+)\.conv1\.noise_const',           lambda r:   tf_params[f'synthesis/noise{int(np.log2(int(r)))*2-4}'][0, 0],
+        r'synthesis\.b(\d+)\.conv1\.noise_strength',        lambda r:   tf_params[f'synthesis/{r}x{r}/Conv1/noise_strength'],
+        r'synthesis\.b(\d+)\.conv1\.affine\.weight',        lambda r:   tf_params[f'synthesis/{r}x{r}/Conv1/mod_weight'].transpose(),
+        r'synthesis\.b(\d+)\.conv1\.affine\.bias',          lambda r:   tf_params[f'synthesis/{r}x{r}/Conv1/mod_bias'] + 1,
+        r'synthesis\.b(\d+)\.torgb\.weight',                lambda r:   tf_params[f'synthesis/{r}x{r}/ToRGB/weight'].transpose(3, 2, 0, 1),
+        r'synthesis\.b(\d+)\.torgb\.bias',                  lambda r:   tf_params[f'synthesis/{r}x{r}/ToRGB/bias'],
+        r'synthesis\.b(\d+)\.torgb\.affine\.weight',        lambda r:   tf_params[f'synthesis/{r}x{r}/ToRGB/mod_weight'].transpose(),
+        r'synthesis\.b(\d+)\.torgb\.affine\.bias',          lambda r:   tf_params[f'synthesis/{r}x{r}/ToRGB/mod_bias'] + 1,
+        r'synthesis\.b(\d+)\.skip\.weight',                 lambda r:   tf_params[f'synthesis/{r}x{r}/Skip/weight'][::-1, ::-1].transpose(3, 2, 0, 1),
+        r'.*\.resample_filter',                             None,
+    )
+    return G
+
+#----------------------------------------------------------------------------
+
+def convert_tf_discriminator(tf_D):
+    if tf_D.version < 4:
+        raise ValueError('TensorFlow pickle version too low')
+
+    # Collect kwargs.
+    tf_kwargs = tf_D.static_kwargs
+    known_kwargs = set()
+    def kwarg(tf_name, default=None):
+        known_kwargs.add(tf_name)
+        return tf_kwargs.get(tf_name, default)
+
+    # Convert kwargs.
+    kwargs = dnnlib.EasyDict(
+        c_dim                   = kwarg('label_size',           0),
+        img_resolution          = kwarg('resolution',           1024),
+        img_channels            = kwarg('num_channels',         3),
+        architecture            = kwarg('architecture',         'resnet'),
+        channel_base            = kwarg('fmap_base',            16384) * 2,
+        channel_max             = kwarg('fmap_max',             512),
+        num_fp16_res            = kwarg('num_fp16_res',         0),
+        conv_clamp              = kwarg('conv_clamp',           None),
+        cmap_dim                = kwarg('mapping_fmaps',        None),
+        block_kwargs = dnnlib.EasyDict(
+            activation          = kwarg('nonlinearity',         'lrelu'),
+            resample_filter     = kwarg('resample_kernel',      [1,3,3,1]),
+            freeze_layers       = kwarg('freeze_layers',        0),
+        ),
+        mapping_kwargs = dnnlib.EasyDict(
+            num_layers          = kwarg('mapping_layers',       0),
+            embed_features      = kwarg('mapping_fmaps',        None),
+            layer_features      = kwarg('mapping_fmaps',        None),
+            activation          = kwarg('nonlinearity',         'lrelu'),
+            lr_multiplier       = kwarg('mapping_lrmul',        0.1),
+        ),
+        epilogue_kwargs = dnnlib.EasyDict(
+            mbstd_group_size    = kwarg('mbstd_group_size',     None),
+            mbstd_num_channels  = kwarg('mbstd_num_features',   1),
+            activation          = kwarg('nonlinearity',         'lrelu'),
+        ),
+    )
+
+    # Check for unknown kwargs.
+    kwarg('structure')
+    unknown_kwargs = list(set(tf_kwargs.keys()) - known_kwargs)
+    if len(unknown_kwargs) > 0:
+        raise ValueError('Unknown TensorFlow kwarg', unknown_kwargs[0])
+
+    # Collect params.
+    tf_params = _collect_tf_params(tf_D)
+    for name, value in list(tf_params.items()):
+        match = re.fullmatch(r'FromRGB_lod(\d+)/(.*)', name)
+        if match:
+            r = kwargs.img_resolution // (2 ** int(match.group(1)))
+            tf_params[f'{r}x{r}/FromRGB/{match.group(2)}'] = value
+            kwargs.architecture = 'orig'
+    #for name, value in tf_params.items(): print(f'{name:<50s}{list(value.shape)}')
+
+    # Convert params.
+    from training import networks
+    D = networks.Discriminator(**kwargs).eval().requires_grad_(False)
+    # pylint: disable=unnecessary-lambda
+    _populate_module_params(D,
+        r'b(\d+)\.fromrgb\.weight',     lambda r:       tf_params[f'{r}x{r}/FromRGB/weight'].transpose(3, 2, 0, 1),
+        r'b(\d+)\.fromrgb\.bias',       lambda r:       tf_params[f'{r}x{r}/FromRGB/bias'],
+        r'b(\d+)\.conv(\d+)\.weight',   lambda r, i:    tf_params[f'{r}x{r}/Conv{i}{["","_down"][int(i)]}/weight'].transpose(3, 2, 0, 1),
+        r'b(\d+)\.conv(\d+)\.bias',     lambda r, i:    tf_params[f'{r}x{r}/Conv{i}{["","_down"][int(i)]}/bias'],
+        r'b(\d+)\.skip\.weight',        lambda r:       tf_params[f'{r}x{r}/Skip/weight'].transpose(3, 2, 0, 1),
+        r'mapping\.embed\.weight',      lambda:         tf_params[f'LabelEmbed/weight'].transpose(),
+        r'mapping\.embed\.bias',        lambda:         tf_params[f'LabelEmbed/bias'],
+        r'mapping\.fc(\d+)\.weight',    lambda i:       tf_params[f'Mapping{i}/weight'].transpose(),
+        r'mapping\.fc(\d+)\.bias',      lambda i:       tf_params[f'Mapping{i}/bias'],
+        r'b4\.conv\.weight',            lambda:         tf_params[f'4x4/Conv/weight'].transpose(3, 2, 0, 1),
+        r'b4\.conv\.bias',              lambda:         tf_params[f'4x4/Conv/bias'],
+        r'b4\.fc\.weight',              lambda:         tf_params[f'4x4/Dense0/weight'].transpose(),
+        r'b4\.fc\.bias',                lambda:         tf_params[f'4x4/Dense0/bias'],
+        r'b4\.out\.weight',             lambda:         tf_params[f'Output/weight'].transpose(),
+        r'b4\.out\.bias',               lambda:         tf_params[f'Output/bias'],
+        r'.*\.resample_filter',         None,
+    )
+    return D
+
+#----------------------------------------------------------------------------
+
+@click.command()
+@click.option('--source', help='Input pickle', required=True, metavar='PATH')
+@click.option('--dest', help='Output pickle', required=True, metavar='PATH')
+@click.option('--force-fp16', help='Force the networks to use FP16', type=bool, default=False, metavar='BOOL', show_default=True)
+def convert_network_pickle(source, dest, force_fp16):
+    """Convert legacy network pickle into the native PyTorch format.
+
+    The tool is able to load the main network configurations exported using the TensorFlow version of StyleGAN2 or StyleGAN2-ADA.
+    It does not support e.g. StyleGAN2-ADA comparison methods, StyleGAN2 configs A-D, or StyleGAN1 networks.
+
+    Example:
+
+    \b
+    python legacy.py \\
+        --source=https://nvlabs-fi-cdn.nvidia.com/stylegan2/networks/stylegan2-cat-config-f.pkl \\
+        --dest=stylegan2-cat-config-f.pkl
+    """
+    print(f'Loading "{source}"...')
+    with dnnlib.util.open_url(source) as f:
+        data = load_network_pkl(f, force_fp16=force_fp16)
+    print(f'Saving "{dest}"...')
+    with open(dest, 'wb') as f:
+        pickle.dump(data, f)
+    print('Done.')
+
+#----------------------------------------------------------------------------
+
+if __name__ == "__main__":
+    convert_network_pickle() # pylint: disable=no-value-for-parameter
+
+#----------------------------------------------------------------------------

losses/loss.py

@@ -0,0 +1,170 @@
+# Copyright (c) 2021, NVIDIA CORPORATION.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+import numpy as np
+import torch
+from torch_utils import training_stats
+from torch_utils import misc
+from torch_utils.ops import conv2d_gradfix
+from losses.pcp import PerceptualLoss
+
+#----------------------------------------------------------------------------
+
+class Loss:
+    def accumulate_gradients(self, phase, real_img, real_c, gen_z, gen_c, sync, gain): # to be overridden by subclass
+        raise NotImplementedError()
+
+#----------------------------------------------------------------------------
+
+class TwoStageLoss(Loss):
+    def __init__(self, device, G_mapping, G_synthesis, D, augment_pipe=None, style_mixing_prob=0.9, r1_gamma=10, pl_batch_shrink=2, pl_decay=0.01, pl_weight=2, truncation_psi=1, pcp_ratio=1.0):
+        super().__init__()
+        self.device = device
+        self.G_mapping = G_mapping
+        self.G_synthesis = G_synthesis
+        self.D = D
+        self.augment_pipe = augment_pipe
+        self.style_mixing_prob = style_mixing_prob
+        self.r1_gamma = r1_gamma
+        self.pl_batch_shrink = pl_batch_shrink
+        self.pl_decay = pl_decay
+        self.pl_weight = pl_weight
+        self.pl_mean = torch.zeros([], device=device)
+        self.truncation_psi = truncation_psi
+        self.pcp = PerceptualLoss(layer_weights=dict(conv4_4=1/4, conv5_4=1/2)).to(device)
+        self.pcp_ratio = pcp_ratio
+
+    def run_G(self, img_in, mask_in, z, c, sync):
+        with misc.ddp_sync(self.G_mapping, sync):
+            ws = self.G_mapping(z, c, truncation_psi=self.truncation_psi)
+            if self.style_mixing_prob > 0:
+                with torch.autograd.profiler.record_function('style_mixing'):
+                    cutoff = torch.empty([], dtype=torch.int64, device=ws.device).random_(1, ws.shape[1])
+                    cutoff = torch.where(torch.rand([], device=ws.device) < self.style_mixing_prob, cutoff, torch.full_like(cutoff, ws.shape[1]))
+                    ws[:, cutoff:] = self.G_mapping(torch.randn_like(z), c, truncation_psi=self.truncation_psi, skip_w_avg_update=True)[:, cutoff:]
+        with misc.ddp_sync(self.G_synthesis, sync):
+            img, img_stg1 = self.G_synthesis(img_in, mask_in, ws, return_stg1=True)
+        return img, ws, img_stg1
+
+    def run_D(self, img, mask, img_stg1, c, sync):
+        # if self.augment_pipe is not None:
+        #     # img = self.augment_pipe(img)
+        #     # !!!!! have to remove the color transform
+        #     tmp_img = torch.cat([img, mask], dim=1)
+        #     tmp_img = self.augment_pipe(tmp_img)
+        #     img, mask = torch.split(tmp_img, [3, 1])
+        with misc.ddp_sync(self.D, sync):
+            logits, logits_stg1 = self.D(img, mask, img_stg1, c)
+        return logits, logits_stg1
+
+    def accumulate_gradients(self, phase, real_img, mask, real_c, gen_z, gen_c, sync, gain):
+        assert phase in ['Gmain', 'Greg', 'Gboth', 'Dmain', 'Dreg', 'Dboth']
+        do_Gmain = (phase in ['Gmain', 'Gboth'])
+        do_Dmain = (phase in ['Dmain', 'Dboth'])
+        do_Gpl   = (phase in ['Greg', 'Gboth']) and (self.pl_weight != 0)
+        do_Dr1   = (phase in ['Dreg', 'Dboth']) and (self.r1_gamma != 0)
+
+        # Gmain: Maximize logits for generated images.
+        if do_Gmain:
+            with torch.autograd.profiler.record_function('Gmain_forward'):
+                gen_img, _gen_ws, gen_img_stg1 = self.run_G(real_img, mask, gen_z, gen_c, sync=(sync and not do_Gpl)) # May get synced by Gpl.
+                gen_logits, gen_logits_stg1 = self.run_D(gen_img, mask, gen_img_stg1, gen_c, sync=False)
+                training_stats.report('Loss/scores/fake', gen_logits)
+                training_stats.report('Loss/signs/fake', gen_logits.sign())
+                training_stats.report('Loss/scores/fake_s1', gen_logits_stg1)
+                training_stats.report('Loss/signs/fake_s1', gen_logits_stg1.sign())
+                loss_Gmain = torch.nn.functional.softplus(-gen_logits) # -log(sigmoid(gen_logits))
+                training_stats.report('Loss/G/loss', loss_Gmain)
+                loss_Gmain_stg1 = torch.nn.functional.softplus(-gen_logits_stg1)
+                training_stats.report('Loss/G/loss_s1', loss_Gmain_stg1)
+                # just for showing
+                l1_loss = torch.mean(torch.abs(gen_img - real_img))
+                training_stats.report('Loss/G/l1_loss', l1_loss)
+                pcp_loss, _ = self.pcp(gen_img, real_img)
+                training_stats.report('Loss/G/pcp_loss', pcp_loss)
+            with torch.autograd.profiler.record_function('Gmain_backward'):
+                loss_Gmain_all = loss_Gmain + loss_Gmain_stg1 + pcp_loss * self.pcp_ratio
+                loss_Gmain_all.mean().mul(gain).backward()
+
+        # # Gpl: Apply path length regularization.
+        # if do_Gpl:
+        #     with torch.autograd.profiler.record_function('Gpl_forward'):
+        #         batch_size = gen_z.shape[0] // self.pl_batch_shrink
+        #         gen_img, gen_ws = self.run_G(real_img[:batch_size], mask[:batch_size], gen_z[:batch_size], gen_c[:batch_size], sync=sync)
+        #         pl_noise = torch.randn_like(gen_img) / np.sqrt(gen_img.shape[2] * gen_img.shape[3])
+        #         with torch.autograd.profiler.record_function('pl_grads'), conv2d_gradfix.no_weight_gradients():
+        #             pl_grads = torch.autograd.grad(outputs=[(gen_img * pl_noise).sum()], inputs=[gen_ws], create_graph=True, only_inputs=True)[0]
+        #         pl_lengths = pl_grads.square().sum(2).mean(1).sqrt()
+        #         pl_mean = self.pl_mean.lerp(pl_lengths.mean(), self.pl_decay)
+        #         self.pl_mean.copy_(pl_mean.detach())
+        #         pl_penalty = (pl_lengths - pl_mean).square()
+        #         training_stats.report('Loss/pl_penalty', pl_penalty)
+        #         loss_Gpl = pl_penalty * self.pl_weight
+        #         training_stats.report('Loss/G/reg', loss_Gpl)
+        #     with torch.autograd.profiler.record_function('Gpl_backward'):
+        #         (gen_img[:, 0, 0, 0] * 0 + loss_Gpl).mean().mul(gain).backward()
+
+        # Dmain: Minimize logits for generated images.
+        loss_Dgen = 0
+        loss_Dgen_stg1 = 0
+        if do_Dmain:
+            with torch.autograd.profiler.record_function('Dgen_forward'):
+                gen_img, _gen_ws, gen_img_stg1 = self.run_G(real_img, mask, gen_z, gen_c, sync=False)
+                gen_logits, gen_logits_stg1 = self.run_D(gen_img, mask, gen_img_stg1, gen_c, sync=False) # Gets synced by loss_Dreal.
+                training_stats.report('Loss/scores/fake', gen_logits)
+                training_stats.report('Loss/signs/fake', gen_logits.sign())
+                loss_Dgen = torch.nn.functional.softplus(gen_logits) # -log(1 - sigmoid(gen_logits))
+                training_stats.report('Loss/scores/fake_s1', gen_logits_stg1)
+                training_stats.report('Loss/signs/fake_s1', gen_logits_stg1.sign())
+                loss_Dgen_stg1 = torch.nn.functional.softplus(gen_logits_stg1)  # -log(1 - sigmoid(gen_logits))
+            with torch.autograd.profiler.record_function('Dgen_backward'):
+                loss_Dgen_all = loss_Dgen + loss_Dgen_stg1
+                loss_Dgen_all.mean().mul(gain).backward()
+
+        # Dmain: Maximize logits for real images.
+        # Dr1: Apply R1 regularization.
+        if do_Dmain or do_Dr1:
+            name = 'Dreal_Dr1' if do_Dmain and do_Dr1 else 'Dreal' if do_Dmain else 'Dr1'
+            with torch.autograd.profiler.record_function(name + '_forward'):
+                real_img_tmp = real_img.detach().requires_grad_(do_Dr1)
+                mask_tmp = mask.detach().requires_grad_(do_Dr1)
+                real_img_tmp_stg1 = real_img.detach().requires_grad_(do_Dr1)
+                real_logits, real_logits_stg1 = self.run_D(real_img_tmp, mask_tmp, real_img_tmp_stg1, real_c, sync=sync)
+                training_stats.report('Loss/scores/real', real_logits)
+                training_stats.report('Loss/signs/real', real_logits.sign())
+                training_stats.report('Loss/scores/real_s1', real_logits_stg1)
+                training_stats.report('Loss/signs/real_s1', real_logits_stg1.sign())
+
+                loss_Dreal = 0
+                loss_Dreal_stg1 = 0
+                if do_Dmain:
+                    loss_Dreal = torch.nn.functional.softplus(-real_logits) # -log(sigmoid(real_logits))
+                    loss_Dreal_stg1 = torch.nn.functional.softplus(-real_logits_stg1)  # -log(sigmoid(real_logits))
+                    training_stats.report('Loss/D/loss', loss_Dgen + loss_Dreal)
+                    training_stats.report('Loss/D/loss_s1', loss_Dgen_stg1 + loss_Dreal_stg1)
+
+                loss_Dr1 = 0
+                loss_Dr1_stg1 = 0
+                if do_Dr1:
+                    with torch.autograd.profiler.record_function('r1_grads'), conv2d_gradfix.no_weight_gradients():
+                        r1_grads = torch.autograd.grad(outputs=[real_logits.sum()], inputs=[real_img_tmp], create_graph=True, only_inputs=True)[0]
+                        r1_grads_stg1 = torch.autograd.grad(outputs=[real_logits_stg1.sum()], inputs=[real_img_tmp_stg1], create_graph=True, only_inputs=True)[0]
+                    r1_penalty = r1_grads.square().sum([1,2,3])
+                    loss_Dr1 = r1_penalty * (self.r1_gamma / 2)
+                    training_stats.report('Loss/r1_penalty', r1_penalty)
+                    training_stats.report('Loss/D/reg', loss_Dr1)
+
+                    r1_penalty_stg1 = r1_grads_stg1.square().sum([1, 2, 3])
+                    loss_Dr1_stg1 = r1_penalty_stg1 * (self.r1_gamma / 2)
+                    training_stats.report('Loss/r1_penalty_s1', r1_penalty_stg1)
+                    training_stats.report('Loss/D/reg_s1', loss_Dr1_stg1)
+
+            with torch.autograd.profiler.record_function(name + '_backward'):
+                ((real_logits + real_logits_stg1) * 0 + loss_Dreal + loss_Dreal_stg1 + loss_Dr1 + loss_Dr1_stg1).mean().mul(gain).backward()
+
+#----------------------------------------------------------------------------

losses/pcp.py

@@ -0,0 +1,126 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from losses.vggNet import VGGFeatureExtractor
+import numpy as np
+
+
+class PerceptualLoss(nn.Module):
+    """Perceptual loss with commonly used style loss.
+
+    Args:
+        layer_weights (dict): The weight for each layer of vgg feature.
+            Here is an example: {'conv5_4': 1.}, which means the conv5_4
+            feature layer (before relu5_4) will be extracted with weight
+            1.0 in calculting losses.
+        vgg_type (str): The type of vgg network used as feature extractor.
+            Default: 'vgg19'.
+        use_input_norm (bool):  If True, normalize the input image in vgg.
+            Default: True.
+        perceptual_weight (float): If `perceptual_weight > 0`, the perceptual
+            loss will be calculated and the loss will multiplied by the
+            weight. Default: 1.0.
+        style_weight (float): If `style_weight > 0`, the style loss will be
+            calculated and the loss will multiplied by the weight.
+            Default: 0.
+        norm_img (bool): If True, the image will be normed to [0, 1]. Note that
+            this is different from the `use_input_norm` which norm the input in
+            in forward function of vgg according to the statistics of dataset.
+            Importantly, the input image must be in range [-1, 1].
+            Default: False.
+        criterion (str): Criterion used for perceptual loss. Default: 'l1'.
+    """
+
+    def __init__(self,
+                 layer_weights,
+                 vgg_type='vgg19',
+                 use_input_norm=True,
+                 use_pcp_loss=True,
+                 use_style_loss=False,
+                 norm_img=True,
+                 criterion='l1'):
+        super(PerceptualLoss, self).__init__()
+        self.norm_img = norm_img
+        self.use_pcp_loss = use_pcp_loss
+        self.use_style_loss = use_style_loss
+        self.layer_weights = layer_weights
+        self.vgg = VGGFeatureExtractor(
+            layer_name_list=list(layer_weights.keys()),
+            vgg_type=vgg_type,
+            use_input_norm=use_input_norm)
+
+        self.criterion_type = criterion
+        if self.criterion_type == 'l1':
+            self.criterion = torch.nn.L1Loss()
+        elif self.criterion_type == 'l2':
+            self.criterion = torch.nn.L2loss()
+        elif self.criterion_type == 'fro':
+            self.criterion = None
+        else:
+            raise NotImplementedError('%s criterion has not been supported.' % self.criterion_type)
+
+    def forward(self, x, gt):
+        """Forward function.
+
+        Args:
+            x (Tensor): Input tensor with shape (n, c, h, w).
+            gt (Tensor): Ground-truth tensor with shape (n, c, h, w).
+
+        Returns:
+            Tensor: Forward results.
+        """
+
+        if self.norm_img:
+            x = (x + 1.) * 0.5
+            gt = (gt + 1.) * 0.5
+
+        # extract vgg features
+        x_features = self.vgg(x)
+        gt_features = self.vgg(gt.detach())
+
+        # calculate perceptual loss
+        if self.use_pcp_loss:
+            percep_loss = 0
+            for k in x_features.keys():
+                if self.criterion_type == 'fro':
+                    percep_loss += torch.norm(
+                        x_features[k] - gt_features[k],
+                        p='fro') * self.layer_weights[k]
+                else:
+                    percep_loss += self.criterion(x_features[k], gt_features[k]) * self.layer_weights[k]
+        else:
+            percep_loss = None
+
+        # calculate style loss
+        if self.use_style_loss:
+            style_loss = 0
+            for k in x_features.keys():
+                if self.criterion_type == 'fro':
+                    style_loss += torch.norm(
+                        self._gram_mat(x_features[k]) -
+                        self._gram_mat(gt_features[k]),
+                        p='fro') * self.layer_weights[k]
+                else:
+                    style_loss += self.criterion(self._gram_mat(x_features[k]), self._gram_mat(gt_features[k])) \
+                                  * self.layer_weights[k]
+        else:
+            style_loss = None
+
+        return percep_loss, style_loss
+
+    def _gram_mat(self, x):
+        """Calculate Gram matrix.
+
+        Args:
+            x (torch.Tensor): Tensor with shape of (n, c, h, w).
+
+        Returns:
+            torch.Tensor: Gram matrix.
+        """
+        n, c, h, w = x.size()
+        features = x.view(n, c, w * h)
+        features_t = features.transpose(1, 2)
+        gram = features.bmm(features_t) / (c * h * w)
+        return gram
+

losses/vggNet.py

@@ -0,0 +1,178 @@
+from collections import OrderedDict
+
+import torch
+import torch.nn as nn
+from torch.nn import DataParallel
+from torch.nn.parallel import DistributedDataParallel
+from torchvision.models import vgg as vgg
+
+
+NAMES = {
+    'vgg11': [
+        'conv1_1', 'relu1_1', 'pool1', 'conv2_1', 'relu2_1', 'pool2',
+        'conv3_1', 'relu3_1', 'conv3_2', 'relu3_2', 'pool3', 'conv4_1',
+        'relu4_1', 'conv4_2', 'relu4_2', 'pool4', 'conv5_1', 'relu5_1',
+        'conv5_2', 'relu5_2', 'pool5'
+    ],
+    'vgg13': [
+        'conv1_1', 'relu1_1', 'conv1_2', 'relu1_2', 'pool1', 'conv2_1',
+        'relu2_1', 'conv2_2', 'relu2_2', 'pool2', 'conv3_1', 'relu3_1',
+        'conv3_2', 'relu3_2', 'pool3', 'conv4_1', 'relu4_1', 'conv4_2',
+        'relu4_2', 'pool4', 'conv5_1', 'relu5_1', 'conv5_2', 'relu5_2', 'pool5'
+    ],
+    'vgg16': [
+        'conv1_1', 'relu1_1', 'conv1_2', 'relu1_2', 'pool1', 'conv2_1',
+        'relu2_1', 'conv2_2', 'relu2_2', 'pool2', 'conv3_1', 'relu3_1',
+        'conv3_2', 'relu3_2', 'conv3_3', 'relu3_3', 'pool3', 'conv4_1',
+        'relu4_1', 'conv4_2', 'relu4_2', 'conv4_3', 'relu4_3', 'pool4',
+        'conv5_1', 'relu5_1', 'conv5_2', 'relu5_2', 'conv5_3', 'relu5_3',
+        'pool5'
+    ],
+    'vgg19': [
+        'conv1_1', 'relu1_1', 'conv1_2', 'relu1_2', 'pool1', 'conv2_1',
+        'relu2_1', 'conv2_2', 'relu2_2', 'pool2', 'conv3_1', 'relu3_1',
+        'conv3_2', 'relu3_2', 'conv3_3', 'relu3_3', 'conv3_4', 'relu3_4',
+        'pool3', 'conv4_1', 'relu4_1', 'conv4_2', 'relu4_2', 'conv4_3',
+        'relu4_3', 'conv4_4', 'relu4_4', 'pool4', 'conv5_1', 'relu5_1',
+        'conv5_2', 'relu5_2', 'conv5_3', 'relu5_3', 'conv5_4', 'relu5_4',
+        'pool5'
+    ]
+}
+
+
+# MODEL_PATH = {
+#     'vgg19': 'losses/pretrained/vgg19-dcbb9e9d.pth'
+# }
+
+
+def load_model(model, model_path, strict=True, cpu=False):
+    if isinstance(model, DataParallel) or isinstance(model, DistributedDataParallel):
+        model = model.module
+    if cpu:
+        loaded_model = torch.load(model_path, map_location='cpu')
+    else:
+        loaded_model = torch.load(model_path)
+    model.load_state_dict(loaded_model, strict=strict)
+
+
+def insert_bn(names):
+    """Insert bn layer after each conv.
+
+    Args:
+        names (list): The list of layer names.
+
+    Returns:
+        list: The list of layer names with bn layers.
+    """
+    names_bn = []
+    for name in names:
+        names_bn.append(name)
+        if 'conv' in name:
+            position = name.replace('conv', '')
+            names_bn.append('bn' + position)
+    return names_bn
+
+
+class VGGFeatureExtractor(nn.Module):
+    """VGG network for feature extraction.
+
+    In this implementation, we allow users to choose whether use normalization
+    in the input feature and the type of vgg network. Note that the pretrained
+    path must fit the vgg type.
+
+    Args:
+        layer_name_list (list[str]): Forward function returns the corresponding
+            features according to the layer_name_list.
+            Example: {'relu1_1', 'relu2_1', 'relu3_1'}.
+        vgg_type (str): Set the type of vgg network. Default: 'vgg19'.
+        use_input_norm (bool): If True, normalize the input image. Importantly,
+            the input feature must in the range [0, 1]. Default: True.
+        requires_grad (bool): If true, the parameters of VGG network will be
+            optimized. Default: False.
+        remove_pooling (bool): If true, the max pooling operations in VGG net
+            will be removed. Default: False.
+        pooling_stride (int): The stride of max pooling operation. Default: 2.
+    """
+
+    def __init__(self,
+                 layer_name_list,
+                 vgg_type='vgg19',
+                 use_input_norm=True,
+                 requires_grad=False,
+                 remove_pooling=False,
+                 pooling_stride=2):
+        super(VGGFeatureExtractor, self).__init__()
+
+        self.layer_name_list = layer_name_list
+        self.use_input_norm = use_input_norm
+
+        self.names = NAMES[vgg_type.replace('_bn', '')]
+        if 'bn' in vgg_type:
+            self.names = insert_bn(self.names)
+
+        # only borrow layers that will be used to avoid unused params
+        max_idx = 0
+        for v in layer_name_list:
+            idx = self.names.index(v)
+            if idx > max_idx:
+                max_idx = idx
+
+        features = getattr(vgg, vgg_type)(pretrained=True).features[:max_idx + 1]
+        # vgg_model = getattr(vgg, vgg_type)(pretrained=False)
+        # load_model(vgg_model, MODEL_PATH[vgg_type], strict=True)
+        # features = vgg_model.features[:max_idx + 1]
+
+        modified_net = OrderedDict()
+        for k, v in zip(self.names, features):
+            if 'pool' in k:
+                # if remove_pooling is true, pooling operation will be removed
+                if remove_pooling:
+                    continue
+                else:
+                    # in some cases, we may want to change the default stride
+                    modified_net[k] = nn.MaxPool2d(
+                        kernel_size=2, stride=pooling_stride)
+            else:
+                modified_net[k] = v
+
+        self.vgg_net = nn.Sequential(modified_net)
+
+        if not requires_grad:
+            self.vgg_net.eval()
+            for param in self.parameters():
+                param.requires_grad = False
+        else:
+            self.vgg_net.train()
+            for param in self.parameters():
+                param.requires_grad = True
+
+        if self.use_input_norm:
+            # the mean is for image with range [0, 1]
+            self.register_buffer(
+                'mean',
+                torch.Tensor([0.485, 0.456, 0.406]).view(1, 3, 1, 1))
+            # the std is for image with range [0, 1]
+            self.register_buffer(
+                'std',
+                torch.Tensor([0.229, 0.224, 0.225]).view(1, 3, 1, 1))
+
+    def forward(self, x):
+        """Forward function.
+
+        Args:
+            x (Tensor): Input tensor with shape (n, c, h, w).
+
+        Returns:
+            Tensor: Forward results.
+        """
+
+        if self.use_input_norm:
+            x = (x - self.mean) / self.std
+
+        output = {}
+        for key, layer in self.vgg_net._modules.items():
+            x = layer(x)
+            if key in self.layer_name_list:
+                output[key] = x.clone()
+
+        return output

metrics/__init__.py

@@ -0,0 +1,9 @@
+# Copyright (c) 2021, NVIDIA CORPORATION.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+# empty

metrics/frechet_inception_distance.py

@@ -0,0 +1,41 @@
+# Copyright (c) 2021, NVIDIA CORPORATION.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+"""Frechet Inception Distance (FID) from the paper
+"GANs trained by a two time-scale update rule converge to a local Nash
+equilibrium". Matches the original implementation by Heusel et al. at
+https://github.com/bioinf-jku/TTUR/blob/master/fid.py"""
+
+import numpy as np
+import scipy.linalg
+from . import metric_utils
+
+#----------------------------------------------------------------------------
+
+def compute_fid(opts, max_real, num_gen):
+    # Direct TorchScript translation of http://download.tensorflow.org/models/image/imagenet/inception-2015-12-05.tgz
+    detector_url = 'https://nvlabs-fi-cdn.nvidia.com/stylegan2-ada-pytorch/pretrained/metrics/inception-2015-12-05.pt'
+    detector_kwargs = dict(return_features=True) # Return raw features before the softmax layer.
+
+    mu_real, sigma_real = metric_utils.compute_feature_stats_for_dataset(
+        opts=opts, detector_url=detector_url, detector_kwargs=detector_kwargs,
+        rel_lo=0, rel_hi=0, capture_mean_cov=True, max_items=max_real).get_mean_cov()
+
+    mu_gen, sigma_gen = metric_utils.compute_feature_stats_for_generator(
+        opts=opts, detector_url=detector_url, detector_kwargs=detector_kwargs,
+        rel_lo=0, rel_hi=1, capture_mean_cov=True, max_items=num_gen).get_mean_cov()
+
+    if opts.rank != 0:
+        return float('nan')
+
+    m = np.square(mu_gen - mu_real).sum()
+    s, _ = scipy.linalg.sqrtm(np.dot(sigma_gen, sigma_real), disp=False) # pylint: disable=no-member
+    fid = np.real(m + np.trace(sigma_gen + sigma_real - s * 2))
+    return float(fid)
+
+#----------------------------------------------------------------------------

metrics/inception_discriminative_score.py

@@ -0,0 +1,37 @@
+
+import numpy as np
+import scipy.linalg
+from . import metric_utils
+import sklearn.svm
+
+#----------------------------------------------------------------------------
+
+def compute_ids(opts, max_real, num_gen):
+    # Direct TorchScript translation of http://download.tensorflow.org/models/image/imagenet/inception-2015-12-05.tgz
+    detector_url = 'https://nvlabs-fi-cdn.nvidia.com/stylegan2-ada-pytorch/pretrained/metrics/inception-2015-12-05.pt'
+    detector_kwargs = dict(return_features=True) # Return raw features before the softmax layer.
+
+    real_activations = metric_utils.compute_feature_stats_for_dataset(
+        opts=opts, detector_url=detector_url, detector_kwargs=detector_kwargs,
+        rel_lo=0, rel_hi=0, capture_all=True, max_items=max_real).get_all()
+
+    fake_activations = metric_utils.compute_feature_stats_for_generator(
+        opts=opts, detector_url=detector_url, detector_kwargs=detector_kwargs,
+        rel_lo=0, rel_hi=1, capture_all=True, max_items=num_gen).get_all()
+
+    if opts.rank != 0:
+        return float('nan')
+
+    svm = sklearn.svm.LinearSVC(dual=False)
+    svm_inputs = np.concatenate([real_activations, fake_activations])
+    svm_targets = np.array([1] * real_activations.shape[0] + [0] * fake_activations.shape[0])
+    print('Fitting ...')
+    svm.fit(svm_inputs, svm_targets)
+    u_ids = 1 - svm.score(svm_inputs, svm_targets)
+    real_outputs = svm.decision_function(real_activations)
+    fake_outputs = svm.decision_function(fake_activations)
+    p_ids = np.mean(fake_outputs > real_outputs)
+
+    return float(u_ids), float(p_ids)
+
+#----------------------------------------------------------------------------

metrics/inception_score.py

@@ -0,0 +1,38 @@
+# Copyright (c) 2021, NVIDIA CORPORATION.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+"""Inception Score (IS) from the paper "Improved techniques for training
+GANs". Matches the original implementation by Salimans et al. at
+https://github.com/openai/improved-gan/blob/master/inception_score/model.py"""
+
+import numpy as np
+from . import metric_utils
+
+#----------------------------------------------------------------------------
+
+def compute_is(opts, num_gen, num_splits):
+    # Direct TorchScript translation of http://download.tensorflow.org/models/image/imagenet/inception-2015-12-05.tgz
+    detector_url = 'https://nvlabs-fi-cdn.nvidia.com/stylegan2-ada-pytorch/pretrained/metrics/inception-2015-12-05.pt'
+    detector_kwargs = dict(no_output_bias=True) # Match the original implementation by not applying bias in the softmax layer.
+
+    gen_probs = metric_utils.compute_feature_stats_for_generator(
+        opts=opts, detector_url=detector_url, detector_kwargs=detector_kwargs,
+        capture_all=True, max_items=num_gen).get_all()
+
+    if opts.rank != 0:
+        return float('nan'), float('nan')
+
+    scores = []
+    for i in range(num_splits):
+        part = gen_probs[i * num_gen // num_splits : (i + 1) * num_gen // num_splits]
+        kl = part * (np.log(part) - np.log(np.mean(part, axis=0, keepdims=True)))
+        kl = np.mean(np.sum(kl, axis=1))
+        scores.append(np.exp(kl))
+    return float(np.mean(scores)), float(np.std(scores))
+
+#----------------------------------------------------------------------------

metrics/kernel_inception_distance.py

@@ -0,0 +1,46 @@
+# Copyright (c) 2021, NVIDIA CORPORATION.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+"""Kernel Inception Distance (KID) from the paper "Demystifying MMD
+GANs". Matches the original implementation by Binkowski et al. at
+https://github.com/mbinkowski/MMD-GAN/blob/master/gan/compute_scores.py"""
+
+import numpy as np
+from . import metric_utils
+
+#----------------------------------------------------------------------------
+
+def compute_kid(opts, max_real, num_gen, num_subsets, max_subset_size):
+    # Direct TorchScript translation of http://download.tensorflow.org/models/image/imagenet/inception-2015-12-05.tgz
+    detector_url = 'https://nvlabs-fi-cdn.nvidia.com/stylegan2-ada-pytorch/pretrained/metrics/inception-2015-12-05.pt'
+    detector_kwargs = dict(return_features=True) # Return raw features before the softmax layer.
+
+    real_features = metric_utils.compute_feature_stats_for_dataset(
+        opts=opts, detector_url=detector_url, detector_kwargs=detector_kwargs,
+        rel_lo=0, rel_hi=0, capture_all=True, max_items=max_real).get_all()
+
+    gen_features = metric_utils.compute_feature_stats_for_generator(
+        opts=opts, detector_url=detector_url, detector_kwargs=detector_kwargs,
+        rel_lo=0, rel_hi=1, capture_all=True, max_items=num_gen).get_all()
+
+    if opts.rank != 0:
+        return float('nan')
+
+    n = real_features.shape[1]
+    m = min(min(real_features.shape[0], gen_features.shape[0]), max_subset_size)
+    t = 0
+    for _subset_idx in range(num_subsets):
+        x = gen_features[np.random.choice(gen_features.shape[0], m, replace=False)]
+        y = real_features[np.random.choice(real_features.shape[0], m, replace=False)]
+        a = (x @ x.T / n + 1) ** 3 + (y @ y.T / n + 1) ** 3
+        b = (x @ y.T / n + 1) ** 3
+        t += (a.sum() - np.diag(a).sum()) / (m - 1) - b.sum() * 2 / m
+    kid = t / num_subsets / m
+    return float(kid)
+
+#----------------------------------------------------------------------------

metrics/metric_main.py

@@ -0,0 +1,184 @@
+# Copyright (c) 2021, NVIDIA CORPORATION.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+import os
+import time
+import json
+import torch
+import dnnlib
+
+from . import metric_utils
+from . import frechet_inception_distance
+from . import kernel_inception_distance
+from . import precision_recall
+from . import perceptual_path_length
+from . import inception_score
+from . import psnr_ssim_l1
+from . import inception_discriminative_score
+
+#----------------------------------------------------------------------------
+
+_metric_dict = dict() # name => fn
+
+def register_metric(fn):
+    assert callable(fn)
+    _metric_dict[fn.__name__] = fn
+    return fn
+
+def is_valid_metric(metric):
+    return metric in _metric_dict
+
+def list_valid_metrics():
+    return list(_metric_dict.keys())
+
+#----------------------------------------------------------------------------
+
+def calc_metric(metric, **kwargs): # See metric_utils.MetricOptions for the full list of arguments.
+    assert is_valid_metric(metric)
+    opts = metric_utils.MetricOptions(**kwargs)
+
+    # Calculate.
+    start_time = time.time()
+    results = _metric_dict[metric](opts)
+    total_time = time.time() - start_time
+
+    # Broadcast results.
+    for key, value in list(results.items()):
+        if opts.num_gpus > 1:
+            value = torch.as_tensor(value, dtype=torch.float64, device=opts.device)
+            torch.distributed.broadcast(tensor=value, src=0)
+            value = float(value.cpu())
+        results[key] = value
+
+    # Decorate with metadata.
+    return dnnlib.EasyDict(
+        results         = dnnlib.EasyDict(results),
+        metric          = metric,
+        total_time      = total_time,
+        total_time_str  = dnnlib.util.format_time(total_time),
+        num_gpus        = opts.num_gpus,
+    )
+
+#----------------------------------------------------------------------------
+
+def report_metric(result_dict, run_dir=None, snapshot_pkl=None):
+    metric = result_dict['metric']
+    assert is_valid_metric(metric)
+    if run_dir is not None and snapshot_pkl is not None:
+        snapshot_pkl = os.path.relpath(snapshot_pkl, run_dir)
+
+    jsonl_line = json.dumps(dict(result_dict, snapshot_pkl=snapshot_pkl, timestamp=time.time()))
+    print(jsonl_line)
+    if run_dir is not None and os.path.isdir(run_dir):
+        with open(os.path.join(run_dir, f'metric-{metric}.jsonl'), 'at') as f:
+            f.write(jsonl_line + '\n')
+
+#----------------------------------------------------------------------------
+# Primary metrics.
+
+@register_metric
+def fid2993_full(opts):
+    opts.dataset_kwargs.update(max_size=None, xflip=False)
+    fid = frechet_inception_distance.compute_fid(opts, max_real=2993, num_gen=2993)
+    return dict(fid2993_full=fid)
+
+@register_metric
+def fid36k5_full(opts):
+    opts.dataset_kwargs.update(max_size=None, xflip=False)
+    fid = frechet_inception_distance.compute_fid(opts, max_real=36500, num_gen=36500)
+    return dict(fid36k5_full=fid)
+
+@register_metric
+def fid_places(opts):
+    opts.dataset_kwargs.update(max_size=None, xflip=False)
+    fid = frechet_inception_distance.compute_fid(opts, max_real=36500, num_gen=36500)
+    return dict(fid36k5_full=fid)
+
+@register_metric
+def ids_places(opts):
+    opts.dataset_kwargs.update(max_size=None, xflip=False)
+    u_ids, p_ids = inception_discriminative_score.compute_ids(opts, max_real=36500, num_gen=36500)
+    return dict(u_ids=u_ids, p_ids=p_ids)
+
+@register_metric
+def psnr36k5_full(opts):
+    opts.dataset_kwargs.update(max_size=None, xflip=False)
+    psnr, ssim, l1 = psnr_ssim_l1.compute_psnr(opts, max_real=36500)
+    return dict(psnr=psnr, ssim=ssim, l1=l1)
+
+@register_metric
+def fid50k_full(opts):
+    opts.dataset_kwargs.update(max_size=None, xflip=False)
+    fid = frechet_inception_distance.compute_fid(opts, max_real=None, num_gen=50000)
+    return dict(fid50k_full=fid)
+
+@register_metric
+def kid50k_full(opts):
+    opts.dataset_kwargs.update(max_size=None, xflip=False)
+    kid = kernel_inception_distance.compute_kid(opts, max_real=1000000, num_gen=50000, num_subsets=100, max_subset_size=1000)
+    return dict(kid50k_full=kid)
+
+@register_metric
+def pr50k3_full(opts):
+    opts.dataset_kwargs.update(max_size=None, xflip=False)
+    precision, recall = precision_recall.compute_pr(opts, max_real=200000, num_gen=50000, nhood_size=3, row_batch_size=10000, col_batch_size=10000)
+    return dict(pr50k3_full_precision=precision, pr50k3_full_recall=recall)
+
+@register_metric
+def ppl2_wend(opts):
+    ppl = perceptual_path_length.compute_ppl(opts, num_samples=50000, epsilon=1e-4, space='w', sampling='end', crop=False, batch_size=2)
+    return dict(ppl2_wend=ppl)
+
+@register_metric
+def is50k(opts):
+    opts.dataset_kwargs.update(max_size=None, xflip=False)
+    mean, std = inception_score.compute_is(opts, num_gen=50000, num_splits=10)
+    return dict(is50k_mean=mean, is50k_std=std)
+
+#----------------------------------------------------------------------------
+# Legacy metrics.
+
+@register_metric
+def fid50k(opts):
+    opts.dataset_kwargs.update(max_size=None)
+    fid = frechet_inception_distance.compute_fid(opts, max_real=50000, num_gen=50000)
+    return dict(fid50k=fid)
+
+@register_metric
+def kid50k(opts):
+    opts.dataset_kwargs.update(max_size=None)
+    kid = kernel_inception_distance.compute_kid(opts, max_real=50000, num_gen=50000, num_subsets=100, max_subset_size=1000)
+    return dict(kid50k=kid)
+
+@register_metric
+def pr50k3(opts):
+    opts.dataset_kwargs.update(max_size=None)
+    precision, recall = precision_recall.compute_pr(opts, max_real=50000, num_gen=50000, nhood_size=3, row_batch_size=10000, col_batch_size=10000)
+    return dict(pr50k3_precision=precision, pr50k3_recall=recall)
+
+@register_metric
+def ppl_zfull(opts):
+    ppl = perceptual_path_length.compute_ppl(opts, num_samples=50000, epsilon=1e-4, space='z', sampling='full', crop=True, batch_size=2)
+    return dict(ppl_zfull=ppl)
+
+@register_metric
+def ppl_wfull(opts):
+    ppl = perceptual_path_length.compute_ppl(opts, num_samples=50000, epsilon=1e-4, space='w', sampling='full', crop=True, batch_size=2)
+    return dict(ppl_wfull=ppl)
+
+@register_metric
+def ppl_zend(opts):
+    ppl = perceptual_path_length.compute_ppl(opts, num_samples=50000, epsilon=1e-4, space='z', sampling='end', crop=True, batch_size=2)
+    return dict(ppl_zend=ppl)
+
+@register_metric
+def ppl_wend(opts):
+    ppl = perceptual_path_length.compute_ppl(opts, num_samples=50000, epsilon=1e-4, space='w', sampling='end', crop=True, batch_size=2)
+    return dict(ppl_wend=ppl)
+
+#----------------------------------------------------------------------------

metrics/metric_utils.py

@@ -0,0 +1,434 @@
+# Copyright (c) 2021, NVIDIA CORPORATION.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+import os
+import time
+import hashlib
+import pickle
+import copy
+import uuid
+import numpy as np
+import torch
+import dnnlib
+import math
+import cv2
+
+#----------------------------------------------------------------------------
+
+class MetricOptions:
+    def __init__(self, G=None, G_kwargs={}, dataset_kwargs={}, num_gpus=1, rank=0, device=None, progress=None, cache=True):
+        assert 0 <= rank < num_gpus
+        self.G              = G
+        self.G_kwargs       = dnnlib.EasyDict(G_kwargs)
+        self.dataset_kwargs = dnnlib.EasyDict(dataset_kwargs)
+        self.num_gpus       = num_gpus
+        self.rank           = rank
+        self.device         = device if device is not None else torch.device('cuda', rank)
+        self.progress       = progress.sub() if progress is not None and rank == 0 else ProgressMonitor()
+        self.cache          = cache
+
+#----------------------------------------------------------------------------
+
+_feature_detector_cache = dict()
+
+def get_feature_detector_name(url):
+    return os.path.splitext(url.split('/')[-1])[0]
+
+def get_feature_detector(url, device=torch.device('cpu'), num_gpus=1, rank=0, verbose=False):
+    assert 0 <= rank < num_gpus
+    key = (url, device)
+    if key not in _feature_detector_cache:
+        is_leader = (rank == 0)
+        if not is_leader and num_gpus > 1:
+            torch.distributed.barrier() # leader goes first
+        with dnnlib.util.open_url(url, verbose=(verbose and is_leader)) as f:
+            _feature_detector_cache[key] = torch.jit.load(f).eval().to(device)
+        if is_leader and num_gpus > 1:
+            torch.distributed.barrier() # others follow
+    return _feature_detector_cache[key]
+
+#----------------------------------------------------------------------------
+
+class FeatureStats:
+    def __init__(self, capture_all=False, capture_mean_cov=False, max_items=None):
+        self.capture_all = capture_all
+        self.capture_mean_cov = capture_mean_cov
+        self.max_items = max_items
+        self.num_items = 0
+        self.num_features = None
+        self.all_features = None
+        self.raw_mean = None
+        self.raw_cov = None
+
+    def set_num_features(self, num_features):
+        if self.num_features is not None:
+            assert num_features == self.num_features
+        else:
+            self.num_features = num_features
+            self.all_features = []
+            self.raw_mean = np.zeros([num_features], dtype=np.float64)
+            self.raw_cov = np.zeros([num_features, num_features], dtype=np.float64)
+
+    def is_full(self):
+        return (self.max_items is not None) and (self.num_items >= self.max_items)
+
+    def append(self, x):
+        x = np.asarray(x, dtype=np.float32)
+        assert x.ndim == 2
+        if (self.max_items is not None) and (self.num_items + x.shape[0] > self.max_items):
+            if self.num_items >= self.max_items:
+                return
+            x = x[:self.max_items - self.num_items]
+
+        self.set_num_features(x.shape[1])
+        self.num_items += x.shape[0]
+        if self.capture_all:
+            self.all_features.append(x)
+        if self.capture_mean_cov:
+            x64 = x.astype(np.float64)
+            self.raw_mean += x64.sum(axis=0)
+            self.raw_cov += x64.T @ x64
+
+    def append_torch(self, x, num_gpus=1, rank=0):
+        assert isinstance(x, torch.Tensor) and x.ndim == 2
+        assert 0 <= rank < num_gpus
+        if num_gpus > 1:
+            ys = []
+            for src in range(num_gpus):
+                y = x.clone()
+                torch.distributed.broadcast(y, src=src)
+                ys.append(y)
+            x = torch.stack(ys, dim=1).flatten(0, 1) # interleave samples
+        self.append(x.cpu().numpy())
+
+    def get_all(self):
+        assert self.capture_all
+        return np.concatenate(self.all_features, axis=0)
+
+    def get_all_torch(self):
+        return torch.from_numpy(self.get_all())
+
+    def get_mean_cov(self):
+        assert self.capture_mean_cov
+        mean = self.raw_mean / self.num_items
+        cov = self.raw_cov / self.num_items
+        cov = cov - np.outer(mean, mean)
+        return mean, cov
+
+    def save(self, pkl_file):
+        with open(pkl_file, 'wb') as f:
+            pickle.dump(self.__dict__, f)
+
+    @staticmethod
+    def load(pkl_file):
+        with open(pkl_file, 'rb') as f:
+            s = dnnlib.EasyDict(pickle.load(f))
+        obj = FeatureStats(capture_all=s.capture_all, max_items=s.max_items)
+        obj.__dict__.update(s)
+        return obj
+
+#----------------------------------------------------------------------------
+
+class ProgressMonitor:
+    def __init__(self, tag=None, num_items=None, flush_interval=1000, verbose=False, progress_fn=None, pfn_lo=0, pfn_hi=1000, pfn_total=1000):
+        self.tag = tag
+        self.num_items = num_items
+        self.verbose = verbose
+        self.flush_interval = flush_interval
+        self.progress_fn = progress_fn
+        self.pfn_lo = pfn_lo
+        self.pfn_hi = pfn_hi
+        self.pfn_total = pfn_total
+        self.start_time = time.time()
+        self.batch_time = self.start_time
+        self.batch_items = 0
+        if self.progress_fn is not None:
+            self.progress_fn(self.pfn_lo, self.pfn_total)
+
+    def update(self, cur_items):
+        assert (self.num_items is None) or (cur_items <= self.num_items)
+        if (cur_items < self.batch_items + self.flush_interval) and (self.num_items is None or cur_items < self.num_items):
+            return
+        cur_time = time.time()
+        total_time = cur_time - self.start_time
+        time_per_item = (cur_time - self.batch_time) / max(cur_items - self.batch_items, 1)
+        if (self.verbose) and (self.tag is not None):
+            print(f'{self.tag:<19s} items {cur_items:<7d} time {dnnlib.util.format_time(total_time):<12s} ms/item {time_per_item*1e3:.2f}')
+        self.batch_time = cur_time
+        self.batch_items = cur_items
+
+        if (self.progress_fn is not None) and (self.num_items is not None):
+            self.progress_fn(self.pfn_lo + (self.pfn_hi - self.pfn_lo) * (cur_items / self.num_items), self.pfn_total)
+
+    def sub(self, tag=None, num_items=None, flush_interval=1000, rel_lo=0, rel_hi=1):
+        return ProgressMonitor(
+            tag             = tag,
+            num_items       = num_items,
+            flush_interval  = flush_interval,
+            verbose         = self.verbose,
+            progress_fn     = self.progress_fn,
+            pfn_lo          = self.pfn_lo + (self.pfn_hi - self.pfn_lo) * rel_lo,
+            pfn_hi          = self.pfn_lo + (self.pfn_hi - self.pfn_lo) * rel_hi,
+            pfn_total       = self.pfn_total,
+        )
+
+#----------------------------------------------------------------------------
+
+def compute_feature_stats_for_dataset(opts, detector_url, detector_kwargs, rel_lo=0, rel_hi=1, batch_size=64, data_loader_kwargs=None, max_items=None, **stats_kwargs):
+    dataset = dnnlib.util.construct_class_by_name(**opts.dataset_kwargs)
+    if data_loader_kwargs is None:
+        data_loader_kwargs = dict(pin_memory=True, num_workers=3, prefetch_factor=2)
+
+    # Try to lookup from cache.
+    cache_file = None
+    if opts.cache:
+        # Choose cache file name.
+        args = dict(dataset_kwargs=opts.dataset_kwargs, detector_url=detector_url, detector_kwargs=detector_kwargs, stats_kwargs=stats_kwargs)
+        md5 = hashlib.md5(repr(sorted(args.items())).encode('utf-8'))
+        cache_tag = f'{dataset.name}-{get_feature_detector_name(detector_url)}-{md5.hexdigest()}'
+        cache_file = dnnlib.make_cache_dir_path('gan-metrics', cache_tag + '.pkl')
+
+        # Check if the file exists (all processes must agree).
+        flag = os.path.isfile(cache_file) if opts.rank == 0 else False
+        if opts.num_gpus > 1:
+            flag = torch.as_tensor(flag, dtype=torch.float32, device=opts.device)
+            torch.distributed.broadcast(tensor=flag, src=0)
+            flag = (float(flag.cpu()) != 0)
+
+        # Load.
+        if flag:
+            return FeatureStats.load(cache_file)
+
+    # Initialize.
+    num_items = len(dataset)
+    if max_items is not None:
+        num_items = min(num_items, max_items)
+    stats = FeatureStats(max_items=num_items, **stats_kwargs)
+    progress = opts.progress.sub(tag='dataset features', num_items=num_items, rel_lo=rel_lo, rel_hi=rel_hi)
+    detector = get_feature_detector(url=detector_url, device=opts.device, num_gpus=opts.num_gpus, rank=opts.rank, verbose=progress.verbose)
+
+    # Main loop.
+    item_subset = [(i * opts.num_gpus + opts.rank) % num_items for i in range((num_items - 1) // opts.num_gpus + 1)]
+    # for images, _labels in torch.utils.data.DataLoader(dataset=dataset, sampler=item_subset, batch_size=batch_size, **data_loader_kwargs):
+    # adaptation to inpainting
+    for images, masks, _labels in torch.utils.data.DataLoader(dataset=dataset, sampler=item_subset, batch_size=batch_size,
+                                                       **data_loader_kwargs):
+    # --------------------------------
+        if images.shape[1] == 1:
+            images = images.repeat([1, 3, 1, 1])
+        features = detector(images.to(opts.device), **detector_kwargs)
+        stats.append_torch(features, num_gpus=opts.num_gpus, rank=opts.rank)
+        progress.update(stats.num_items)
+
+    # Save to cache.
+    if cache_file is not None and opts.rank == 0:
+        os.makedirs(os.path.dirname(cache_file), exist_ok=True)
+        temp_file = cache_file + '.' + uuid.uuid4().hex
+        stats.save(temp_file)
+        os.replace(temp_file, cache_file) # atomic
+    return stats
+
+#----------------------------------------------------------------------------
+
+def compute_feature_stats_for_generator(opts, detector_url, detector_kwargs, rel_lo=0, rel_hi=1, batch_size=64, batch_gen=None, jit=False, data_loader_kwargs=None, **stats_kwargs):
+    if data_loader_kwargs is None:
+        data_loader_kwargs = dict(pin_memory=True, num_workers=3, prefetch_factor=2)
+
+    if batch_gen is None:
+        batch_gen = min(batch_size, 4)
+    assert batch_size % batch_gen == 0
+
+    # Setup generator and load labels.
+    G = copy.deepcopy(opts.G).eval().requires_grad_(False).to(opts.device)
+    dataset = dnnlib.util.construct_class_by_name(**opts.dataset_kwargs)
+
+    # Image generation func.
+    def run_generator(img_in, mask_in, z, c):
+        img = G(img_in, mask_in, z, c, **opts.G_kwargs)
+        # img = (img * 127.5 + 128).clamp(0, 255).to(torch.uint8)
+        img = ((img + 1.0) * 127.5).clamp(0, 255).round().to(torch.uint8)
+        return img
+
+    # # JIT.
+    # if jit:
+    #     z = torch.zeros([batch_gen, G.z_dim], device=opts.device)
+    #     c = torch.zeros([batch_gen, G.c_dim], device=opts.device)
+    #     run_generator = torch.jit.trace(run_generator, [z, c], check_trace=False)
+
+    # Initialize.
+    stats = FeatureStats(**stats_kwargs)
+    assert stats.max_items is not None
+    progress = opts.progress.sub(tag='generator features', num_items=stats.max_items, rel_lo=rel_lo, rel_hi=rel_hi)
+    detector = get_feature_detector(url=detector_url, device=opts.device, num_gpus=opts.num_gpus, rank=opts.rank, verbose=progress.verbose)
+
+    # Main loop.
+    item_subset = [(i * opts.num_gpus + opts.rank) % stats.max_items for i in range((stats.max_items - 1) // opts.num_gpus + 1)]
+    for imgs_batch, masks_batch, labels_batch in torch.utils.data.DataLoader(dataset=dataset, sampler=item_subset,
+                                                              batch_size=batch_size,
+                                                              **data_loader_kwargs):
+        images = []
+        imgs_gen = (imgs_batch.to(opts.device).to(torch.float32) / 127.5 - 1).split(batch_gen)
+        masks_gen = masks_batch.to(opts.device).to(torch.float32).split(batch_gen)
+        for img_in, mask_in in zip(imgs_gen, masks_gen):
+            z = torch.randn([img_in.shape[0], G.z_dim], device=opts.device)
+            c = [dataset.get_label(np.random.randint(len(dataset))) for _i in range(img_in.shape[0])]
+            c = torch.from_numpy(np.stack(c)).pin_memory().to(opts.device)
+            images.append(run_generator(img_in, mask_in, z, c))
+        images = torch.cat(images)
+        if images.shape[1] == 1:
+            images = images.repeat([1, 3, 1, 1])
+        features = detector(images, **detector_kwargs)
+        stats.append_torch(features, num_gpus=opts.num_gpus, rank=opts.rank)
+        progress.update(stats.num_items)
+    return stats
+
+#----------------------------------------------------------------------------
+
+def compute_image_stats_for_generator(opts, rel_lo=0, rel_hi=1, batch_size=64, batch_gen=None, jit=False, data_loader_kwargs=None, **stats_kwargs):
+    if data_loader_kwargs is None:
+        data_loader_kwargs = dict(pin_memory=True, num_workers=3, prefetch_factor=2)
+
+    if batch_gen is None:
+        batch_gen = min(batch_size, 4)
+    assert batch_size % batch_gen == 0
+
+    # Setup generator and load labels.
+    G = copy.deepcopy(opts.G).eval().requires_grad_(False).to(opts.device)
+    dataset = dnnlib.util.construct_class_by_name(**opts.dataset_kwargs)
+
+    # Image generation func.
+    def run_generator(img_in, mask_in, z, c):
+        img = G(img_in, mask_in, z, c, **opts.G_kwargs)
+        # img = (img * 127.5 + 128).clamp(0, 255).to(torch.uint8)
+        img = ((img + 1.0) * 127.5).clamp(0, 255).round().to(torch.uint8)
+        return img
+
+    # Initialize.
+    stats = FeatureStats(**stats_kwargs)
+    assert stats.max_items is not None
+    progress = opts.progress.sub(tag='generator images', num_items=stats.max_items, rel_lo=rel_lo, rel_hi=rel_hi)
+
+    # Main loop.
+    item_subset = [(i * opts.num_gpus + opts.rank) % stats.max_items for i in range((stats.max_items - 1) // opts.num_gpus + 1)]
+    for imgs_batch, masks_batch, labels_batch in torch.utils.data.DataLoader(dataset=dataset, sampler=item_subset,
+                                                                             batch_size=batch_size,
+                                                                             **data_loader_kwargs):
+        images = []
+        imgs_gen = (imgs_batch.to(opts.device).to(torch.float32) / 127.5 - 1).split(batch_gen)
+        masks_gen = masks_batch.to(opts.device).to(torch.float32).split(batch_gen)
+        for img_in, mask_in in zip(imgs_gen, masks_gen):
+            z = torch.randn([img_in.shape[0], G.z_dim], device=opts.device)
+            c = [dataset.get_label(np.random.randint(len(dataset))) for _i in range(img_in.shape[0])]
+            c = torch.from_numpy(np.stack(c)).pin_memory().to(opts.device)
+            images.append(run_generator(img_in, mask_in, z, c))
+        images = torch.cat(images)
+        if images.shape[1] == 1:
+            images = images.repeat([1, 3, 1, 1])
+
+        assert imgs_batch.shape == images.shape
+        metrics = []
+        for i in range(imgs_batch.shape[0]):
+            img_real = np.transpose(imgs_batch[i].cpu().numpy(), [1, 2, 0])
+            img_gen = np.transpose(images[i].cpu().numpy(), [1, 2, 0])
+            psnr = calculate_psnr(img_gen, img_real)
+            ssim = calculate_ssim(img_gen, img_real)
+            l1 = calculate_l1(img_gen, img_real)
+            metrics.append([psnr, ssim, l1])
+        metrics = torch.from_numpy(np.array(metrics)).to(torch.float32).to(opts.device)
+
+        stats.append_torch(metrics, num_gpus=opts.num_gpus, rank=opts.rank)
+        progress.update(stats.num_items)
+    return stats
+
+
+def calculate_psnr(img1, img2):
+    # img1 and img2 have range [0, 255]
+    img1 = img1.astype(np.float64)
+    img2 = img2.astype(np.float64)
+    mse = np.mean((img1 - img2) ** 2)
+    if mse == 0:
+        return float('inf')
+
+    return 20 * math.log10(255.0 / math.sqrt(mse))
+
+
+def calculate_ssim(img1, img2):
+    C1 = (0.01 * 255) ** 2
+    C2 = (0.03 * 255) ** 2
+
+    img1 = img1.astype(np.float64)
+    img2 = img2.astype(np.float64)
+    kernel = cv2.getGaussianKernel(11, 1.5)
+    window = np.outer(kernel, kernel.transpose())
+
+    mu1 = cv2.filter2D(img1, -1, window)[5:-5, 5:-5]
+    mu2 = cv2.filter2D(img2, -1, window)[5:-5, 5:-5]
+    mu1_sq = mu1 ** 2
+    mu2_sq = mu2 ** 2
+    mu1_mu2 = mu1 * mu2
+    sigma1_sq = cv2.filter2D(img1 ** 2, -1, window)[5:-5, 5:-5] - mu1_sq
+    sigma2_sq = cv2.filter2D(img2 ** 2, -1, window)[5:-5, 5:-5] - mu2_sq
+    sigma12 = cv2.filter2D(img1 * img2, -1, window)[5:-5, 5:-5] - mu1_mu2
+
+    ssim_map = ((2 * mu1_mu2 + C1) * (2 * sigma12 + C2)) / ((mu1_sq + mu2_sq + C1) * (sigma1_sq + sigma2_sq + C2))
+
+    return ssim_map.mean()
+
+
+def calculate_l1(img1, img2):
+    img1 = img1.astype(np.float64) / 255.0
+    img2 = img2.astype(np.float64) / 255.0
+    l1 = np.mean(np.abs(img1 - img2))
+
+    return l1
+
+
+# def compute_feature_stats_for_generator(opts, detector_url, detector_kwargs, rel_lo=0, rel_hi=1, batch_size=64, batch_gen=None, jit=False, **stats_kwargs):
+#     if batch_gen is None:
+#         batch_gen = min(batch_size, 4)
+#     assert batch_size % batch_gen == 0
+#
+#     # Setup generator and load labels.
+#     G = copy.deepcopy(opts.G).eval().requires_grad_(False).to(opts.device)
+#     dataset = dnnlib.util.construct_class_by_name(**opts.dataset_kwargs)
+#
+#     # Image generation func.
+#     def run_generator(z, c):
+#         img = G(z=z, c=c, **opts.G_kwargs)
+#         img = (img * 127.5 + 128).clamp(0, 255).to(torch.uint8)
+#         return img
+#
+#     # JIT.
+#     if jit:
+#         z = torch.zeros([batch_gen, G.z_dim], device=opts.device)
+#         c = torch.zeros([batch_gen, G.c_dim], device=opts.device)
+#         run_generator = torch.jit.trace(run_generator, [z, c], check_trace=False)
+#
+#     # Initialize.
+#     stats = FeatureStats(**stats_kwargs)
+#     assert stats.max_items is not None
+#     progress = opts.progress.sub(tag='generator features', num_items=stats.max_items, rel_lo=rel_lo, rel_hi=rel_hi)
+#     detector = get_feature_detector(url=detector_url, device=opts.device, num_gpus=opts.num_gpus, rank=opts.rank, verbose=progress.verbose)
+#
+#     # Main loop.
+#     while not stats.is_full():
+#         images = []
+#         for _i in range(batch_size // batch_gen):
+#             z = torch.randn([batch_gen, G.z_dim], device=opts.device)
+#             c = [dataset.get_label(np.random.randint(len(dataset))) for _i in range(batch_gen)]
+#             c = torch.from_numpy(np.stack(c)).pin_memory().to(opts.device)
+#             images.append(run_generator(z, c))
+#         images = torch.cat(images)
+#         if images.shape[1] == 1:
+#             images = images.repeat([1, 3, 1, 1])
+#         features = detector(images, **detector_kwargs)
+#         stats.append_torch(features, num_gpus=opts.num_gpus, rank=opts.rank)
+#         progress.update(stats.num_items)
+#     return stats
+#
+# #----------------------------------------------------------------------------

metrics/perceptual_path_length.py

@@ -0,0 +1,131 @@
+# Copyright (c) 2021, NVIDIA CORPORATION.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+"""Perceptual Path Length (PPL) from the paper "A Style-Based Generator
+Architecture for Generative Adversarial Networks". Matches the original
+implementation by Karras et al. at
+https://github.com/NVlabs/stylegan/blob/master/metrics/perceptual_path_length.py"""
+
+import copy
+import numpy as np
+import torch
+import dnnlib
+from . import metric_utils
+
+#----------------------------------------------------------------------------
+
+# Spherical interpolation of a batch of vectors.
+def slerp(a, b, t):
+    a = a / a.norm(dim=-1, keepdim=True)
+    b = b / b.norm(dim=-1, keepdim=True)
+    d = (a * b).sum(dim=-1, keepdim=True)
+    p = t * torch.acos(d)
+    c = b - d * a
+    c = c / c.norm(dim=-1, keepdim=True)
+    d = a * torch.cos(p) + c * torch.sin(p)
+    d = d / d.norm(dim=-1, keepdim=True)
+    return d
+
+#----------------------------------------------------------------------------
+
+class PPLSampler(torch.nn.Module):
+    def __init__(self, G, G_kwargs, epsilon, space, sampling, crop, vgg16):
+        assert space in ['z', 'w']
+        assert sampling in ['full', 'end']
+        super().__init__()
+        self.G = copy.deepcopy(G)
+        self.G_kwargs = G_kwargs
+        self.epsilon = epsilon
+        self.space = space
+        self.sampling = sampling
+        self.crop = crop
+        self.vgg16 = copy.deepcopy(vgg16)
+
+    def forward(self, c):
+        # Generate random latents and interpolation t-values.
+        t = torch.rand([c.shape[0]], device=c.device) * (1 if self.sampling == 'full' else 0)
+        z0, z1 = torch.randn([c.shape[0] * 2, self.G.z_dim], device=c.device).chunk(2)
+
+        # Interpolate in W or Z.
+        if self.space == 'w':
+            w0, w1 = self.G.mapping(z=torch.cat([z0,z1]), c=torch.cat([c,c])).chunk(2)
+            wt0 = w0.lerp(w1, t.unsqueeze(1).unsqueeze(2))
+            wt1 = w0.lerp(w1, t.unsqueeze(1).unsqueeze(2) + self.epsilon)
+        else: # space == 'z'
+            zt0 = slerp(z0, z1, t.unsqueeze(1))
+            zt1 = slerp(z0, z1, t.unsqueeze(1) + self.epsilon)
+            wt0, wt1 = self.G.mapping(z=torch.cat([zt0,zt1]), c=torch.cat([c,c])).chunk(2)
+
+        # Randomize noise buffers.
+        for name, buf in self.G.named_buffers():
+            if name.endswith('.noise_const'):
+                buf.copy_(torch.randn_like(buf))
+
+        # Generate images.
+        img = self.G.synthesis(ws=torch.cat([wt0,wt1]), noise_mode='const', force_fp32=True, **self.G_kwargs)
+
+        # Center crop.
+        if self.crop:
+            assert img.shape[2] == img.shape[3]
+            c = img.shape[2] // 8
+            img = img[:, :, c*3 : c*7, c*2 : c*6]
+
+        # Downsample to 256x256.
+        factor = self.G.img_resolution // 256
+        if factor > 1:
+            img = img.reshape([-1, img.shape[1], img.shape[2] // factor, factor, img.shape[3] // factor, factor]).mean([3, 5])
+
+        # Scale dynamic range from [-1,1] to [0,255].
+        img = (img + 1) * (255 / 2)
+        if self.G.img_channels == 1:
+            img = img.repeat([1, 3, 1, 1])
+
+        # Evaluate differential LPIPS.
+        lpips_t0, lpips_t1 = self.vgg16(img, resize_images=False, return_lpips=True).chunk(2)
+        dist = (lpips_t0 - lpips_t1).square().sum(1) / self.epsilon ** 2
+        return dist
+
+#----------------------------------------------------------------------------
+
+def compute_ppl(opts, num_samples, epsilon, space, sampling, crop, batch_size, jit=False):
+    dataset = dnnlib.util.construct_class_by_name(**opts.dataset_kwargs)
+    vgg16_url = 'https://nvlabs-fi-cdn.nvidia.com/stylegan2-ada-pytorch/pretrained/metrics/vgg16.pt'
+    vgg16 = metric_utils.get_feature_detector(vgg16_url, num_gpus=opts.num_gpus, rank=opts.rank, verbose=opts.progress.verbose)
+
+    # Setup sampler.
+    sampler = PPLSampler(G=opts.G, G_kwargs=opts.G_kwargs, epsilon=epsilon, space=space, sampling=sampling, crop=crop, vgg16=vgg16)
+    sampler.eval().requires_grad_(False).to(opts.device)
+    if jit:
+        c = torch.zeros([batch_size, opts.G.c_dim], device=opts.device)
+        sampler = torch.jit.trace(sampler, [c], check_trace=False)
+
+    # Sampling loop.
+    dist = []
+    progress = opts.progress.sub(tag='ppl sampling', num_items=num_samples)
+    for batch_start in range(0, num_samples, batch_size * opts.num_gpus):
+        progress.update(batch_start)
+        c = [dataset.get_label(np.random.randint(len(dataset))) for _i in range(batch_size)]
+        c = torch.from_numpy(np.stack(c)).pin_memory().to(opts.device)
+        x = sampler(c)
+        for src in range(opts.num_gpus):
+            y = x.clone()
+            if opts.num_gpus > 1:
+                torch.distributed.broadcast(y, src=src)
+            dist.append(y)
+    progress.update(num_samples)
+
+    # Compute PPL.
+    if opts.rank != 0:
+        return float('nan')
+    dist = torch.cat(dist)[:num_samples].cpu().numpy()
+    lo = np.percentile(dist, 1, interpolation='lower')
+    hi = np.percentile(dist, 99, interpolation='higher')
+    ppl = np.extract(np.logical_and(dist >= lo, dist <= hi), dist).mean()
+    return float(ppl)
+
+#----------------------------------------------------------------------------

metrics/precision_recall.py

@@ -0,0 +1,62 @@
+# Copyright (c) 2021, NVIDIA CORPORATION.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+"""Precision/Recall (PR) from the paper "Improved Precision and Recall
+Metric for Assessing Generative Models". Matches the original implementation
+by Kynkaanniemi et al. at
+https://github.com/kynkaat/improved-precision-and-recall-metric/blob/master/precision_recall.py"""
+
+import torch
+from . import metric_utils
+
+#----------------------------------------------------------------------------
+
+def compute_distances(row_features, col_features, num_gpus, rank, col_batch_size):
+    assert 0 <= rank < num_gpus
+    num_cols = col_features.shape[0]
+    num_batches = ((num_cols - 1) // col_batch_size // num_gpus + 1) * num_gpus
+    col_batches = torch.nn.functional.pad(col_features, [0, 0, 0, -num_cols % num_batches]).chunk(num_batches)
+    dist_batches = []
+    for col_batch in col_batches[rank :: num_gpus]:
+        dist_batch = torch.cdist(row_features.unsqueeze(0), col_batch.unsqueeze(0))[0]
+        for src in range(num_gpus):
+            dist_broadcast = dist_batch.clone()
+            if num_gpus > 1:
+                torch.distributed.broadcast(dist_broadcast, src=src)
+            dist_batches.append(dist_broadcast.cpu() if rank == 0 else None)
+    return torch.cat(dist_batches, dim=1)[:, :num_cols] if rank == 0 else None
+
+#----------------------------------------------------------------------------
+
+def compute_pr(opts, max_real, num_gen, nhood_size, row_batch_size, col_batch_size):
+    detector_url = 'https://nvlabs-fi-cdn.nvidia.com/stylegan2-ada-pytorch/pretrained/metrics/vgg16.pt'
+    detector_kwargs = dict(return_features=True)
+
+    real_features = metric_utils.compute_feature_stats_for_dataset(
+        opts=opts, detector_url=detector_url, detector_kwargs=detector_kwargs,
+        rel_lo=0, rel_hi=0, capture_all=True, max_items=max_real).get_all_torch().to(torch.float16).to(opts.device)
+
+    gen_features = metric_utils.compute_feature_stats_for_generator(
+        opts=opts, detector_url=detector_url, detector_kwargs=detector_kwargs,
+        rel_lo=0, rel_hi=1, capture_all=True, max_items=num_gen).get_all_torch().to(torch.float16).to(opts.device)
+
+    results = dict()
+    for name, manifold, probes in [('precision', real_features, gen_features), ('recall', gen_features, real_features)]:
+        kth = []
+        for manifold_batch in manifold.split(row_batch_size):
+            dist = compute_distances(row_features=manifold_batch, col_features=manifold, num_gpus=opts.num_gpus, rank=opts.rank, col_batch_size=col_batch_size)
+            kth.append(dist.to(torch.float32).kthvalue(nhood_size + 1).values.to(torch.float16) if opts.rank == 0 else None)
+        kth = torch.cat(kth) if opts.rank == 0 else None
+        pred = []
+        for probes_batch in probes.split(row_batch_size):
+            dist = compute_distances(row_features=probes_batch, col_features=manifold, num_gpus=opts.num_gpus, rank=opts.rank, col_batch_size=col_batch_size)
+            pred.append((dist <= kth).any(dim=1) if opts.rank == 0 else None)
+        results[name] = float(torch.cat(pred).to(torch.float32).mean() if opts.rank == 0 else 'nan')
+    return results['precision'], results['recall']
+
+#----------------------------------------------------------------------------

metrics/psnr_ssim_l1.py

@@ -0,0 +1,19 @@
+import numpy as np
+import scipy.linalg
+from . import metric_utils
+import math
+import cv2
+
+
+def compute_psnr(opts, max_real):
+    # stats: numpy, [N, 3]
+    stats = metric_utils.compute_image_stats_for_generator(opts=opts, capture_all=True, max_items=max_real).get_all()
+
+    if opts.rank != 0:
+        return float('nan'), float('nan'), float('nan')
+
+    print('Number of samples: %d' % stats.shape[0])
+    avg_psnr = stats[:, 0].sum() / stats.shape[0]
+    avg_ssim = stats[:, 1].sum() / stats.shape[0]
+    avg_l1 = stats[:, 2].sum() / stats.shape[0]
+    return avg_psnr, avg_ssim, avg_l1

models/Places_512_FullData+LAION300k.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0230b8b39287e4a1ec4c53a7c724188cf0fe6dab2610bf79cdff3756b8517291
+size 661315824

models/Places_512_FullData.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d960c4e6b3266b6b9fa74ee4458a9482160d54c06d7738696bc9a9e2b34c66dc
+size 661420475

networks/basic_module.py

@@ -0,0 +1,583 @@
+import sys
+sys.path.insert(0, '../')
+from collections import OrderedDict
+import numpy as np
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch_utils import misc
+from torch_utils import persistence
+from torch_utils.ops import conv2d_resample
+from torch_utils.ops import upfirdn2d
+from torch_utils.ops import bias_act
+
+#----------------------------------------------------------------------------
+
+@misc.profiled_function
+def normalize_2nd_moment(x, dim=1, eps=1e-8):
+    return x * (x.square().mean(dim=dim, keepdim=True) + eps).rsqrt()
+
+#----------------------------------------------------------------------------
+
+@persistence.persistent_class
+class FullyConnectedLayer(nn.Module):
+    def __init__(self,
+                 in_features,                # Number of input features.
+                 out_features,               # Number of output features.
+                 bias            = True,     # Apply additive bias before the activation function?
+                 activation      = 'linear', # Activation function: 'relu', 'lrelu', etc.
+                 lr_multiplier   = 1,        # Learning rate multiplier.
+                 bias_init       = 0,        # Initial value for the additive bias.
+                 ):
+        super().__init__()
+        self.weight = torch.nn.Parameter(torch.randn([out_features, in_features]) / lr_multiplier)
+        self.bias = torch.nn.Parameter(torch.full([out_features], np.float32(bias_init))) if bias else None
+        self.activation = activation
+
+        self.weight_gain = lr_multiplier / np.sqrt(in_features)
+        self.bias_gain = lr_multiplier
+
+    def forward(self, x):
+        w = self.weight * self.weight_gain
+        b = self.bias
+        if b is not None and self.bias_gain != 1:
+            b = b * self.bias_gain
+
+        if self.activation == 'linear' and b is not None:
+            # out = torch.addmm(b.unsqueeze(0), x, w.t())
+            x = x.matmul(w.t())
+            out = x + b.reshape([-1 if i == x.ndim-1 else 1 for i in range(x.ndim)])
+        else:
+            x = x.matmul(w.t())
+            out = bias_act.bias_act(x, b, act=self.activation, dim=x.ndim-1)
+        return out
+
+#----------------------------------------------------------------------------
+
+@persistence.persistent_class
+class Conv2dLayer(nn.Module):
+    def __init__(self,
+                 in_channels,                    # Number of input channels.
+                 out_channels,                   # Number of output channels.
+                 kernel_size,                    # Width and height of the convolution kernel.
+                 bias            = True,         # Apply additive bias before the activation function?
+                 activation      = 'linear',     # Activation function: 'relu', 'lrelu', etc.
+                 up              = 1,            # Integer upsampling factor.
+                 down            = 1,            # Integer downsampling factor.
+                 resample_filter = [1,3,3,1],    # Low-pass filter to apply when resampling activations.
+                 conv_clamp      = None,         # Clamp the output to +-X, None = disable clamping.
+                 trainable       = True,         # Update the weights of this layer during training?
+                 ):
+        super().__init__()
+        self.activation = activation
+        self.up = up
+        self.down = down
+        self.register_buffer('resample_filter', upfirdn2d.setup_filter(resample_filter))
+        self.conv_clamp = conv_clamp
+        self.padding = kernel_size // 2
+        self.weight_gain = 1 / np.sqrt(in_channels * (kernel_size ** 2))
+        self.act_gain = bias_act.activation_funcs[activation].def_gain
+
+        weight = torch.randn([out_channels, in_channels, kernel_size, kernel_size])
+        bias = torch.zeros([out_channels]) if bias else None
+        if trainable:
+            self.weight = torch.nn.Parameter(weight)
+            self.bias = torch.nn.Parameter(bias) if bias is not None else None
+        else:
+            self.register_buffer('weight', weight)
+            if bias is not None:
+                self.register_buffer('bias', bias)
+            else:
+                self.bias = None
+
+    def forward(self, x, gain=1):
+        w = self.weight * self.weight_gain
+        x = conv2d_resample.conv2d_resample(x=x, w=w, f=self.resample_filter, up=self.up, down=self.down,
+                                            padding=self.padding)
+
+        act_gain = self.act_gain * gain
+        act_clamp = self.conv_clamp * gain if self.conv_clamp is not None else None
+        out = bias_act.bias_act(x, self.bias, act=self.activation, gain=act_gain, clamp=act_clamp)
+        return out
+
+#----------------------------------------------------------------------------
+
+@persistence.persistent_class
+class ModulatedConv2d(nn.Module):
+    def __init__(self,
+                 in_channels,                   # Number of input channels.
+                 out_channels,                  # Number of output channels.
+                 kernel_size,                   # Width and height of the convolution kernel.
+                 style_dim,                     # dimension of the style code
+                 demodulate=True,               # perfrom demodulation
+                 up=1,                          # Integer upsampling factor.
+                 down=1,                        # Integer downsampling factor.
+                 resample_filter=[1,3,3,1],  # Low-pass filter to apply when resampling activations.
+                 conv_clamp=None,               # Clamp the output to +-X, None = disable clamping.
+                 ):
+        super().__init__()
+        self.demodulate = demodulate
+
+        self.weight = torch.nn.Parameter(torch.randn([1, out_channels, in_channels, kernel_size, kernel_size]))
+        self.out_channels = out_channels
+        self.kernel_size = kernel_size
+        self.weight_gain = 1 / np.sqrt(in_channels * (kernel_size ** 2))
+        self.padding = self.kernel_size // 2
+        self.up = up
+        self.down = down
+        self.register_buffer('resample_filter', upfirdn2d.setup_filter(resample_filter))
+        self.conv_clamp = conv_clamp
+
+        self.affine = FullyConnectedLayer(style_dim, in_channels, bias_init=1)
+
+    def forward(self, x, style):
+        batch, in_channels, height, width = x.shape
+        style = self.affine(style).view(batch, 1, in_channels, 1, 1)
+        weight = self.weight * self.weight_gain * style
+
+        if self.demodulate:
+            decoefs = (weight.pow(2).sum(dim=[2, 3, 4]) + 1e-8).rsqrt()
+            weight = weight * decoefs.view(batch, self.out_channels, 1, 1, 1)
+
+        weight = weight.view(batch * self.out_channels, in_channels, self.kernel_size, self.kernel_size)
+        x = x.view(1, batch * in_channels, height, width)
+        x = conv2d_resample.conv2d_resample(x=x, w=weight, f=self.resample_filter, up=self.up, down=self.down,
+                                            padding=self.padding, groups=batch)
+        out = x.view(batch, self.out_channels, *x.shape[2:])
+
+        return out
+
+#----------------------------------------------------------------------------
+
+@persistence.persistent_class
+class StyleConv(torch.nn.Module):
+    def __init__(self,
+        in_channels,                    # Number of input channels.
+        out_channels,                   # Number of output channels.
+        style_dim,                      # Intermediate latent (W) dimensionality.
+        resolution,                     # Resolution of this layer.
+        kernel_size     = 3,            # Convolution kernel size.
+        up              = 1,            # Integer upsampling factor.
+        use_noise       = True,         # Enable noise input?
+        activation      = 'lrelu',      # Activation function: 'relu', 'lrelu', etc.
+        resample_filter = [1,3,3,1],    # Low-pass filter to apply when resampling activations.
+        conv_clamp      = None,         # Clamp the output of convolution layers to +-X, None = disable clamping.
+        demodulate      = True,         # perform demodulation
+    ):
+        super().__init__()
+
+        self.conv = ModulatedConv2d(in_channels=in_channels,
+                                    out_channels=out_channels,
+                                    kernel_size=kernel_size,
+                                    style_dim=style_dim,
+                                    demodulate=demodulate,
+                                    up=up,
+                                    resample_filter=resample_filter,
+                                    conv_clamp=conv_clamp)
+
+        self.use_noise = use_noise
+        self.resolution = resolution
+        if use_noise:
+            self.register_buffer('noise_const', torch.randn([resolution, resolution]))
+            self.noise_strength = torch.nn.Parameter(torch.zeros([]))
+
+        self.bias = torch.nn.Parameter(torch.zeros([out_channels]))
+        self.activation = activation
+        self.act_gain = bias_act.activation_funcs[activation].def_gain
+        self.conv_clamp = conv_clamp
+
+    def forward(self, x, style, noise_mode='random', gain=1):
+        x = self.conv(x, style)
+
+        assert noise_mode in ['random', 'const', 'none']
+
+        if self.use_noise:
+            if noise_mode == 'random':
+                xh, xw = x.size()[-2:]
+                noise = torch.randn([x.shape[0], 1, xh, xw], device=x.device) \
+                        * self.noise_strength
+            if noise_mode == 'const':
+                noise = self.noise_const * self.noise_strength
+            x = x + noise
+
+        act_gain = self.act_gain * gain
+        act_clamp = self.conv_clamp * gain if self.conv_clamp is not None else None
+        out = bias_act.bias_act(x, self.bias, act=self.activation, gain=act_gain, clamp=act_clamp)
+
+        return out
+
+#----------------------------------------------------------------------------
+
+@persistence.persistent_class
+class ToRGB(torch.nn.Module):
+    def __init__(self,
+                 in_channels,
+                 out_channels,
+                 style_dim,
+                 kernel_size=1,
+                 resample_filter=[1,3,3,1],
+                 conv_clamp=None,
+                 demodulate=False):
+        super().__init__()
+
+        self.conv = ModulatedConv2d(in_channels=in_channels,
+                                    out_channels=out_channels,
+                                    kernel_size=kernel_size,
+                                    style_dim=style_dim,
+                                    demodulate=demodulate,
+                                    resample_filter=resample_filter,
+                                    conv_clamp=conv_clamp)
+        self.bias = torch.nn.Parameter(torch.zeros([out_channels]))
+        self.register_buffer('resample_filter', upfirdn2d.setup_filter(resample_filter))
+        self.conv_clamp = conv_clamp
+
+    def forward(self, x, style, skip=None):
+        x = self.conv(x, style)
+        out = bias_act.bias_act(x, self.bias, clamp=self.conv_clamp)
+
+        if skip is not None:
+            if skip.shape != out.shape:
+                skip = upfirdn2d.upsample2d(skip, self.resample_filter)
+            out = out + skip
+
+        return out
+
+#----------------------------------------------------------------------------
+
+@misc.profiled_function
+def get_style_code(a, b):
+    return torch.cat([a, b], dim=1)
+
+#----------------------------------------------------------------------------
+
+@persistence.persistent_class
+class DecBlockFirst(nn.Module):
+    def __init__(self, in_channels, out_channels, activation, style_dim, use_noise, demodulate, img_channels):
+        super().__init__()
+        self.fc = FullyConnectedLayer(in_features=in_channels*2,
+                                      out_features=in_channels*4**2,
+                                      activation=activation)
+        self.conv = StyleConv(in_channels=in_channels,
+                              out_channels=out_channels,
+                              style_dim=style_dim,
+                              resolution=4,
+                              kernel_size=3,
+                              use_noise=use_noise,
+                              activation=activation,
+                              demodulate=demodulate,
+                              )
+        self.toRGB = ToRGB(in_channels=out_channels,
+                           out_channels=img_channels,
+                           style_dim=style_dim,
+                           kernel_size=1,
+                           demodulate=False,
+                           )
+
+    def forward(self, x, ws, gs, E_features, noise_mode='random'):
+        x = self.fc(x).view(x.shape[0], -1, 4, 4)
+        x = x + E_features[2]
+        style = get_style_code(ws[:, 0], gs)
+        x = self.conv(x, style, noise_mode=noise_mode)
+        style = get_style_code(ws[:, 1], gs)
+        img = self.toRGB(x, style, skip=None)
+
+        return x, img
+
+
+@persistence.persistent_class
+class DecBlockFirstV2(nn.Module):
+    def __init__(self, in_channels, out_channels, activation, style_dim, use_noise, demodulate, img_channels):
+        super().__init__()
+        self.conv0 = Conv2dLayer(in_channels=in_channels,
+                                out_channels=in_channels,
+                                kernel_size=3,
+                                activation=activation,
+                                )
+        self.conv1 = StyleConv(in_channels=in_channels,
+                              out_channels=out_channels,
+                              style_dim=style_dim,
+                              resolution=4,
+                              kernel_size=3,
+                              use_noise=use_noise,
+                              activation=activation,
+                              demodulate=demodulate,
+                              )
+        self.toRGB = ToRGB(in_channels=out_channels,
+                           out_channels=img_channels,
+                           style_dim=style_dim,
+                           kernel_size=1,
+                           demodulate=False,
+                           )
+
+    def forward(self, x, ws, gs, E_features, noise_mode='random'):
+        # x = self.fc(x).view(x.shape[0], -1, 4, 4)
+        x = self.conv0(x)
+        x = x + E_features[2]
+        style = get_style_code(ws[:, 0], gs)
+        x = self.conv1(x, style, noise_mode=noise_mode)
+        style = get_style_code(ws[:, 1], gs)
+        img = self.toRGB(x, style, skip=None)
+
+        return x, img
+
+#----------------------------------------------------------------------------
+
+@persistence.persistent_class
+class DecBlock(nn.Module):
+    def __init__(self, res, in_channels, out_channels, activation, style_dim, use_noise, demodulate, img_channels):  # res = 2, ..., resolution_log2
+        super().__init__()
+        self.res = res
+
+        self.conv0 = StyleConv(in_channels=in_channels,
+                               out_channels=out_channels,
+                               style_dim=style_dim,
+                               resolution=2**res,
+                               kernel_size=3,
+                               up=2,
+                               use_noise=use_noise,
+                               activation=activation,
+                               demodulate=demodulate,
+                               )
+        self.conv1 = StyleConv(in_channels=out_channels,
+                               out_channels=out_channels,
+                               style_dim=style_dim,
+                               resolution=2**res,
+                               kernel_size=3,
+                               use_noise=use_noise,
+                               activation=activation,
+                               demodulate=demodulate,
+                               )
+        self.toRGB = ToRGB(in_channels=out_channels,
+                           out_channels=img_channels,
+                           style_dim=style_dim,
+                           kernel_size=1,
+                           demodulate=False,
+                           )
+
+    def forward(self, x, img, ws, gs, E_features, noise_mode='random'):
+        style = get_style_code(ws[:, self.res * 2 - 5], gs)
+        x = self.conv0(x, style, noise_mode=noise_mode)
+        x = x + E_features[self.res]
+        style = get_style_code(ws[:, self.res * 2 - 4], gs)
+        x = self.conv1(x, style, noise_mode=noise_mode)
+        style = get_style_code(ws[:, self.res * 2 - 3], gs)
+        img = self.toRGB(x, style, skip=img)
+
+        return x, img
+
+#----------------------------------------------------------------------------
+
+@persistence.persistent_class
+class MappingNet(torch.nn.Module):
+    def __init__(self,
+        z_dim,                      # Input latent (Z) dimensionality, 0 = no latent.
+        c_dim,                      # Conditioning label (C) dimensionality, 0 = no label.
+        w_dim,                      # Intermediate latent (W) dimensionality.
+        num_ws,                     # Number of intermediate latents to output, None = do not broadcast.
+        num_layers      = 8,        # Number of mapping layers.
+        embed_features  = None,     # Label embedding dimensionality, None = same as w_dim.
+        layer_features  = None,     # Number of intermediate features in the mapping layers, None = same as w_dim.
+        activation      = 'lrelu',  # Activation function: 'relu', 'lrelu', etc.
+        lr_multiplier   = 0.01,     # Learning rate multiplier for the mapping layers.
+        w_avg_beta      = 0.995,    # Decay for tracking the moving average of W during training, None = do not track.
+    ):
+        super().__init__()
+        self.z_dim = z_dim
+        self.c_dim = c_dim
+        self.w_dim = w_dim
+        self.num_ws = num_ws
+        self.num_layers = num_layers
+        self.w_avg_beta = w_avg_beta
+
+        if embed_features is None:
+            embed_features = w_dim
+        if c_dim == 0:
+            embed_features = 0
+        if layer_features is None:
+            layer_features = w_dim
+        features_list = [z_dim + embed_features] + [layer_features] * (num_layers - 1) + [w_dim]
+
+        if c_dim > 0:
+            self.embed = FullyConnectedLayer(c_dim, embed_features)
+        for idx in range(num_layers):
+            in_features = features_list[idx]
+            out_features = features_list[idx + 1]
+            layer = FullyConnectedLayer(in_features, out_features, activation=activation, lr_multiplier=lr_multiplier)
+            setattr(self, f'fc{idx}', layer)
+
+        if num_ws is not None and w_avg_beta is not None:
+            self.register_buffer('w_avg', torch.zeros([w_dim]))
+
+    def forward(self, z, c, truncation_psi=1, truncation_cutoff=None, skip_w_avg_update=False):
+        # Embed, normalize, and concat inputs.
+        x = None
+        with torch.autograd.profiler.record_function('input'):
+            if self.z_dim > 0:
+                x = normalize_2nd_moment(z.to(torch.float32))
+            if self.c_dim > 0:
+                y = normalize_2nd_moment(self.embed(c.to(torch.float32)))
+                x = torch.cat([x, y], dim=1) if x is not None else y
+
+        # Main layers.
+        for idx in range(self.num_layers):
+            layer = getattr(self, f'fc{idx}')
+            x = layer(x)
+
+        # Update moving average of W.
+        if self.w_avg_beta is not None and self.training and not skip_w_avg_update:
+            with torch.autograd.profiler.record_function('update_w_avg'):
+                self.w_avg.copy_(x.detach().mean(dim=0).lerp(self.w_avg, self.w_avg_beta))
+
+        # Broadcast.
+        if self.num_ws is not None:
+            with torch.autograd.profiler.record_function('broadcast'):
+                x = x.unsqueeze(1).repeat([1, self.num_ws, 1])
+
+        # Apply truncation.
+        if truncation_psi != 1:
+            with torch.autograd.profiler.record_function('truncate'):
+                assert self.w_avg_beta is not None
+                if self.num_ws is None or truncation_cutoff is None:
+                    x = self.w_avg.lerp(x, truncation_psi)
+                else:
+                    x[:, :truncation_cutoff] = self.w_avg.lerp(x[:, :truncation_cutoff], truncation_psi)
+
+        return x
+
+#----------------------------------------------------------------------------
+
+@persistence.persistent_class
+class DisFromRGB(nn.Module):
+    def __init__(self, in_channels, out_channels, activation):  # res = 2, ..., resolution_log2
+        super().__init__()
+        self.conv = Conv2dLayer(in_channels=in_channels,
+                                out_channels=out_channels,
+                                kernel_size=1,
+                                activation=activation,
+                                )
+
+    def forward(self, x):
+        return self.conv(x)
+
+#----------------------------------------------------------------------------
+
+@persistence.persistent_class
+class DisBlock(nn.Module):
+    def __init__(self, in_channels, out_channels, activation):  # res = 2, ..., resolution_log2
+        super().__init__()
+        self.conv0 = Conv2dLayer(in_channels=in_channels,
+                                 out_channels=in_channels,
+                                 kernel_size=3,
+                                 activation=activation,
+                                 )
+        self.conv1 = Conv2dLayer(in_channels=in_channels,
+                                 out_channels=out_channels,
+                                 kernel_size=3,
+                                 down=2,
+                                 activation=activation,
+                                 )
+        self.skip = Conv2dLayer(in_channels=in_channels,
+                                out_channels=out_channels,
+                                kernel_size=1,
+                                down=2,
+                                bias=False,
+                             )
+
+    def forward(self, x):
+        skip = self.skip(x, gain=np.sqrt(0.5))
+        x = self.conv0(x)
+        x = self.conv1(x, gain=np.sqrt(0.5))
+        out = skip + x
+
+        return out
+
+#----------------------------------------------------------------------------
+
+@persistence.persistent_class
+class MinibatchStdLayer(torch.nn.Module):
+    def __init__(self, group_size, num_channels=1):
+        super().__init__()
+        self.group_size = group_size
+        self.num_channels = num_channels
+
+    def forward(self, x):
+        N, C, H, W = x.shape
+        with misc.suppress_tracer_warnings():  # as_tensor results are registered as constants
+            G = torch.min(torch.as_tensor(self.group_size),
+                          torch.as_tensor(N)) if self.group_size is not None else N
+        F = self.num_channels
+        c = C // F
+
+        y = x.reshape(G, -1, F, c, H,
+                      W)  # [GnFcHW] Split minibatch N into n groups of size G, and channels C into F groups of size c.
+        y = y - y.mean(dim=0)  # [GnFcHW] Subtract mean over group.
+        y = y.square().mean(dim=0)  # [nFcHW]  Calc variance over group.
+        y = (y + 1e-8).sqrt()  # [nFcHW]  Calc stddev over group.
+        y = y.mean(dim=[2, 3, 4])  # [nF]     Take average over channels and pixels.
+        y = y.reshape(-1, F, 1, 1)  # [nF11]   Add missing dimensions.
+        y = y.repeat(G, 1, H, W)  # [NFHW]   Replicate over group and pixels.
+        x = torch.cat([x, y], dim=1)  # [NCHW]   Append to input as new channels.
+        return x
+
+#----------------------------------------------------------------------------
+
+@persistence.persistent_class
+class Discriminator(torch.nn.Module):
+    def __init__(self,
+                 c_dim,                        # Conditioning label (C) dimensionality.
+                 img_resolution,               # Input resolution.
+                 img_channels,                 # Number of input color channels.
+                 channel_base       = 32768,    # Overall multiplier for the number of channels.
+                 channel_max        = 512,      # Maximum number of channels in any layer.
+                 channel_decay      = 1,
+                 cmap_dim           = None,     # Dimensionality of mapped conditioning label, None = default.
+                 activation         = 'lrelu',
+                 mbstd_group_size   = 4,        # Group size for the minibatch standard deviation layer, None = entire minibatch.
+                 mbstd_num_channels = 1,        # Number of features for the minibatch standard deviation layer, 0 = disable.
+                 ):
+        super().__init__()
+        self.c_dim = c_dim
+        self.img_resolution = img_resolution
+        self.img_channels = img_channels
+
+        resolution_log2 = int(np.log2(img_resolution))
+        assert img_resolution == 2 ** resolution_log2 and img_resolution >= 4
+        self.resolution_log2 = resolution_log2
+
+        def nf(stage):
+            return np.clip(int(channel_base / 2 ** (stage * channel_decay)), 1, channel_max)
+
+        if cmap_dim == None:
+            cmap_dim = nf(2)
+        if c_dim == 0:
+            cmap_dim = 0
+        self.cmap_dim = cmap_dim
+
+        if c_dim > 0:
+            self.mapping = MappingNet(z_dim=0, c_dim=c_dim, w_dim=cmap_dim, num_ws=None, w_avg_beta=None)
+
+        Dis = [DisFromRGB(img_channels+1, nf(resolution_log2), activation)]
+        for res in range(resolution_log2, 2, -1):
+            Dis.append(DisBlock(nf(res), nf(res-1), activation))
+
+        if mbstd_num_channels > 0:
+            Dis.append(MinibatchStdLayer(group_size=mbstd_group_size, num_channels=mbstd_num_channels))
+        Dis.append(Conv2dLayer(nf(2) + mbstd_num_channels, nf(2), kernel_size=3, activation=activation))
+        self.Dis = nn.Sequential(*Dis)
+
+        self.fc0 = FullyConnectedLayer(nf(2)*4**2, nf(2), activation=activation)
+        self.fc1 = FullyConnectedLayer(nf(2), 1 if cmap_dim == 0 else cmap_dim)
+
+    def forward(self, images_in, masks_in, c):
+        x = torch.cat([masks_in - 0.5, images_in], dim=1)
+        x = self.Dis(x)
+        x = self.fc1(self.fc0(x.flatten(start_dim=1)))
+
+        if self.c_dim > 0:
+            cmap = self.mapping(None, c)
+
+        if self.cmap_dim > 0:
+            x = (x * cmap).sum(dim=1, keepdim=True) * (1 / np.sqrt(self.cmap_dim))
+
+        return x

networks/mat.py

@@ -0,0 +1,996 @@
+import numpy as np
+import math
+import sys
+sys.path.insert(0, '../')
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torch.utils.checkpoint as checkpoint
+from timm.models.layers import DropPath, to_2tuple, trunc_normal_
+
+from torch_utils import misc
+from torch_utils import persistence
+from networks.basic_module import FullyConnectedLayer, Conv2dLayer, MappingNet, MinibatchStdLayer, DisFromRGB, DisBlock, StyleConv, ToRGB, get_style_code
+
+
+@misc.profiled_function
+def nf(stage, channel_base=32768, channel_decay=1.0, channel_max=512):
+    NF = {512: 64, 256: 128, 128: 256, 64: 512, 32: 512, 16: 512, 8: 512, 4: 512}
+    return NF[2 ** stage]
+
+
+@persistence.persistent_class
+class Mlp(nn.Module):
+    def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):
+        super().__init__()
+        out_features = out_features or in_features
+        hidden_features = hidden_features or in_features
+        self.fc1 = FullyConnectedLayer(in_features=in_features, out_features=hidden_features, activation='lrelu')
+        self.fc2 = FullyConnectedLayer(in_features=hidden_features, out_features=out_features)
+
+    def forward(self, x):
+        x = self.fc1(x)
+        x = self.fc2(x)
+        return x
+
+
+@misc.profiled_function
+def window_partition(x, window_size):
+    """
+    Args:
+        x: (B, H, W, C)
+        window_size (int): window size
+    Returns:
+        windows: (num_windows*B, window_size, window_size, C)
+    """
+    B, H, W, C = x.shape
+    x = x.view(B, H // window_size, window_size, W // window_size, window_size, C)
+    windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C)
+    return windows
+
+
+@misc.profiled_function
+def window_reverse(windows, window_size, H, W):
+    """
+    Args:
+        windows: (num_windows*B, window_size, window_size, C)
+        window_size (int): Window size
+        H (int): Height of image
+        W (int): Width of image
+    Returns:
+        x: (B, H, W, C)
+    """
+    B = int(windows.shape[0] / (H * W / window_size / window_size))
+    x = windows.view(B, H // window_size, W // window_size, window_size, window_size, -1)
+    x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, -1)
+    return x
+
+
+@persistence.persistent_class
+class Conv2dLayerPartial(nn.Module):
+    def __init__(self,
+                 in_channels,                    # Number of input channels.
+                 out_channels,                   # Number of output channels.
+                 kernel_size,                    # Width and height of the convolution kernel.
+                 bias            = True,         # Apply additive bias before the activation function?
+                 activation      = 'linear',     # Activation function: 'relu', 'lrelu', etc.
+                 up              = 1,            # Integer upsampling factor.
+                 down            = 1,            # Integer downsampling factor.
+                 resample_filter = [1,3,3,1],    # Low-pass filter to apply when resampling activations.
+                 conv_clamp      = None,         # Clamp the output to +-X, None = disable clamping.
+                 trainable       = True,         # Update the weights of this layer during training?
+                 ):
+        super().__init__()
+        self.conv = Conv2dLayer(in_channels, out_channels, kernel_size, bias, activation, up, down, resample_filter,
+                                conv_clamp, trainable)
+
+        self.weight_maskUpdater = torch.ones(1, 1, kernel_size, kernel_size)
+        self.slide_winsize = kernel_size ** 2
+        self.stride = down
+        self.padding = kernel_size // 2 if kernel_size % 2 == 1 else 0
+
+    def forward(self, x, mask=None):
+        if mask is not None:
+            with torch.no_grad():
+                if self.weight_maskUpdater.type() != x.type():
+                    self.weight_maskUpdater = self.weight_maskUpdater.to(x)
+                update_mask = F.conv2d(mask, self.weight_maskUpdater, bias=None, stride=self.stride, padding=self.padding)
+                mask_ratio = self.slide_winsize / (update_mask + 1e-8)
+                update_mask = torch.clamp(update_mask, 0, 1)  # 0 or 1
+                mask_ratio = torch.mul(mask_ratio, update_mask)
+            x = self.conv(x)
+            x = torch.mul(x, mask_ratio)
+            return x, update_mask
+        else:
+            x = self.conv(x)
+            return x, None
+
+
+@persistence.persistent_class
+class WindowAttention(nn.Module):
+    r""" Window based multi-head self attention (W-MSA) module with relative position bias.
+    It supports both of shifted and non-shifted window.
+    Args:
+        dim (int): Number of input channels.
+        window_size (tuple[int]): The height and width of the window.
+        num_heads (int): Number of attention heads.
+        qkv_bias (bool, optional):  If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set
+        attn_drop (float, optional): Dropout ratio of attention weight. Default: 0.0
+        proj_drop (float, optional): Dropout ratio of output. Default: 0.0
+    """
+
+    def __init__(self, dim, window_size, num_heads, down_ratio=1, qkv_bias=True, qk_scale=None, attn_drop=0., proj_drop=0.):
+
+        super().__init__()
+        self.dim = dim
+        self.window_size = window_size  # Wh, Ww
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+        self.scale = qk_scale or head_dim ** -0.5
+
+        self.q = FullyConnectedLayer(in_features=dim, out_features=dim)
+        self.k = FullyConnectedLayer(in_features=dim, out_features=dim)
+        self.v = FullyConnectedLayer(in_features=dim, out_features=dim)
+        self.proj = FullyConnectedLayer(in_features=dim, out_features=dim)
+
+        self.softmax = nn.Softmax(dim=-1)
+
+    def forward(self, x, mask_windows=None, mask=None):
+        """
+        Args:
+            x: input features with shape of (num_windows*B, N, C)
+            mask: (0/-inf) mask with shape of (num_windows, Wh*Ww, Wh*Ww) or None
+        """
+        B_, N, C = x.shape
+        norm_x = F.normalize(x, p=2.0, dim=-1)
+        q = self.q(norm_x).reshape(B_, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)
+        k = self.k(norm_x).view(B_, -1, self.num_heads, C // self.num_heads).permute(0, 2, 3, 1)
+        v = self.v(x).view(B_, -1, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)
+
+        attn = (q @ k) * self.scale
+
+        if mask is not None:
+            nW = mask.shape[0]
+            attn = attn.view(B_ // nW, nW, self.num_heads, N, N) + mask.unsqueeze(1).unsqueeze(0)
+            attn = attn.view(-1, self.num_heads, N, N)
+
+        if mask_windows is not None:
+            attn_mask_windows = mask_windows.squeeze(-1).unsqueeze(1).unsqueeze(1)
+            attn = attn + attn_mask_windows.masked_fill(attn_mask_windows == 0, float(-100.0)).masked_fill(
+                attn_mask_windows == 1, float(0.0))
+            with torch.no_grad():
+                mask_windows = torch.clamp(torch.sum(mask_windows, dim=1, keepdim=True), 0, 1).repeat(1, N, 1)
+
+        attn = self.softmax(attn)
+
+        x = (attn @ v).transpose(1, 2).reshape(B_, N, C)
+        x = self.proj(x)
+        return x, mask_windows
+
+
+@persistence.persistent_class
+class SwinTransformerBlock(nn.Module):
+    r""" Swin Transformer Block.
+    Args:
+        dim (int): Number of input channels.
+        input_resolution (tuple[int]): Input resulotion.
+        num_heads (int): Number of attention heads.
+        window_size (int): Window size.
+        shift_size (int): Shift size for SW-MSA.
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
+        qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
+        drop (float, optional): Dropout rate. Default: 0.0
+        attn_drop (float, optional): Attention dropout rate. Default: 0.0
+        drop_path (float, optional): Stochastic depth rate. Default: 0.0
+        act_layer (nn.Module, optional): Activation layer. Default: nn.GELU
+        norm_layer (nn.Module, optional): Normalization layer.  Default: nn.LayerNorm
+    """
+
+    def __init__(self, dim, input_resolution, num_heads, down_ratio=1, window_size=7, shift_size=0,
+                 mlp_ratio=4., qkv_bias=True, qk_scale=None, drop=0., attn_drop=0., drop_path=0.,
+                 act_layer=nn.GELU, norm_layer=nn.LayerNorm):
+        super().__init__()
+        self.dim = dim
+        self.input_resolution = input_resolution
+        self.num_heads = num_heads
+        self.window_size = window_size
+        self.shift_size = shift_size
+        self.mlp_ratio = mlp_ratio
+        if min(self.input_resolution) <= self.window_size:
+            # if window size is larger than input resolution, we don't partition windows
+            self.shift_size = 0
+            self.window_size = min(self.input_resolution)
+        assert 0 <= self.shift_size < self.window_size, "shift_size must in 0-window_size"
+
+        if self.shift_size > 0:
+            down_ratio = 1
+        self.attn = WindowAttention(dim, window_size=to_2tuple(self.window_size), num_heads=num_heads,
+                                    down_ratio=down_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop,
+                                    proj_drop=drop)
+
+        self.fuse = FullyConnectedLayer(in_features=dim * 2, out_features=dim, activation='lrelu')
+
+        mlp_hidden_dim = int(dim * mlp_ratio)
+        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
+
+        if self.shift_size > 0:
+            attn_mask = self.calculate_mask(self.input_resolution)
+        else:
+            attn_mask = None
+
+        self.register_buffer("attn_mask", attn_mask)
+
+    def calculate_mask(self, x_size):
+        # calculate attention mask for SW-MSA
+        H, W = x_size
+        img_mask = torch.zeros((1, H, W, 1))  # 1 H W 1
+        h_slices = (slice(0, -self.window_size),
+                    slice(-self.window_size, -self.shift_size),
+                    slice(-self.shift_size, None))
+        w_slices = (slice(0, -self.window_size),
+                    slice(-self.window_size, -self.shift_size),
+                    slice(-self.shift_size, None))
+        cnt = 0
+        for h in h_slices:
+            for w in w_slices:
+                img_mask[:, h, w, :] = cnt
+                cnt += 1
+
+        mask_windows = window_partition(img_mask, self.window_size)  # nW, window_size, window_size, 1
+        mask_windows = mask_windows.view(-1, self.window_size * self.window_size)
+        attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2)
+        attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(attn_mask == 0, float(0.0))
+
+        return attn_mask
+
+    def forward(self, x, x_size, mask=None):
+        # H, W = self.input_resolution
+        H, W = x_size
+        B, L, C = x.shape
+        assert L == H * W, "input feature has wrong size"
+
+        shortcut = x
+        x = x.view(B, H, W, C)
+        if mask is not None:
+            mask = mask.view(B, H, W, 1)
+
+        # cyclic shift
+        if self.shift_size > 0:
+            shifted_x = torch.roll(x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2))
+            if mask is not None:
+                shifted_mask = torch.roll(mask, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2))
+        else:
+            shifted_x = x
+            if mask is not None:
+                shifted_mask = mask
+
+        # partition windows
+        x_windows = window_partition(shifted_x, self.window_size)  # nW*B, window_size, window_size, C
+        x_windows = x_windows.view(-1, self.window_size * self.window_size, C)  # nW*B, window_size*window_size, C
+        if mask is not None:
+            mask_windows = window_partition(shifted_mask, self.window_size)
+            mask_windows = mask_windows.view(-1, self.window_size * self.window_size, 1)
+        else:
+            mask_windows = None
+
+        # W-MSA/SW-MSA (to be compatible for testing on images whose shapes are the multiple of window size
+        if self.input_resolution == x_size:
+            attn_windows, mask_windows = self.attn(x_windows, mask_windows, mask=self.attn_mask)  # nW*B, window_size*window_size, C
+        else:
+            attn_windows, mask_windows = self.attn(x_windows, mask_windows, mask=self.calculate_mask(x_size).to(x.device))  # nW*B, window_size*window_size, C
+
+        # merge windows
+        attn_windows = attn_windows.view(-1, self.window_size, self.window_size, C)
+        shifted_x = window_reverse(attn_windows, self.window_size, H, W)  # B H' W' C
+        if mask is not None:
+            mask_windows = mask_windows.view(-1, self.window_size, self.window_size, 1)
+            shifted_mask = window_reverse(mask_windows, self.window_size, H, W)
+
+        # reverse cyclic shift
+        if self.shift_size > 0:
+            x = torch.roll(shifted_x, shifts=(self.shift_size, self.shift_size), dims=(1, 2))
+            if mask is not None:
+                mask = torch.roll(shifted_mask, shifts=(self.shift_size, self.shift_size), dims=(1, 2))
+        else:
+            x = shifted_x
+            if mask is not None:
+                mask = shifted_mask
+        x = x.view(B, H * W, C)
+        if mask is not None:
+            mask = mask.view(B, H * W, 1)
+
+        # FFN
+        x = self.fuse(torch.cat([shortcut, x], dim=-1))
+        x = self.mlp(x)
+
+        return x, mask
+
+
+@persistence.persistent_class
+class PatchMerging(nn.Module):
+    def __init__(self, in_channels, out_channels, down=2):
+        super().__init__()
+        self.conv = Conv2dLayerPartial(in_channels=in_channels,
+                                       out_channels=out_channels,
+                                       kernel_size=3,
+                                       activation='lrelu',
+                                       down=down,
+                                       )
+        self.down = down
+
+    def forward(self, x, x_size, mask=None):
+        x = token2feature(x, x_size)
+        if mask is not None:
+            mask = token2feature(mask, x_size)
+        x, mask = self.conv(x, mask)
+        if self.down != 1:
+            ratio = 1 / self.down
+            x_size = (int(x_size[0] * ratio), int(x_size[1] * ratio))
+        x = feature2token(x)
+        if mask is not None:
+            mask = feature2token(mask)
+        return x, x_size, mask
+
+
+@persistence.persistent_class
+class PatchUpsampling(nn.Module):
+    def __init__(self, in_channels, out_channels, up=2):
+        super().__init__()
+        self.conv = Conv2dLayerPartial(in_channels=in_channels,
+                                       out_channels=out_channels,
+                                       kernel_size=3,
+                                       activation='lrelu',
+                                       up=up,
+                                       )
+        self.up = up
+
+    def forward(self, x, x_size, mask=None):
+        x = token2feature(x, x_size)
+        if mask is not None:
+            mask = token2feature(mask, x_size)
+        x, mask = self.conv(x, mask)
+        if self.up != 1:
+            x_size = (int(x_size[0] * self.up), int(x_size[1] * self.up))
+        x = feature2token(x)
+        if mask is not None:
+            mask = feature2token(mask)
+        return x, x_size, mask
+
+
+
+@persistence.persistent_class
+class BasicLayer(nn.Module):
+    """ A basic Swin Transformer layer for one stage.
+    Args:
+        dim (int): Number of input channels.
+        input_resolution (tuple[int]): Input resolution.
+        depth (int): Number of blocks.
+        num_heads (int): Number of attention heads.
+        window_size (int): Local window size.
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
+        qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
+        drop (float, optional): Dropout rate. Default: 0.0
+        attn_drop (float, optional): Attention dropout rate. Default: 0.0
+        drop_path (float | tuple[float], optional): Stochastic depth rate. Default: 0.0
+        norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
+        downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default: None
+        use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False.
+    """
+
+    def __init__(self, dim, input_resolution, depth, num_heads, window_size, down_ratio=1,
+                 mlp_ratio=2., qkv_bias=True, qk_scale=None, drop=0., attn_drop=0.,
+                 drop_path=0., norm_layer=nn.LayerNorm, downsample=None, use_checkpoint=False):
+
+        super().__init__()
+        self.dim = dim
+        self.input_resolution = input_resolution
+        self.depth = depth
+        self.use_checkpoint = use_checkpoint
+
+        # patch merging layer
+        if downsample is not None:
+            # self.downsample = downsample(input_resolution, dim=dim, norm_layer=norm_layer)
+            self.downsample = downsample
+        else:
+            self.downsample = None
+
+        # build blocks
+        self.blocks = nn.ModuleList([
+            SwinTransformerBlock(dim=dim, input_resolution=input_resolution,
+                                 num_heads=num_heads, down_ratio=down_ratio, window_size=window_size,
+                                 shift_size=0 if (i % 2 == 0) else window_size // 2,
+                                 mlp_ratio=mlp_ratio,
+                                 qkv_bias=qkv_bias, qk_scale=qk_scale,
+                                 drop=drop, attn_drop=attn_drop,
+                                 drop_path=drop_path[i] if isinstance(drop_path, list) else drop_path,
+                                 norm_layer=norm_layer)
+            for i in range(depth)])
+
+        self.conv = Conv2dLayerPartial(in_channels=dim, out_channels=dim, kernel_size=3, activation='lrelu')
+
+    def forward(self, x, x_size, mask=None):
+        if self.downsample is not None:
+            x, x_size, mask = self.downsample(x, x_size, mask)
+        identity = x
+        for blk in self.blocks:
+            if self.use_checkpoint:
+                x, mask = checkpoint.checkpoint(blk, x, x_size, mask)
+            else:
+                x, mask = blk(x, x_size, mask)
+        if mask is not None:
+            mask = token2feature(mask, x_size)
+        x, mask = self.conv(token2feature(x, x_size), mask)
+        x = feature2token(x) + identity
+        if mask is not None:
+            mask = feature2token(mask)
+        return x, x_size, mask
+
+
+@persistence.persistent_class
+class ToToken(nn.Module):
+    def __init__(self, in_channels=3, dim=128, kernel_size=5, stride=1):
+        super().__init__()
+
+        self.proj = Conv2dLayerPartial(in_channels=in_channels, out_channels=dim, kernel_size=kernel_size, activation='lrelu')
+
+    def forward(self, x, mask):
+        x, mask = self.proj(x, mask)
+
+        return x, mask
+
+#----------------------------------------------------------------------------
+
+@persistence.persistent_class
+class EncFromRGB(nn.Module):
+    def __init__(self, in_channels, out_channels, activation):  # res = 2, ..., resolution_log2
+        super().__init__()
+        self.conv0 = Conv2dLayer(in_channels=in_channels,
+                                out_channels=out_channels,
+                                kernel_size=1,
+                                activation=activation,
+                                )
+        self.conv1 = Conv2dLayer(in_channels=out_channels,
+                                out_channels=out_channels,
+                                kernel_size=3,
+                                activation=activation,
+                                )
+
+    def forward(self, x):
+        x = self.conv0(x)
+        x = self.conv1(x)
+
+        return x
+
+@persistence.persistent_class
+class ConvBlockDown(nn.Module):
+    def __init__(self, in_channels, out_channels, activation):  # res = 2, ..., resolution_log
+        super().__init__()
+
+        self.conv0 = Conv2dLayer(in_channels=in_channels,
+                                 out_channels=out_channels,
+                                 kernel_size=3,
+                                 activation=activation,
+                                 down=2,
+                                 )
+        self.conv1 = Conv2dLayer(in_channels=out_channels,
+                                 out_channels=out_channels,
+                                 kernel_size=3,
+                                 activation=activation,
+                                 )
+
+    def forward(self, x):
+        x = self.conv0(x)
+        x = self.conv1(x)
+
+        return x
+
+
+def token2feature(x, x_size):
+    B, N, C = x.shape
+    h, w = x_size
+    x = x.permute(0, 2, 1).reshape(B, C, h, w)
+    return x
+
+
+def feature2token(x):
+    B, C, H, W = x.shape
+    x = x.view(B, C, -1).transpose(1, 2)
+    return x
+
+
+@persistence.persistent_class
+class Encoder(nn.Module):
+    def __init__(self, res_log2, img_channels, activation, patch_size=5, channels=16, drop_path_rate=0.1):
+        super().__init__()
+
+        self.resolution = []
+
+        for idx, i in enumerate(range(res_log2, 3, -1)):  # from input size to 16x16
+            res = 2 ** i
+            self.resolution.append(res)
+            if i == res_log2:
+                block = EncFromRGB(img_channels * 2 + 1, nf(i), activation)
+            else:
+                block = ConvBlockDown(nf(i+1), nf(i), activation)
+            setattr(self, 'EncConv_Block_%dx%d' % (res, res), block)
+
+    def forward(self, x):
+        out = {}
+        for res in self.resolution:
+            res_log2 = int(np.log2(res))
+            x = getattr(self, 'EncConv_Block_%dx%d' % (res, res))(x)
+            out[res_log2] = x
+
+        return out
+
+
+@persistence.persistent_class
+class ToStyle(nn.Module):
+    def __init__(self, in_channels, out_channels, activation, drop_rate):
+        super().__init__()
+        self.conv = nn.Sequential(
+                Conv2dLayer(in_channels=in_channels, out_channels=in_channels, kernel_size=3, activation=activation, down=2),
+                Conv2dLayer(in_channels=in_channels, out_channels=in_channels, kernel_size=3, activation=activation, down=2),
+                Conv2dLayer(in_channels=in_channels, out_channels=in_channels, kernel_size=3, activation=activation, down=2),
+                )
+
+        self.pool = nn.AdaptiveAvgPool2d(1)
+        self.fc = FullyConnectedLayer(in_features=in_channels,
+                                      out_features=out_channels,
+                                      activation=activation)
+        # self.dropout = nn.Dropout(drop_rate)
+
+    def forward(self, x):
+        x = self.conv(x)
+        x = self.pool(x)
+        x = self.fc(x.flatten(start_dim=1))
+        # x = self.dropout(x)
+
+        return x
+
+
+@persistence.persistent_class
+class DecBlockFirstV2(nn.Module):
+    def __init__(self, res, in_channels, out_channels, activation, style_dim, use_noise, demodulate, img_channels):
+        super().__init__()
+        self.res = res
+
+        self.conv0 = Conv2dLayer(in_channels=in_channels,
+                                out_channels=in_channels,
+                                kernel_size=3,
+                                activation=activation,
+                                )
+        self.conv1 = StyleConv(in_channels=in_channels,
+                              out_channels=out_channels,
+                              style_dim=style_dim,
+                              resolution=2**res,
+                              kernel_size=3,
+                              use_noise=use_noise,
+                              activation=activation,
+                              demodulate=demodulate,
+                              )
+        self.toRGB = ToRGB(in_channels=out_channels,
+                           out_channels=img_channels,
+                           style_dim=style_dim,
+                           kernel_size=1,
+                           demodulate=False,
+                           )
+
+    def forward(self, x, ws, gs, E_features, noise_mode='random'):
+        # x = self.fc(x).view(x.shape[0], -1, 4, 4)
+        x = self.conv0(x)
+        x = x + E_features[self.res]
+        style = get_style_code(ws[:, 0], gs)
+        x = self.conv1(x, style, noise_mode=noise_mode)
+        style = get_style_code(ws[:, 1], gs)
+        img = self.toRGB(x, style, skip=None)
+
+        return x, img
+
+#----------------------------------------------------------------------------
+
+@persistence.persistent_class
+class DecBlock(nn.Module):
+    def __init__(self, res, in_channels, out_channels, activation, style_dim, use_noise, demodulate, img_channels):  # res = 4, ..., resolution_log2
+        super().__init__()
+        self.res = res
+
+        self.conv0 = StyleConv(in_channels=in_channels,
+                               out_channels=out_channels,
+                               style_dim=style_dim,
+                               resolution=2**res,
+                               kernel_size=3,
+                               up=2,
+                               use_noise=use_noise,
+                               activation=activation,
+                               demodulate=demodulate,
+                               )
+        self.conv1 = StyleConv(in_channels=out_channels,
+                               out_channels=out_channels,
+                               style_dim=style_dim,
+                               resolution=2**res,
+                               kernel_size=3,
+                               use_noise=use_noise,
+                               activation=activation,
+                               demodulate=demodulate,
+                               )
+        self.toRGB = ToRGB(in_channels=out_channels,
+                           out_channels=img_channels,
+                           style_dim=style_dim,
+                           kernel_size=1,
+                           demodulate=False,
+                           )
+
+    def forward(self, x, img, ws, gs, E_features, noise_mode='random'):
+        style = get_style_code(ws[:, self.res * 2 - 9], gs)
+        x = self.conv0(x, style, noise_mode=noise_mode)
+        x = x + E_features[self.res]
+        style = get_style_code(ws[:, self.res * 2 - 8], gs)
+        x = self.conv1(x, style, noise_mode=noise_mode)
+        style = get_style_code(ws[:, self.res * 2 - 7], gs)
+        img = self.toRGB(x, style, skip=img)
+
+        return x, img
+
+
+@persistence.persistent_class
+class Decoder(nn.Module):
+    def __init__(self, res_log2, activation, style_dim, use_noise, demodulate, img_channels):
+        super().__init__()
+        self.Dec_16x16 = DecBlockFirstV2(4, nf(4), nf(4), activation, style_dim, use_noise, demodulate, img_channels)
+        for res in range(5, res_log2 + 1):
+            setattr(self, 'Dec_%dx%d' % (2 ** res, 2 ** res),
+                    DecBlock(res, nf(res - 1), nf(res), activation, style_dim, use_noise, demodulate, img_channels))
+        self.res_log2 = res_log2
+
+    def forward(self, x, ws, gs, E_features, noise_mode='random'):
+        x, img = self.Dec_16x16(x, ws, gs, E_features, noise_mode=noise_mode)
+        for res in range(5, self.res_log2 + 1):
+            block = getattr(self, 'Dec_%dx%d' % (2 ** res, 2 ** res))
+            x, img = block(x, img, ws, gs, E_features, noise_mode=noise_mode)
+
+        return img
+
+
+@persistence.persistent_class
+class DecStyleBlock(nn.Module):
+    def __init__(self, res, in_channels, out_channels, activation, style_dim, use_noise, demodulate, img_channels):
+        super().__init__()
+        self.res = res
+
+        self.conv0 = StyleConv(in_channels=in_channels,
+                               out_channels=out_channels,
+                               style_dim=style_dim,
+                               resolution=2**res,
+                               kernel_size=3,
+                               up=2,
+                               use_noise=use_noise,
+                               activation=activation,
+                               demodulate=demodulate,
+                               )
+        self.conv1 = StyleConv(in_channels=out_channels,
+                               out_channels=out_channels,
+                               style_dim=style_dim,
+                               resolution=2**res,
+                               kernel_size=3,
+                               use_noise=use_noise,
+                               activation=activation,
+                               demodulate=demodulate,
+                               )
+        self.toRGB = ToRGB(in_channels=out_channels,
+                           out_channels=img_channels,
+                           style_dim=style_dim,
+                           kernel_size=1,
+                           demodulate=False,
+                           )
+
+    def forward(self, x, img, style, skip, noise_mode='random'):
+        x = self.conv0(x, style, noise_mode=noise_mode)
+        x = x + skip
+        x = self.conv1(x, style, noise_mode=noise_mode)
+        img = self.toRGB(x, style, skip=img)
+
+        return x, img
+
+
+@persistence.persistent_class
+class FirstStage(nn.Module):
+    def __init__(self, img_channels, img_resolution=256, dim=180, w_dim=512, use_noise=False, demodulate=True, activation='lrelu'):
+        super().__init__()
+        res = 64
+
+        self.conv_first = Conv2dLayerPartial(in_channels=img_channels+1, out_channels=dim, kernel_size=3, activation=activation)
+        self.enc_conv = nn.ModuleList()
+        down_time = int(np.log2(img_resolution // res))
+        for i in range(down_time):  # from input size to 64
+            self.enc_conv.append(
+                Conv2dLayerPartial(in_channels=dim, out_channels=dim, kernel_size=3, down=2, activation=activation)
+            )
+
+        # from 64 -> 16 -> 64
+        depths = [2, 3, 4, 3, 2]
+        ratios = [1, 1/2, 1/2, 2, 2]
+        num_heads = 6
+        window_sizes = [8, 16, 16, 16, 8]
+        drop_path_rate = 0.1
+        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))]
+
+        self.tran = nn.ModuleList()
+        for i, depth in enumerate(depths):
+            res = int(res * ratios[i])
+            if ratios[i] < 1:
+                merge = PatchMerging(dim, dim, down=int(1/ratios[i]))
+            elif ratios[i] > 1:
+                merge = PatchUpsampling(dim, dim, up=ratios[i])
+            else:
+                merge = None
+            self.tran.append(
+                BasicLayer(dim=dim, input_resolution=[res, res], depth=depth, num_heads=num_heads,
+                           window_size=window_sizes[i], drop_path=dpr[sum(depths[:i]):sum(depths[:i + 1])],
+                           downsample=merge)
+            )
+
+        # global style
+        down_conv = []
+        for i in range(int(np.log2(16))):
+            down_conv.append(Conv2dLayer(in_channels=dim, out_channels=dim, kernel_size=3, down=2, activation=activation))
+        down_conv.append(nn.AdaptiveAvgPool2d((1, 1)))
+        self.down_conv = nn.Sequential(*down_conv)
+        self.to_style = FullyConnectedLayer(in_features=dim, out_features=dim*2, activation=activation)
+        self.ws_style = FullyConnectedLayer(in_features=w_dim, out_features=dim, activation=activation)
+        self.to_square = FullyConnectedLayer(in_features=dim, out_features=16*16, activation=activation)
+
+        style_dim = dim * 3
+        self.dec_conv = nn.ModuleList()
+        for i in range(down_time):  # from 64 to input size
+            res = res * 2
+            self.dec_conv.append(DecStyleBlock(res, dim, dim, activation, style_dim, use_noise, demodulate, img_channels))
+
+    def forward(self, images_in, masks_in, ws, noise_mode='random'):
+        x = torch.cat([masks_in - 0.5, images_in * masks_in], dim=1)
+
+        skips = []
+        x, mask = self.conv_first(x, masks_in)  # input size
+        skips.append(x)
+        for i, block in enumerate(self.enc_conv):  # input size to 64
+            x, mask = block(x, mask)
+            if i != len(self.enc_conv) - 1:
+                skips.append(x)
+
+        x_size = x.size()[-2:]
+        x = feature2token(x)
+        mask = feature2token(mask)
+        mid = len(self.tran) // 2
+        for i, block in enumerate(self.tran):  # 64 to 16
+            if i < mid:
+                x, x_size, mask = block(x, x_size, mask)
+                skips.append(x)
+            elif i > mid:
+                x, x_size, mask = block(x, x_size, None)
+                x = x + skips[mid - i]
+            else:
+                x, x_size, mask = block(x, x_size, None)
+
+                mul_map = torch.ones_like(x) * 0.5
+                mul_map = F.dropout(mul_map, training=True)
+                ws = self.ws_style(ws[:, -1])
+                add_n = self.to_square(ws).unsqueeze(1)
+                add_n = F.interpolate(add_n, size=x.size(1), mode='linear', align_corners=False).squeeze(1).unsqueeze(-1)
+                x = x * mul_map + add_n * (1 - mul_map)
+                gs = self.to_style(self.down_conv(token2feature(x, x_size)).flatten(start_dim=1))
+                style = torch.cat([gs, ws], dim=1)
+
+        x = token2feature(x, x_size).contiguous()
+        img = None
+        for i, block in enumerate(self.dec_conv):
+            x, img = block(x, img, style, skips[len(self.dec_conv)-i-1], noise_mode=noise_mode)
+
+        # ensemble
+        img = img * (1 - masks_in) + images_in * masks_in
+
+        return img
+
+
+@persistence.persistent_class
+class SynthesisNet(nn.Module):
+    def __init__(self,
+                 w_dim,                     # Intermediate latent (W) dimensionality.
+                 img_resolution,            # Output image resolution.
+                 img_channels   = 3,        # Number of color channels.
+                 channel_base   = 32768,    # Overall multiplier for the number of channels.
+                 channel_decay  = 1.0,
+                 channel_max    = 512,      # Maximum number of channels in any layer.
+                 activation     = 'lrelu',  # Activation function: 'relu', 'lrelu', etc.
+                 drop_rate      = 0.5,
+                 use_noise      = True,
+                 demodulate     = True,
+                 ):
+        super().__init__()
+        resolution_log2 = int(np.log2(img_resolution))
+        assert img_resolution == 2 ** resolution_log2 and img_resolution >= 4
+
+        self.num_layers = resolution_log2 * 2 - 3 * 2
+        self.img_resolution = img_resolution
+        self.resolution_log2 = resolution_log2
+
+        # first stage
+        self.first_stage = FirstStage(img_channels, img_resolution=img_resolution, w_dim=w_dim, use_noise=False, demodulate=demodulate)
+
+        # second stage
+        self.enc = Encoder(resolution_log2, img_channels, activation, patch_size=5, channels=16)
+        self.to_square = FullyConnectedLayer(in_features=w_dim, out_features=16*16, activation=activation)
+        self.to_style = ToStyle(in_channels=nf(4), out_channels=nf(2) * 2, activation=activation, drop_rate=drop_rate)
+        style_dim = w_dim + nf(2) * 2
+        self.dec = Decoder(resolution_log2, activation, style_dim, use_noise, demodulate, img_channels)
+
+    def forward(self, images_in, masks_in, ws, noise_mode='random', return_stg1=False):
+        out_stg1 = self.first_stage(images_in, masks_in, ws, noise_mode=noise_mode)
+
+        # encoder
+        x = images_in * masks_in + out_stg1 * (1 - masks_in)
+        x = torch.cat([masks_in - 0.5, x, images_in * masks_in], dim=1)
+        E_features = self.enc(x)
+
+        fea_16 = E_features[4]
+        mul_map = torch.ones_like(fea_16) * 0.5
+        mul_map = F.dropout(mul_map, training=True)
+        add_n = self.to_square(ws[:, 0]).view(-1, 16, 16).unsqueeze(1)
+        add_n = F.interpolate(add_n, size=fea_16.size()[-2:], mode='bilinear', align_corners=False)
+        fea_16 = fea_16 * mul_map + add_n * (1 - mul_map)
+        E_features[4] = fea_16
+
+        # style
+        gs = self.to_style(fea_16)
+
+        # decoder
+        img = self.dec(fea_16, ws, gs, E_features, noise_mode=noise_mode)
+
+        # ensemble
+        img = img * (1 - masks_in) + images_in * masks_in
+
+        if not return_stg1:
+            return img
+        else:
+            return img, out_stg1
+
+
+@persistence.persistent_class
+class Generator(nn.Module):
+    def __init__(self,
+                 z_dim,                  # Input latent (Z) dimensionality, 0 = no latent.
+                 c_dim,                  # Conditioning label (C) dimensionality, 0 = no label.
+                 w_dim,                  # Intermediate latent (W) dimensionality.
+                 img_resolution,         # resolution of generated image
+                 img_channels,           # Number of input color channels.
+                 synthesis_kwargs = {},  # Arguments for SynthesisNetwork.
+                 mapping_kwargs   = {},  # Arguments for MappingNetwork.
+                 ):
+        super().__init__()
+        self.z_dim = z_dim
+        self.c_dim = c_dim
+        self.w_dim = w_dim
+        self.img_resolution = img_resolution
+        self.img_channels = img_channels
+
+        self.synthesis = SynthesisNet(w_dim=w_dim,
+                                      img_resolution=img_resolution,
+                                      img_channels=img_channels,
+                                      **synthesis_kwargs)
+        self.mapping = MappingNet(z_dim=z_dim,
+                                  c_dim=c_dim,
+                                  w_dim=w_dim,
+                                  num_ws=self.synthesis.num_layers,
+                                  **mapping_kwargs)
+
+    def forward(self, images_in, masks_in, z, c, truncation_psi=1, truncation_cutoff=None, skip_w_avg_update=False,
+                noise_mode='random', return_stg1=False):
+        ws = self.mapping(z, c, truncation_psi=truncation_psi, truncation_cutoff=truncation_cutoff,
+                          skip_w_avg_update=skip_w_avg_update)
+
+        if not return_stg1:
+            img = self.synthesis(images_in, masks_in, ws, noise_mode=noise_mode)
+            return img
+        else:
+            img, out_stg1 = self.synthesis(images_in, masks_in, ws, noise_mode=noise_mode, return_stg1=True)
+            return img, out_stg1
+
+
+@persistence.persistent_class
+class Discriminator(torch.nn.Module):
+    def __init__(self,
+                 c_dim,                        # Conditioning label (C) dimensionality.
+                 img_resolution,               # Input resolution.
+                 img_channels,                 # Number of input color channels.
+                 channel_base       = 32768,    # Overall multiplier for the number of channels.
+                 channel_max        = 512,      # Maximum number of channels in any layer.
+                 channel_decay      = 1,
+                 cmap_dim           = None,     # Dimensionality of mapped conditioning label, None = default.
+                 activation         = 'lrelu',
+                 mbstd_group_size   = 4,        # Group size for the minibatch standard deviation layer, None = entire minibatch.
+                 mbstd_num_channels = 1,        # Number of features for the minibatch standard deviation layer, 0 = disable.
+                 ):
+        super().__init__()
+        self.c_dim = c_dim
+        self.img_resolution = img_resolution
+        self.img_channels = img_channels
+
+        resolution_log2 = int(np.log2(img_resolution))
+        assert img_resolution == 2 ** resolution_log2 and img_resolution >= 4
+        self.resolution_log2 = resolution_log2
+
+        if cmap_dim == None:
+            cmap_dim = nf(2)
+        if c_dim == 0:
+            cmap_dim = 0
+        self.cmap_dim = cmap_dim
+
+        if c_dim > 0:
+            self.mapping = MappingNet(z_dim=0, c_dim=c_dim, w_dim=cmap_dim, num_ws=None, w_avg_beta=None)
+
+        Dis = [DisFromRGB(img_channels+1, nf(resolution_log2), activation)]
+        for res in range(resolution_log2, 2, -1):
+            Dis.append(DisBlock(nf(res), nf(res-1), activation))
+
+        if mbstd_num_channels > 0:
+            Dis.append(MinibatchStdLayer(group_size=mbstd_group_size, num_channels=mbstd_num_channels))
+        Dis.append(Conv2dLayer(nf(2) + mbstd_num_channels, nf(2), kernel_size=3, activation=activation))
+        self.Dis = nn.Sequential(*Dis)
+
+        self.fc0 = FullyConnectedLayer(nf(2)*4**2, nf(2), activation=activation)
+        self.fc1 = FullyConnectedLayer(nf(2), 1 if cmap_dim == 0 else cmap_dim)
+
+        # for 64x64
+        Dis_stg1 = [DisFromRGB(img_channels+1, nf(resolution_log2) // 2, activation)]
+        for res in range(resolution_log2, 2, -1):
+            Dis_stg1.append(DisBlock(nf(res) // 2, nf(res - 1) // 2, activation))
+
+        if mbstd_num_channels > 0:
+            Dis_stg1.append(MinibatchStdLayer(group_size=mbstd_group_size, num_channels=mbstd_num_channels))
+        Dis_stg1.append(Conv2dLayer(nf(2) // 2 + mbstd_num_channels, nf(2) // 2, kernel_size=3, activation=activation))
+        self.Dis_stg1 = nn.Sequential(*Dis_stg1)
+
+        self.fc0_stg1 = FullyConnectedLayer(nf(2) // 2 * 4 ** 2, nf(2) // 2, activation=activation)
+        self.fc1_stg1 = FullyConnectedLayer(nf(2) // 2, 1 if cmap_dim == 0 else cmap_dim)
+
+    def forward(self, images_in, masks_in, images_stg1, c):
+        x = self.Dis(torch.cat([masks_in - 0.5, images_in], dim=1))
+        x = self.fc1(self.fc0(x.flatten(start_dim=1)))
+
+        x_stg1 = self.Dis_stg1(torch.cat([masks_in - 0.5, images_stg1], dim=1))
+        x_stg1 = self.fc1_stg1(self.fc0_stg1(x_stg1.flatten(start_dim=1)))
+
+        if self.c_dim > 0:
+            cmap = self.mapping(None, c)
+
+        if self.cmap_dim > 0:
+            x = (x * cmap).sum(dim=1, keepdim=True) * (1 / np.sqrt(self.cmap_dim))
+            x_stg1 = (x_stg1 * cmap).sum(dim=1, keepdim=True) * (1 / np.sqrt(self.cmap_dim))
+
+        return x, x_stg1
+
+
+if __name__ == '__main__':
+    device = torch.device('cuda:0')
+    batch = 1
+    res = 512
+    G = Generator(z_dim=512, c_dim=0, w_dim=512, img_resolution=512, img_channels=3).to(device)
+    D = Discriminator(c_dim=0, img_resolution=res, img_channels=3).to(device)
+    img = torch.randn(batch, 3, res, res).to(device)
+    mask = torch.randn(batch, 1, res, res).to(device)
+    z = torch.randn(batch, 512).to(device)
+    G.eval()
+
+    # def count(block):
+    #     return sum(p.numel() for p in block.parameters()) / 10 ** 6
+    # print('Generator', count(G))
+    # print('discriminator', count(D))
+
+    with torch.no_grad():
+        img, img_stg1 = G(img, mask, z, None, return_stg1=True)
+    print('output of G:', img.shape, img_stg1.shape)
+    score, score_stg1 = D(img, mask, img_stg1, None)
+    print('output of D:', score.shape, score_stg1.shape)

requirements.txt

@@ -0,0 +1,16 @@
+easydict
+future
+matplotlib
+numpy
+opencv-python
+scikit-image
+scipy
+click
+requests
+tqdm
+pyspng
+ninja
+imageio-ffmpeg==0.4.3
+timm
+psutil
+scikit-learn

torch_utils/__init__.py

@@ -0,0 +1,9 @@
+# Copyright (c) 2021, NVIDIA CORPORATION.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+# empty

torch_utils/custom_ops.py

@@ -0,0 +1,126 @@
+# Copyright (c) 2021, NVIDIA CORPORATION.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+import os
+import glob
+import torch
+import torch.utils.cpp_extension
+import importlib
+import hashlib
+import shutil
+from pathlib import Path
+
+from torch.utils.file_baton import FileBaton
+
+#----------------------------------------------------------------------------
+# Global options.
+
+verbosity = 'brief' # Verbosity level: 'none', 'brief', 'full'
+
+#----------------------------------------------------------------------------
+# Internal helper funcs.
+
+def _find_compiler_bindir():
+    patterns = [
+        'C:/Program Files (x86)/Microsoft Visual Studio/*/Professional/VC/Tools/MSVC/*/bin/Hostx64/x64',
+        'C:/Program Files (x86)/Microsoft Visual Studio/*/BuildTools/VC/Tools/MSVC/*/bin/Hostx64/x64',
+        'C:/Program Files (x86)/Microsoft Visual Studio/*/Community/VC/Tools/MSVC/*/bin/Hostx64/x64',
+        'C:/Program Files (x86)/Microsoft Visual Studio */vc/bin',
+    ]
+    for pattern in patterns:
+        matches = sorted(glob.glob(pattern))
+        if len(matches):
+            return matches[-1]
+    return None
+
+#----------------------------------------------------------------------------
+# Main entry point for compiling and loading C++/CUDA plugins.
+
+_cached_plugins = dict()
+
+def get_plugin(module_name, sources, **build_kwargs):
+    assert verbosity in ['none', 'brief', 'full']
+
+    # Already cached?
+    if module_name in _cached_plugins:
+        return _cached_plugins[module_name]
+
+    # Print status.
+    if verbosity == 'full':
+        print(f'Setting up PyTorch plugin "{module_name}"...')
+    elif verbosity == 'brief':
+        print(f'Setting up PyTorch plugin "{module_name}"... ', end='', flush=True)
+
+    try: # pylint: disable=too-many-nested-blocks
+        # Make sure we can find the necessary compiler binaries.
+        if os.name == 'nt' and os.system("where cl.exe >nul 2>nul") != 0:
+            compiler_bindir = _find_compiler_bindir()
+            if compiler_bindir is None:
+                raise RuntimeError(f'Could not find MSVC/GCC/CLANG installation on this computer. Check _find_compiler_bindir() in "{__file__}".')
+            os.environ['PATH'] += ';' + compiler_bindir
+
+        # Compile and load.
+        verbose_build = (verbosity == 'full')
+
+        # Incremental build md5sum trickery.  Copies all the input source files
+        # into a cached build directory under a combined md5 digest of the input
+        # source files.  Copying is done only if the combined digest has changed.
+        # This keeps input file timestamps and filenames the same as in previous
+        # extension builds, allowing for fast incremental rebuilds.
+        #
+        # This optimization is done only in case all the source files reside in
+        # a single directory (just for simplicity) and if the TORCH_EXTENSIONS_DIR
+        # environment variable is set (we take this as a signal that the user
+        # actually cares about this.)
+        source_dirs_set = set(os.path.dirname(source) for source in sources)
+        if len(source_dirs_set) == 1 and ('TORCH_EXTENSIONS_DIR' in os.environ):
+            all_source_files = sorted(list(x for x in Path(list(source_dirs_set)[0]).iterdir() if x.is_file()))
+
+            # Compute a combined hash digest for all source files in the same
+            # custom op directory (usually .cu, .cpp, .py and .h files).
+            hash_md5 = hashlib.md5()
+            for src in all_source_files:
+                with open(src, 'rb') as f:
+                    hash_md5.update(f.read())
+            build_dir = torch.utils.cpp_extension._get_build_directory(module_name, verbose=verbose_build) # pylint: disable=protected-access
+            digest_build_dir = os.path.join(build_dir, hash_md5.hexdigest())
+
+            if not os.path.isdir(digest_build_dir):
+                os.makedirs(digest_build_dir, exist_ok=True)
+                baton = FileBaton(os.path.join(digest_build_dir, 'lock'))
+                if baton.try_acquire():
+                    try:
+                        for src in all_source_files:
+                            shutil.copyfile(src, os.path.join(digest_build_dir, os.path.basename(src)))
+                    finally:
+                        baton.release()
+                else:
+                    # Someone else is copying source files under the digest dir,
+                    # wait until done and continue.
+                    baton.wait()
+            digest_sources = [os.path.join(digest_build_dir, os.path.basename(x)) for x in sources]
+            torch.utils.cpp_extension.load(name=module_name, build_directory=build_dir,
+                verbose=verbose_build, sources=digest_sources, **build_kwargs)
+        else:
+            torch.utils.cpp_extension.load(name=module_name, verbose=verbose_build, sources=sources, **build_kwargs)
+        module = importlib.import_module(module_name)
+
+    except:
+        if verbosity == 'brief':
+            print('Failed!')
+        raise
+
+    # Print status and add to cache.
+    if verbosity == 'full':
+        print(f'Done setting up PyTorch plugin "{module_name}".')
+    elif verbosity == 'brief':
+        print('Done.')
+    _cached_plugins[module_name] = module
+    return module
+
+#----------------------------------------------------------------------------

torch_utils/misc.py

@@ -0,0 +1,268 @@
+# Copyright (c) 2021, NVIDIA CORPORATION.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+import re
+import contextlib
+import numpy as np
+import torch
+import warnings
+import dnnlib
+
+#----------------------------------------------------------------------------
+# Cached construction of constant tensors. Avoids CPU=>GPU copy when the
+# same constant is used multiple times.
+
+_constant_cache = dict()
+
+def constant(value, shape=None, dtype=None, device=None, memory_format=None):
+    value = np.asarray(value)
+    if shape is not None:
+        shape = tuple(shape)
+    if dtype is None:
+        dtype = torch.get_default_dtype()
+    if device is None:
+        device = torch.device('cpu')
+    if memory_format is None:
+        memory_format = torch.contiguous_format
+
+    key = (value.shape, value.dtype, value.tobytes(), shape, dtype, device, memory_format)
+    tensor = _constant_cache.get(key, None)
+    if tensor is None:
+        tensor = torch.as_tensor(value.copy(), dtype=dtype, device=device)
+        if shape is not None:
+            tensor, _ = torch.broadcast_tensors(tensor, torch.empty(shape))
+        tensor = tensor.contiguous(memory_format=memory_format)
+        _constant_cache[key] = tensor
+    return tensor
+
+#----------------------------------------------------------------------------
+# Replace NaN/Inf with specified numerical values.
+
+try:
+    nan_to_num = torch.nan_to_num # 1.8.0a0
+except AttributeError:
+    def nan_to_num(input, nan=0.0, posinf=None, neginf=None, *, out=None): # pylint: disable=redefined-builtin
+        assert isinstance(input, torch.Tensor)
+        if posinf is None:
+            posinf = torch.finfo(input.dtype).max
+        if neginf is None:
+            neginf = torch.finfo(input.dtype).min
+        assert nan == 0
+        return torch.clamp(input.unsqueeze(0).nansum(0), min=neginf, max=posinf, out=out)
+
+#----------------------------------------------------------------------------
+# Symbolic assert.
+
+try:
+    symbolic_assert = torch._assert # 1.8.0a0 # pylint: disable=protected-access
+except AttributeError:
+    symbolic_assert = torch.Assert # 1.7.0
+
+#----------------------------------------------------------------------------
+# Context manager to suppress known warnings in torch.jit.trace().
+
+class suppress_tracer_warnings(warnings.catch_warnings):
+    def __enter__(self):
+        super().__enter__()
+        warnings.simplefilter('ignore', category=torch.jit.TracerWarning)
+        return self
+
+#----------------------------------------------------------------------------
+# Assert that the shape of a tensor matches the given list of integers.
+# None indicates that the size of a dimension is allowed to vary.
+# Performs symbolic assertion when used in torch.jit.trace().
+
+def assert_shape(tensor, ref_shape):
+    if tensor.ndim != len(ref_shape):
+        raise AssertionError(f'Wrong number of dimensions: got {tensor.ndim}, expected {len(ref_shape)}')
+    for idx, (size, ref_size) in enumerate(zip(tensor.shape, ref_shape)):
+        if ref_size is None:
+            pass
+        elif isinstance(ref_size, torch.Tensor):
+            with suppress_tracer_warnings(): # as_tensor results are registered as constants
+                symbolic_assert(torch.equal(torch.as_tensor(size), ref_size), f'Wrong size for dimension {idx}')
+        elif isinstance(size, torch.Tensor):
+            with suppress_tracer_warnings(): # as_tensor results are registered as constants
+                symbolic_assert(torch.equal(size, torch.as_tensor(ref_size)), f'Wrong size for dimension {idx}: expected {ref_size}')
+        elif size != ref_size:
+            raise AssertionError(f'Wrong size for dimension {idx}: got {size}, expected {ref_size}')
+
+#----------------------------------------------------------------------------
+# Function decorator that calls torch.autograd.profiler.record_function().
+
+def profiled_function(fn):
+    def decorator(*args, **kwargs):
+        with torch.autograd.profiler.record_function(fn.__name__):
+            return fn(*args, **kwargs)
+    decorator.__name__ = fn.__name__
+    return decorator
+
+#----------------------------------------------------------------------------
+# Sampler for torch.utils.data.DataLoader that loops over the dataset
+# indefinitely, shuffling items as it goes.
+
+class InfiniteSampler(torch.utils.data.Sampler):
+    def __init__(self, dataset, rank=0, num_replicas=1, shuffle=True, seed=0, window_size=0.5):
+        assert len(dataset) > 0
+        assert num_replicas > 0
+        assert 0 <= rank < num_replicas
+        assert 0 <= window_size <= 1
+        super().__init__(dataset)
+        self.dataset = dataset
+        self.rank = rank
+        self.num_replicas = num_replicas
+        self.shuffle = shuffle
+        self.seed = seed
+        self.window_size = window_size
+
+    def __iter__(self):
+        order = np.arange(len(self.dataset))
+        rnd = None
+        window = 0
+        if self.shuffle:
+            rnd = np.random.RandomState(self.seed)
+            rnd.shuffle(order)
+            window = int(np.rint(order.size * self.window_size))
+
+        idx = 0
+        while True:
+            i = idx % order.size
+            if idx % self.num_replicas == self.rank:
+                yield order[i]
+            if window >= 2:
+                j = (i - rnd.randint(window)) % order.size
+                order[i], order[j] = order[j], order[i]
+            idx += 1
+
+#----------------------------------------------------------------------------
+# Utilities for operating with torch.nn.Module parameters and buffers.
+
+def params_and_buffers(module):
+    assert isinstance(module, torch.nn.Module)
+    return list(module.parameters()) + list(module.buffers())
+
+def named_params_and_buffers(module):
+    assert isinstance(module, torch.nn.Module)
+    return list(module.named_parameters()) + list(module.named_buffers())
+
+def copy_params_and_buffers(src_module, dst_module, require_all=False):
+    assert isinstance(src_module, torch.nn.Module)
+    assert isinstance(dst_module, torch.nn.Module)
+    src_tensors = {name: tensor for name, tensor in named_params_and_buffers(src_module)}
+    for name, tensor in named_params_and_buffers(dst_module):
+        assert (name in src_tensors) or (not require_all)
+        if name in src_tensors:
+            tensor.copy_(src_tensors[name].detach()).requires_grad_(tensor.requires_grad)
+
+#----------------------------------------------------------------------------
+# Context manager for easily enabling/disabling DistributedDataParallel
+# synchronization.
+
+@contextlib.contextmanager
+def ddp_sync(module, sync):
+    assert isinstance(module, torch.nn.Module)
+    if sync or not isinstance(module, torch.nn.parallel.DistributedDataParallel):
+        yield
+    else:
+        with module.no_sync():
+            yield
+
+#----------------------------------------------------------------------------
+# Check DistributedDataParallel consistency across processes.
+
+def check_ddp_consistency(module, ignore_regex=None):
+    assert isinstance(module, torch.nn.Module)
+    for name, tensor in named_params_and_buffers(module):
+        fullname = type(module).__name__ + '.' + name
+        flag = False
+        if ignore_regex is not None:
+            for regex in ignore_regex:
+                if re.fullmatch(regex, fullname):
+                    flag = True
+                    break
+        if flag:
+            continue
+        tensor = tensor.detach()
+        other = tensor.clone()
+        torch.distributed.broadcast(tensor=other, src=0)
+        assert (nan_to_num(tensor) == nan_to_num(other)).all(), fullname
+
+#----------------------------------------------------------------------------
+# Print summary table of module hierarchy.
+
+def print_module_summary(module, inputs, max_nesting=3, skip_redundant=True):
+    assert isinstance(module, torch.nn.Module)
+    assert not isinstance(module, torch.jit.ScriptModule)
+    assert isinstance(inputs, (tuple, list))
+
+    # Register hooks.
+    entries = []
+    nesting = [0]
+    def pre_hook(_mod, _inputs):
+        nesting[0] += 1
+    def post_hook(mod, _inputs, outputs):
+        nesting[0] -= 1
+        if nesting[0] <= max_nesting:
+            outputs = list(outputs) if isinstance(outputs, (tuple, list)) else [outputs]
+            outputs = [t for t in outputs if isinstance(t, torch.Tensor)]
+            entries.append(dnnlib.EasyDict(mod=mod, outputs=outputs))
+    hooks = [mod.register_forward_pre_hook(pre_hook) for mod in module.modules()]
+    hooks += [mod.register_forward_hook(post_hook) for mod in module.modules()]
+
+    # Run module.
+    outputs = module(*inputs)
+    for hook in hooks:
+        hook.remove()
+
+    # Identify unique outputs, parameters, and buffers.
+    tensors_seen = set()
+    for e in entries:
+        e.unique_params = [t for t in e.mod.parameters() if id(t) not in tensors_seen]
+        e.unique_buffers = [t for t in e.mod.buffers() if id(t) not in tensors_seen]
+        e.unique_outputs = [t for t in e.outputs if id(t) not in tensors_seen]
+        tensors_seen |= {id(t) for t in e.unique_params + e.unique_buffers + e.unique_outputs}
+
+    # Filter out redundant entries.
+    if skip_redundant:
+        entries = [e for e in entries if len(e.unique_params) or len(e.unique_buffers) or len(e.unique_outputs)]
+
+    # Construct table.
+    rows = [[type(module).__name__, 'Parameters', 'Buffers', 'Output shape', 'Datatype']]
+    rows += [['---'] * len(rows[0])]
+    param_total = 0
+    buffer_total = 0
+    submodule_names = {mod: name for name, mod in module.named_modules()}
+    for e in entries:
+        name = '<top-level>' if e.mod is module else submodule_names[e.mod]
+        param_size = sum(t.numel() for t in e.unique_params)
+        buffer_size = sum(t.numel() for t in e.unique_buffers)
+        output_shapes = [str(list(e.outputs[0].shape)) for t in e.outputs]
+        output_dtypes = [str(t.dtype).split('.')[-1] for t in e.outputs]
+        rows += [[
+            name + (':0' if len(e.outputs) >= 2 else ''),
+            str(param_size) if param_size else '-',
+            str(buffer_size) if buffer_size else '-',
+            (output_shapes + ['-'])[0],
+            (output_dtypes + ['-'])[0],
+        ]]
+        for idx in range(1, len(e.outputs)):
+            rows += [[name + f':{idx}', '-', '-', output_shapes[idx], output_dtypes[idx]]]
+        param_total += param_size
+        buffer_total += buffer_size
+    rows += [['---'] * len(rows[0])]
+    rows += [['Total', str(param_total), str(buffer_total), '-', '-']]
+
+    # Print table.
+    widths = [max(len(cell) for cell in column) for column in zip(*rows)]
+    print()
+    for row in rows:
+        print('  '.join(cell + ' ' * (width - len(cell)) for cell, width in zip(row, widths)))
+    print()
+    return outputs
+
+#----------------------------------------------------------------------------

torch_utils/ops/__init__.py

@@ -0,0 +1,9 @@
+# Copyright (c) 2021, NVIDIA CORPORATION.  All rights reserved.
+#
+# NVIDIA CORPORATION and its licensors retain all intellectual property
+# and proprietary rights in and to this software, related documentation
+# and any modifications thereto.  Any use, reproduction, disclosure or
+# distribution of this software and related documentation without an express
+# license agreement from NVIDIA CORPORATION is strictly prohibited.
+
+# empty