Initial commit

README.md

@@ -1,12 +1,12 @@
 ---
-title: HaMeR Test
-emoji: 📚
-colorFrom: pink
-colorTo: purple
+title: HaMeR
+emoji: 🔥
+colorFrom: yellow
+colorTo: yellow
 sdk: gradio
 sdk_version: 4.8.0
 app_file: app.py
 pinned: false
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

_DATA/data/mano/MANO_RIGHT.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:45d60aa3b27ef9107a7afd4e00808f307fd91111e1cfa35afd5c4a62de264767
+size 3821356

_DATA/data/mano_mean_params.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:efc0ec58e4a5cef78f3abfb4e8f91623b8950be9eff8b8e0dbb0d036ebc63988
+size 1178

_DATA/hamer_ckpts/checkpoints/hamer.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e5cc06f294d88a92dee24e603480aab04de532b49f0e08200804ee7d90e16f53
+size 2689536166

_DATA/hamer_ckpts/dataset_config.yaml

@@ -0,0 +1,42 @@
+COCOW-TRAIN:
+  TYPE: ImageDataset
+  URLS: hamer_training_data/dataset_tars/cocow-train/{000000..000036}.tar
+  epoch_size: 78666
+DEX-TRAIN:
+  TYPE: ImageDataset
+  URLS: hamer_training_data/dataset_tars/dex-train/{000000..000406}.tar
+  epoch_size: 406888
+FREIHAND-MOCAP:
+  DATASET_FILE: hamer_training_data/freihand_mocap.npz
+FREIHAND-TRAIN:
+  TYPE: ImageDataset
+  URLS: hamer_training_data/dataset_tars/freihand-train/{000000..000130}.tar
+  epoch_size: 130240
+H2O3D-TRAIN:
+  TYPE: ImageDataset
+  URLS: hamer_training_data/dataset_tars/h2o3d-train/{000000..000060}.tar
+  epoch_size: 121996
+HALPE-TRAIN:
+  TYPE: ImageDataset
+  URLS: hamer_training_data/dataset_tars/halpe-train/{000000..000022}.tar
+  epoch_size: 34289
+HO3D-TRAIN:
+  TYPE: ImageDataset
+  URLS: hamer_training_data/dataset_tars/ho3d-train/{000000..000083}.tar
+  epoch_size: 83325
+INTERHAND26M-TRAIN:
+  TYPE: ImageDataset
+  URLS: hamer_training_data/dataset_tars/interhand26m-train/{000000..001056}.tar
+  epoch_size: 1424632
+MPIINZSL-TRAIN:
+  TYPE: ImageDataset
+  URLS: hamer_training_data/dataset_tars/mpiinzsl-train/{000000..000015}.tar
+  epoch_size: 15184
+MTC-TRAIN:
+  TYPE: ImageDataset
+  URLS: hamer_training_data/dataset_tars/mtc-train/{000000..000306}.tar
+  epoch_size: 363947
+RHD-TRAIN:
+  TYPE: ImageDataset
+  URLS: hamer_training_data/dataset_tars/rhd-train/{000000..000041}.tar
+  epoch_size: 61705

_DATA/hamer_ckpts/model_config.yaml

@@ -0,0 +1,111 @@
+task_name: train
+tags:
+- dev
+train: true
+test: false
+ckpt_path: null
+seed: null
+DATASETS:
+  TRAIN:
+    FREIHAND-TRAIN:
+      WEIGHT: 0.25
+    INTERHAND26M-TRAIN:
+      WEIGHT: 0.25
+    MTC-TRAIN:
+      WEIGHT: 0.1
+    RHD-TRAIN:
+      WEIGHT: 0.05
+    COCOW-TRAIN:
+      WEIGHT: 0.1
+    HALPE-TRAIN:
+      WEIGHT: 0.05
+    MPIINZSL-TRAIN:
+      WEIGHT: 0.05
+    HO3D-TRAIN:
+      WEIGHT: 0.05
+    H2O3D-TRAIN:
+      WEIGHT: 0.05
+    DEX-TRAIN:
+      WEIGHT: 0.05
+  VAL:
+    FREIHAND-TRAIN:
+      WEIGHT: 1.0
+  MOCAP: FREIHAND-MOCAP
+  BETAS_REG: true
+  CONFIG:
+    SCALE_FACTOR: 0.3
+    ROT_FACTOR: 30
+    TRANS_FACTOR: 0.02
+    COLOR_SCALE: 0.2
+    ROT_AUG_RATE: 0.6
+    TRANS_AUG_RATE: 0.5
+    DO_FLIP: false
+    FLIP_AUG_RATE: 0.0
+    EXTREME_CROP_AUG_RATE: 0.0
+    EXTREME_CROP_AUG_LEVEL: 1
+extras:
+  ignore_warnings: false
+  enforce_tags: true
+  print_config: true
+exp_name: hamer
+MANO:
+  DATA_DIR: _DATA/data/
+  MODEL_PATH: _DATA/data/mano
+  GENDER: neutral
+  NUM_HAND_JOINTS: 15
+  MEAN_PARAMS: _DATA/data/mano_mean_params.npz
+  CREATE_BODY_POSE: false
+EXTRA:
+  FOCAL_LENGTH: 5000
+  NUM_LOG_IMAGES: 4
+  NUM_LOG_SAMPLES_PER_IMAGE: 8
+  PELVIS_IND: 0
+GENERAL:
+  TOTAL_STEPS: 1000000
+  LOG_STEPS: 1000
+  VAL_STEPS: 1000
+  CHECKPOINT_STEPS: 10000
+  CHECKPOINT_SAVE_TOP_K: 1
+  NUM_WORKERS: 8
+  PREFETCH_FACTOR: 2
+TRAIN:
+  LR: 1.0e-05
+  WEIGHT_DECAY: 0.0001
+  BATCH_SIZE: 32
+  LOSS_REDUCTION: mean
+  NUM_TRAIN_SAMPLES: 2
+  NUM_TEST_SAMPLES: 64
+  POSE_2D_NOISE_RATIO: 0.01
+  SMPL_PARAM_NOISE_RATIO: 0.005
+MODEL:
+  IMAGE_SIZE: 256
+  IMAGE_MEAN:
+  - 0.485
+  - 0.456
+  - 0.406
+  IMAGE_STD:
+  - 0.229
+  - 0.224
+  - 0.225
+  BACKBONE:
+    TYPE: vit
+    PRETRAINED_WEIGHTS: hamer_training_data/vitpose_backbone.pth
+  MANO_HEAD:
+    TYPE: transformer_decoder
+    IN_CHANNELS: 2048
+    TRANSFORMER_DECODER:
+      depth: 6
+      heads: 8
+      mlp_dim: 1024
+      dim_head: 64
+      dropout: 0.0
+      emb_dropout: 0.0
+      norm: layer
+      context_dim: 1280
+LOSS_WEIGHTS:
+  KEYPOINTS_3D: 0.05
+  KEYPOINTS_2D: 0.01
+  GLOBAL_ORIENT: 0.001
+  HAND_POSE: 0.001
+  BETAS: 0.0005
+  ADVERSARIAL: 0.0005

_DATA/vitpose_ckpts/vitpose+_huge/wholebody.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b0555e1e2392e6a2be2d9265368f344d70ccbfd656ad480aa5c1de2e604519c9
+size 3807742341

app.py

@@ -0,0 +1,234 @@
+import argparse
+import os
+from pathlib import Path
+import tempfile
+import sys
+import cv2
+import gradio as gr
+import numpy as np
+import torch
+from PIL import Image
+
+# print file path
+print(os.path.abspath(__file__))
+os.environ["PYOPENGL_PLATFORM"] = "egl"
+os.environ["MESA_GL_VERSION_OVERRIDE"] = "4.1"
+os.system('pip install /home/user/app/pyrender')
+sys.path.append('/home/user/app/pyrender')
+
+from hamer.configs import get_config
+from hamer.datasets.vitdet_dataset import (DEFAULT_MEAN, DEFAULT_STD,
+                                          ViTDetDataset)
+from hamer.models import HAMER
+from hamer.utils import recursive_to
+from hamer.utils.renderer import Renderer, cam_crop_to_full
+
+try:
+    import detectron2
+except:
+    import os 
+    os.system('pip install --upgrade pip')
+    os.system('pip install git+https://github.com/facebookresearch/detectron2.git')
+
+#try:
+#    from vitpose_model import ViTPoseModel
+#except:
+#    os.system('pip install -v -e /home/user/app/vendor/ViTPose')
+#    from vitpose_model import ViTPoseModel
+from vitpose_model import ViTPoseModel
+
+OUT_FOLDER = 'demo_out'
+os.makedirs(OUT_FOLDER, exist_ok=True)
+
+# Setup HaMeR model
+LIGHT_BLUE=(0.65098039,  0.74117647,  0.85882353)
+DEFAULT_CHECKPOINT='_DATA/hamer_ckpts/checkpoints/hamer.ckpt'
+device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu')
+model_cfg = str(Path(DEFAULT_CHECKPOINT).parent.parent / 'model_config.yaml')
+model_cfg = get_config(model_cfg)
+model = HAMER.load_from_checkpoint(DEFAULT_CHECKPOINT, strict=False, cfg=model_cfg).to(device)
+model.eval()
+
+
+# Load detector
+from detectron2.config import LazyConfig
+
+from hamer.utils.utils_detectron2 import DefaultPredictor_Lazy
+
+detectron2_cfg = LazyConfig.load(f"vendor/detectron2/projects/ViTDet/configs/COCO/cascade_mask_rcnn_vitdet_h_75ep.py")
+detectron2_cfg.train.init_checkpoint = "https://dl.fbaipublicfiles.com/detectron2/ViTDet/COCO/cascade_mask_rcnn_vitdet_h/f328730692/model_final_f05665.pkl"
+for i in range(3):
+    detectron2_cfg.model.roi_heads.box_predictors[i].test_score_thresh = 0.25
+detector = DefaultPredictor_Lazy(detectron2_cfg)
+
+# Setup the renderer
+renderer = Renderer(model_cfg, faces=model.mano.faces)
+
+# keypoint detector
+cpm = ViTPoseModel(device)
+
+import numpy as np
+
+def infer(in_pil_img, in_threshold=0.8, out_pil_img=None):
+
+    open_cv_image = np.array(in_pil_img)
+    # Convert RGB to BGR
+    open_cv_image = open_cv_image[:, :, ::-1].copy()
+    print("EEEEE", open_cv_image.shape)
+    det_out = detector(open_cv_image)
+    det_instances = det_out['instances']
+    valid_idx = (det_instances.pred_classes==0) & (det_instances.scores > in_threshold)
+    pred_bboxes=det_instances.pred_boxes.tensor[valid_idx].cpu().numpy()
+    pred_scores=det_instances.scores[valid_idx].cpu().numpy()
+
+
+    # Detect human keypoints for each person
+    vitposes_out = cpm.predict_pose(
+        open_cv_image,
+        [np.concatenate([pred_bboxes, pred_scores[:, None]], axis=1)],
+    )
+
+    bboxes = []
+    is_right = []
+
+    # Use hands based on hand keypoint detections
+    for vitposes in vitposes_out:
+        left_hand_keyp = vitposes['keypoints'][-42:-21]
+        right_hand_keyp = vitposes['keypoints'][-21:]
+
+        # Rejecting not confident detections (this could be improved)
+        keyp = left_hand_keyp
+        valid = keyp[:,2] > 0.5
+        if sum(valid) > 3:
+            bbox = [keyp[valid,0].min(), keyp[valid,1].min(), keyp[valid,0].max(), keyp[valid,1].max()]
+            bboxes.append(bbox)
+            is_right.append(0)
+        keyp = right_hand_keyp
+        valid = keyp[:,2] > 0.5
+        if sum(valid) > 3:
+            bbox = [keyp[valid,0].min(), keyp[valid,1].min(), keyp[valid,0].max(), keyp[valid,1].max()]
+            bboxes.append(bbox)
+            is_right.append(1)
+
+    if len(bboxes) == 0:
+        return None, []
+
+    boxes = np.stack(bboxes)
+    right = np.stack(is_right)
+
+
+    # Run HaMeR on all detected humans
+    dataset = ViTDetDataset(model_cfg, open_cv_image, boxes, right)
+    dataloader = torch.utils.data.DataLoader(dataset, batch_size=8, shuffle=False, num_workers=0)
+
+    all_verts = []
+    all_cam_t = []
+    all_right = []
+    all_mesh_paths = []
+
+    temp_name = next(tempfile._get_candidate_names())
+    
+    for batch in dataloader:
+        batch = recursive_to(batch, device)
+        with torch.no_grad():
+            out = model(batch)
+
+        multiplier = (2*batch['right']-1)
+        pred_cam = out['pred_cam']
+        pred_cam[:,1] = multiplier*pred_cam[:,1]
+        box_center = batch["box_center"].float()
+        box_size = batch["box_size"].float()
+        img_size = batch["img_size"].float()
+        multiplier = (2*batch['right']-1)
+        render_size = img_size
+        scaled_focal_length = model_cfg.EXTRA.FOCAL_LENGTH / model_cfg.MODEL.IMAGE_SIZE * img_size.max()
+        pred_cam_t = cam_crop_to_full(pred_cam, box_center, box_size, render_size, scaled_focal_length).detach().cpu().numpy()
+
+        # Render the result
+        batch_size = batch['img'].shape[0]
+        for n in range(batch_size):
+            # Get filename from path img_path
+            # img_fn, _ = os.path.splitext(os.path.basename(img_path))
+            person_id = int(batch['personid'][n])
+            white_img = (torch.ones_like(batch['img'][n]).cpu() - DEFAULT_MEAN[:,None,None]/255) / (DEFAULT_STD[:,None,None]/255)
+            input_patch = batch['img'][n].cpu() * (DEFAULT_STD[:,None,None]/255) + (DEFAULT_MEAN[:,None,None]/255)
+            input_patch = input_patch.permute(1,2,0).numpy()
+
+
+            verts = out['pred_vertices'][n].detach().cpu().numpy()
+            is_right = batch['right'][n].cpu().numpy()
+            verts[:,0] = (2*is_right-1)*verts[:,0]
+            cam_t = pred_cam_t[n]
+
+            all_verts.append(verts)
+            all_cam_t.append(cam_t)
+            all_right.append(is_right)
+
+            # Save all meshes to disk
+            # if args.save_mesh:
+            if True:
+                camera_translation = cam_t.copy()
+                tmesh = renderer.vertices_to_trimesh(verts, camera_translation, LIGHT_BLUE, is_right=is_right)
+
+                temp_path = os.path.join(f'{OUT_FOLDER}/{temp_name}_{person_id}.obj')
+                tmesh.export(temp_path)
+                all_mesh_paths.append(temp_path)
+
+    # Render front view
+    if len(all_verts) > 0:
+        misc_args = dict(
+            mesh_base_color=LIGHT_BLUE,
+            scene_bg_color=(1, 1, 1),
+            focal_length=scaled_focal_length,
+        )
+        cam_view = renderer.render_rgba_multiple(all_verts, cam_t=all_cam_t, render_res=render_size[n], is_right=all_right, **misc_args)
+
+        # Overlay image
+        input_img = open_cv_image.astype(np.float32)[:,:,::-1]/255.0
+        input_img = np.concatenate([input_img, np.ones_like(input_img[:,:,:1])], axis=2) # Add alpha channel
+        input_img_overlay = input_img[:,:,:3] * (1-cam_view[:,:,3:]) + cam_view[:,:,:3] * cam_view[:,:,3:]
+
+        # convert to PIL image
+        out_pil_img =  Image.fromarray((input_img_overlay*255).astype(np.uint8))
+
+        return out_pil_img, all_mesh_paths
+    else:
+        return None, []
+
+
+with gr.Blocks(title="HaMeR", css=".gradio-container") as demo:
+
+    gr.HTML("""<div style="font-weight:bold; text-align:center; color:royalblue;">HaMeR</div>""")
+
+    with gr.Row():
+        with gr.Column():
+            input_image = gr.Image(label="Input image", type="pil")
+        with gr.Column():
+            output_image = gr.Image(label="Reconstructions", type="pil")
+            output_meshes = gr.File(label="3D meshes")
+
+    gr.HTML("""<br/>""")
+
+    with gr.Row():
+        threshold = gr.Slider(0, 1.0, value=0.6, label='Detection Threshold')
+        send_btn = gr.Button("Infer")
+        send_btn.click(fn=infer, inputs=[input_image, threshold], outputs=[output_image, output_meshes])
+
+    # with gr.Row():
+    example_images = gr.Examples([
+        ['/home/user/app/assets/test1.jpg'], 
+        ['/home/user/app/assets/test2.jpg'], 
+        ['/home/user/app/assets/test3.jpg'], 
+        ['/home/user/app/assets/test4.jpg'], 
+        ['/home/user/app/assets/test5.jpg'], 
+        ], 
+        inputs=[input_image, 0.6])
+
+
+#demo.queue()
+demo.launch(debug=True)
+
+
+
+
+### EOF ###

hamer/configs/__init__.py

@@ -0,0 +1,111 @@
+import os
+from typing import Dict
+from yacs.config import CfgNode as CN
+
+CACHE_DIR_HAMER = "./_DATA"
+
+def to_lower(x: Dict) -> Dict:
+    """
+    Convert all dictionary keys to lowercase
+    Args:
+      x (dict): Input dictionary
+    Returns:
+      dict: Output dictionary with all keys converted to lowercase
+    """
+    return {k.lower(): v for k, v in x.items()}
+
+_C = CN(new_allowed=True)
+
+_C.GENERAL = CN(new_allowed=True)
+_C.GENERAL.RESUME = True
+_C.GENERAL.TIME_TO_RUN = 3300
+_C.GENERAL.VAL_STEPS = 100
+_C.GENERAL.LOG_STEPS = 100
+_C.GENERAL.CHECKPOINT_STEPS = 20000
+_C.GENERAL.CHECKPOINT_DIR = "checkpoints"
+_C.GENERAL.SUMMARY_DIR = "tensorboard"
+_C.GENERAL.NUM_GPUS = 1
+_C.GENERAL.NUM_WORKERS = 4
+_C.GENERAL.MIXED_PRECISION = True
+_C.GENERAL.ALLOW_CUDA = True
+_C.GENERAL.PIN_MEMORY = False
+_C.GENERAL.DISTRIBUTED = False
+_C.GENERAL.LOCAL_RANK = 0
+_C.GENERAL.USE_SYNCBN = False
+_C.GENERAL.WORLD_SIZE = 1
+
+_C.TRAIN = CN(new_allowed=True)
+_C.TRAIN.NUM_EPOCHS = 100
+_C.TRAIN.BATCH_SIZE = 32
+_C.TRAIN.SHUFFLE = True
+_C.TRAIN.WARMUP = False
+_C.TRAIN.NORMALIZE_PER_IMAGE = False
+_C.TRAIN.CLIP_GRAD = False
+_C.TRAIN.CLIP_GRAD_VALUE = 1.0
+_C.LOSS_WEIGHTS = CN(new_allowed=True)
+
+_C.DATASETS = CN(new_allowed=True)
+
+_C.MODEL = CN(new_allowed=True)
+_C.MODEL.IMAGE_SIZE = 224
+
+_C.EXTRA = CN(new_allowed=True)
+_C.EXTRA.FOCAL_LENGTH = 5000
+
+_C.DATASETS.CONFIG = CN(new_allowed=True)
+_C.DATASETS.CONFIG.SCALE_FACTOR = 0.3
+_C.DATASETS.CONFIG.ROT_FACTOR = 30
+_C.DATASETS.CONFIG.TRANS_FACTOR = 0.02
+_C.DATASETS.CONFIG.COLOR_SCALE = 0.2
+_C.DATASETS.CONFIG.ROT_AUG_RATE = 0.6
+_C.DATASETS.CONFIG.TRANS_AUG_RATE = 0.5
+_C.DATASETS.CONFIG.DO_FLIP = False
+_C.DATASETS.CONFIG.FLIP_AUG_RATE = 0.5
+_C.DATASETS.CONFIG.EXTREME_CROP_AUG_RATE = 0.10
+
+def default_config() -> CN:
+    """
+    Get a yacs CfgNode object with the default config values.
+    """
+    # Return a clone so that the defaults will not be altered
+    # This is for the "local variable" use pattern
+    return _C.clone()
+
+def dataset_config() -> CN:
+    """
+    Get dataset config file
+    Returns:
+      CfgNode: Dataset config as a yacs CfgNode object.
+    """
+    cfg = CN(new_allowed=True)
+    config_file = os.path.join(os.path.dirname(os.path.realpath(__file__)), 'datasets_tar.yaml')
+    cfg.merge_from_file(config_file)
+    cfg.freeze()
+    return cfg
+
+def get_config(config_file: str, merge: bool = True, update_cachedir: bool = False) -> CN:
+    """
+    Read a config file and optionally merge it with the default config file.
+    Args:
+      config_file (str): Path to config file.
+      merge (bool): Whether to merge with the default config or not.
+    Returns:
+      CfgNode: Config as a yacs CfgNode object.
+    """
+    if merge:
+      cfg = default_config()
+    else:
+      cfg = CN(new_allowed=True)
+    cfg.merge_from_file(config_file)
+
+    if update_cachedir:
+      def update_path(path: str) -> str:
+        if os.path.isabs(path):
+          return path
+        return os.path.join(CACHE_DIR_HAMER, path)
+
+      cfg.MANO.MODEL_PATH = update_path(cfg.MANO.MODEL_PATH)
+      cfg.MANO.MEAN_PARAMS = update_path(cfg.MANO.MEAN_PARAMS)
+
+    cfg.freeze()
+    return cfg

hamer/configs/cascade_mask_rcnn_vitdet_h_75ep.py

@@ -0,0 +1,129 @@
+## coco_loader_lsj.py
+
+import detectron2.data.transforms as T
+from detectron2 import model_zoo
+from detectron2.config import LazyCall as L
+
+# Data using LSJ
+image_size = 1024
+dataloader = model_zoo.get_config("common/data/coco.py").dataloader
+dataloader.train.mapper.augmentations = [
+    L(T.RandomFlip)(horizontal=True),  # flip first
+    L(T.ResizeScale)(
+        min_scale=0.1, max_scale=2.0, target_height=image_size, target_width=image_size
+    ),
+    L(T.FixedSizeCrop)(crop_size=(image_size, image_size), pad=False),
+]
+dataloader.train.mapper.image_format = "RGB"
+dataloader.train.total_batch_size = 64
+# recompute boxes due to cropping
+dataloader.train.mapper.recompute_boxes = True
+
+dataloader.test.mapper.augmentations = [
+    L(T.ResizeShortestEdge)(short_edge_length=image_size, max_size=image_size),
+]
+
+from functools import partial
+from fvcore.common.param_scheduler import MultiStepParamScheduler
+
+from detectron2 import model_zoo
+from detectron2.config import LazyCall as L
+from detectron2.solver import WarmupParamScheduler
+from detectron2.modeling.backbone.vit import get_vit_lr_decay_rate
+
+# mask_rcnn_vitdet_b_100ep.py
+
+model = model_zoo.get_config("common/models/mask_rcnn_vitdet.py").model
+
+# Initialization and trainer settings
+train = model_zoo.get_config("common/train.py").train
+train.amp.enabled = True
+train.ddp.fp16_compression = True
+train.init_checkpoint = "detectron2://ImageNetPretrained/MAE/mae_pretrain_vit_base.pth"
+
+
+# Schedule
+# 100 ep = 184375 iters * 64 images/iter / 118000 images/ep
+train.max_iter = 184375
+
+lr_multiplier = L(WarmupParamScheduler)(
+    scheduler=L(MultiStepParamScheduler)(
+        values=[1.0, 0.1, 0.01],
+        milestones=[163889, 177546],
+        num_updates=train.max_iter,
+    ),
+    warmup_length=250 / train.max_iter,
+    warmup_factor=0.001,
+)
+
+# Optimizer
+optimizer = model_zoo.get_config("common/optim.py").AdamW
+optimizer.params.lr_factor_func = partial(get_vit_lr_decay_rate, num_layers=12, lr_decay_rate=0.7)
+optimizer.params.overrides = {"pos_embed": {"weight_decay": 0.0}}
+
+# cascade_mask_rcnn_vitdet_b_100ep.py
+
+from detectron2.config import LazyCall as L
+from detectron2.layers import ShapeSpec
+from detectron2.modeling.box_regression import Box2BoxTransform
+from detectron2.modeling.matcher import Matcher
+from detectron2.modeling.roi_heads import (
+    FastRCNNOutputLayers,
+    FastRCNNConvFCHead,
+    CascadeROIHeads,
+)
+
+# arguments that don't exist for Cascade R-CNN
+[model.roi_heads.pop(k) for k in ["box_head", "box_predictor", "proposal_matcher"]]
+
+model.roi_heads.update(
+    _target_=CascadeROIHeads,
+    box_heads=[
+        L(FastRCNNConvFCHead)(
+            input_shape=ShapeSpec(channels=256, height=7, width=7),
+            conv_dims=[256, 256, 256, 256],
+            fc_dims=[1024],
+            conv_norm="LN",
+        )
+        for _ in range(3)
+    ],
+    box_predictors=[
+        L(FastRCNNOutputLayers)(
+            input_shape=ShapeSpec(channels=1024),
+            test_score_thresh=0.05,
+            box2box_transform=L(Box2BoxTransform)(weights=(w1, w1, w2, w2)),
+            cls_agnostic_bbox_reg=True,
+            num_classes="${...num_classes}",
+        )
+        for (w1, w2) in [(10, 5), (20, 10), (30, 15)]
+    ],
+    proposal_matchers=[
+        L(Matcher)(thresholds=[th], labels=[0, 1], allow_low_quality_matches=False)
+        for th in [0.5, 0.6, 0.7]
+    ],
+)
+
+# cascade_mask_rcnn_vitdet_h_75ep.py
+
+from functools import partial
+
+train.init_checkpoint = "detectron2://ImageNetPretrained/MAE/mae_pretrain_vit_huge_p14to16.pth"
+
+model.backbone.net.embed_dim = 1280
+model.backbone.net.depth = 32
+model.backbone.net.num_heads = 16
+model.backbone.net.drop_path_rate = 0.5
+# 7, 15, 23, 31 for global attention
+model.backbone.net.window_block_indexes = (
+    list(range(0, 7)) + list(range(8, 15)) + list(range(16, 23)) + list(range(24, 31))
+)
+
+optimizer.params.lr_factor_func = partial(get_vit_lr_decay_rate, lr_decay_rate=0.9, num_layers=32)
+optimizer.params.overrides = {}
+optimizer.params.weight_decay_norm = None
+
+train.max_iter = train.max_iter * 3 // 4  # 100ep -> 75ep
+lr_multiplier.scheduler.milestones = [
+    milestone * 3 // 4 for milestone in lr_multiplier.scheduler.milestones
+]
+lr_multiplier.scheduler.num_updates = train.max_iter

hamer/configs/datasets_tar.yaml

@@ -0,0 +1,42 @@
+FREIHAND-TRAIN:
+    TYPE: ImageDataset
+    URLS: hamer_training_data/dataset_tars/freihand-train/{000000..000130}.tar
+    epoch_size: 130_240
+INTERHAND26M-TRAIN:
+    TYPE: ImageDataset
+    URLS: hamer_training_data/dataset_tars/interhand26m-train/{000000..001056}.tar
+    epoch_size: 1_424_632
+HALPE-TRAIN:
+    TYPE: ImageDataset
+    URLS: hamer_training_data/dataset_tars/halpe-train/{000000..000022}.tar
+    epoch_size: 34_289
+COCOW-TRAIN:
+    TYPE: ImageDataset
+    URLS: hamer_training_data/dataset_tars/cocow-train/{000000..000036}.tar
+    epoch_size: 78_666
+MTC-TRAIN:
+    TYPE: ImageDataset
+    URLS: hamer_training_data/dataset_tars/mtc-train/{000000..000306}.tar
+    epoch_size: 363_947
+RHD-TRAIN:
+    TYPE: ImageDataset
+    URLS: hamer_training_data/dataset_tars/rhd-train/{000000..000041}.tar
+    epoch_size: 61_705
+MPIINZSL-TRAIN:
+    TYPE: ImageDataset
+    URLS: hamer_training_data/dataset_tars/mpiinzsl-train/{000000..000015}.tar
+    epoch_size: 15_184
+HO3D-TRAIN:
+    TYPE: ImageDataset
+    URLS: hamer_training_data/dataset_tars/ho3d-train/{000000..000083}.tar
+    epoch_size: 83_325
+H2O3D-TRAIN:
+    TYPE: ImageDataset
+    URLS: hamer_training_data/dataset_tars/h2o3d-train/{000000..000060}.tar
+    epoch_size: 121_996
+DEX-TRAIN:
+    TYPE: ImageDataset
+    URLS: hamer_training_data/dataset_tars/dex-train/{000000..000406}.tar
+    epoch_size: 406_888
+FREIHAND-MOCAP:
+    DATASET_FILE: hamer_training_data/freihand_mocap.npz

hamer/configs_hydra/data/mix_all.yaml

@@ -0,0 +1,31 @@
+# @package _global_
+defaults:
+  - /data_filtering: low1
+
+DATASETS:
+  TRAIN: 
+    FREIHAND-TRAIN:
+      WEIGHT: 0.25
+    INTERHAND26M-TRAIN:
+      WEIGHT: 0.25
+    MTC-TRAIN:
+      WEIGHT: 0.1
+    RHD-TRAIN:
+      WEIGHT: 0.05
+    COCOW-TRAIN:
+      WEIGHT: 0.1
+    HALPE-TRAIN:
+      WEIGHT: 0.05
+    MPIINZSL-TRAIN:
+      WEIGHT: 0.05
+    HO3D-TRAIN:
+      WEIGHT: 0.05
+    H2O3D-TRAIN:
+      WEIGHT: 0.05
+    DEX-TRAIN:
+      WEIGHT: 0.05
+  VAL:
+    FREIHAND-TRAIN:
+      WEIGHT: 1.0
+
+  MOCAP: FREIHAND-MOCAP

hamer/configs_hydra/data_filtering/low1.yaml

@@ -0,0 +1,13 @@
+# @package _global_
+
+DATASETS:
+  # Data filtering during training
+  SUPPRESS_KP_CONF_THRESH: 0.3
+  FILTER_NUM_KP: 4
+  FILTER_NUM_KP_THRESH: 0.0
+  FILTER_REPROJ_THRESH: 31000
+
+  SUPPRESS_BETAS_THRESH: 3.0
+  SUPPRESS_BAD_POSES: False
+  POSES_BETAS_SIMULTANEOUS: True
+  FILTER_NO_POSES: False # If True, filters images that don't have poses

hamer/configs_hydra/experiment/default.yaml

@@ -0,0 +1,29 @@
+# @package _global_
+
+MANO:
+  DATA_DIR: ${oc.env:HOME}/.cache/4DHumans/data/
+  MODEL_PATH: ${MANO.DATA_DIR}/mano
+  GENDER: neutral
+  NUM_HAND_JOINTS: 15
+  MEAN_PARAMS: ${MANO.DATA_DIR}/mano_mean_params.npz
+  CREATE_BODY_POSE: FALSE
+
+EXTRA:
+  FOCAL_LENGTH: 5000
+  NUM_LOG_IMAGES: 4
+  NUM_LOG_SAMPLES_PER_IMAGE: 8
+  PELVIS_IND: 0
+
+DATASETS:
+  BETAS_REG: True
+  CONFIG:
+    SCALE_FACTOR: 0.3
+    ROT_FACTOR: 30
+    TRANS_FACTOR: 0.02
+    COLOR_SCALE: 0.2
+    ROT_AUG_RATE: 0.6
+    TRANS_AUG_RATE: 0.5
+    DO_FLIP: False
+    FLIP_AUG_RATE: 0.0
+    EXTREME_CROP_AUG_RATE: 0.0
+    EXTREME_CROP_AUG_LEVEL: 1

hamer/configs_hydra/experiment/hamer_vit_transformer.yaml

@@ -0,0 +1,51 @@
+# @package _global_
+
+defaults:
+  - default.yaml
+
+GENERAL:
+  TOTAL_STEPS: 1_000_000
+  LOG_STEPS: 1000
+  VAL_STEPS: 1000
+  CHECKPOINT_STEPS: 1000
+  CHECKPOINT_SAVE_TOP_K: 1
+  NUM_WORKERS: 25
+  PREFETCH_FACTOR: 2
+
+TRAIN:
+  LR: 1e-5
+  WEIGHT_DECAY: 1e-4
+  BATCH_SIZE: 8
+  LOSS_REDUCTION: mean
+  NUM_TRAIN_SAMPLES: 2
+  NUM_TEST_SAMPLES: 64
+  POSE_2D_NOISE_RATIO: 0.01
+  SMPL_PARAM_NOISE_RATIO: 0.005
+
+MODEL:
+  IMAGE_SIZE: 256
+  IMAGE_MEAN: [0.485, 0.456, 0.406]
+  IMAGE_STD: [0.229, 0.224, 0.225]
+  BACKBONE:
+    TYPE: vit
+    PRETRAINED_WEIGHTS: hamer_training_data/vitpose_backbone.pth
+  MANO_HEAD:
+    TYPE: transformer_decoder
+    IN_CHANNELS: 2048
+    TRANSFORMER_DECODER:
+      depth: 6
+      heads: 8
+      mlp_dim: 1024
+      dim_head: 64
+      dropout: 0.0
+      emb_dropout: 0.0
+      norm: layer
+      context_dim: 1280 # from vitpose-H
+
+LOSS_WEIGHTS:
+  KEYPOINTS_3D: 0.05
+  KEYPOINTS_2D: 0.01
+  GLOBAL_ORIENT: 0.001
+  HAND_POSE: 0.001
+  BETAS: 0.0005
+  ADVERSARIAL: 0.0005

hamer/configs_hydra/extras/default.yaml

@@ -0,0 +1,8 @@
+# disable python warnings if they annoy you
+ignore_warnings: False
+
+# ask user for tags if none are provided in the config
+enforce_tags: True
+
+# pretty print config tree at the start of the run using Rich library
+print_config: True

hamer/configs_hydra/hydra/default.yaml

@@ -0,0 +1,26 @@
+# @package _global_
+# https://hydra.cc/docs/configure_hydra/intro/
+
+# enable color logging
+defaults:
+  - override /hydra/hydra_logging: colorlog
+  - override /hydra/job_logging: colorlog
+
+# exp_name: ovrd_${hydra:job.override_dirname}
+exp_name: ${now:%Y-%m-%d}_${now:%H-%M-%S}
+
+hydra:
+  run:
+    dir: ${paths.log_dir}/${task_name}/runs/${exp_name}
+  sweep:
+    dir: ${paths.log_dir}/${task_name}/multiruns/${exp_name}
+    subdir: ${hydra.job.num}
+  job:
+    config:
+      override_dirname:
+        exclude_keys:
+          - trainer
+          - trainer.devices
+          - trainer.num_nodes
+          - callbacks
+          - debug

hamer/configs_hydra/launcher/local.yaml

@@ -0,0 +1,13 @@
+# @package _global_
+
+defaults:
+  - override /hydra/launcher: submitit_local
+
+hydra:
+  launcher:
+    timeout_min: 10_080   # 7 days
+    nodes: 1
+    tasks_per_node: ${trainer.devices}
+    cpus_per_task: 6
+    gpus_per_node: ${trainer.devices}
+    name: hamer

hamer/configs_hydra/launcher/slurm.yaml

@@ -0,0 +1,22 @@
+# @package _global_
+
+defaults:
+  - override /hydra/launcher: submitit_slurm
+
+hydra:
+  launcher:
+    timeout_min: 10_080   # 7 days
+    max_num_timeout: 3
+    partition: g40
+    qos: idle
+    nodes: 1
+    tasks_per_node: ${trainer.devices}
+    gpus_per_task: null
+    cpus_per_task: 12
+    gpus_per_node: ${trainer.devices}
+    cpus_per_gpu: null
+    comment: laion
+    name: hamer
+    setup:
+      - module load cuda openmpi libfabric-aws
+      - export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7

hamer/configs_hydra/paths/default.yaml

@@ -0,0 +1,18 @@
+# path to root directory
+# this requires PROJECT_ROOT environment variable to exist
+# PROJECT_ROOT is inferred and set by pyrootutils package in `train.py` and `eval.py`
+root_dir: ${oc.env:PROJECT_ROOT}
+
+# path to data directory
+data_dir: ${paths.root_dir}/data/
+
+# path to logging directory
+log_dir: logs/
+
+# path to output directory, created dynamically by hydra
+# path generation pattern is specified in `configs/hydra/default.yaml`
+# use it to store all files generated during the run, like ckpts and metrics
+output_dir: ${hydra:runtime.output_dir}
+
+# path to working directory
+work_dir: ${hydra:runtime.cwd}

hamer/configs_hydra/train.yaml

@@ -0,0 +1,47 @@
+# @package _global_
+
+# specify here default configuration
+# order of defaults determines the order in which configs override each other
+defaults:
+  - _self_
+  - data: mix_all.yaml
+  - trainer: ddp.yaml
+  - paths: default.yaml
+  - extras: default.yaml
+  - hydra: default.yaml
+
+  # experiment configs allow for version control of specific hyperparameters
+  # e.g. best hyperparameters for given model and datamodule
+  - experiment: null
+  - texture_exp: null
+
+  # optional local config for machine/user specific settings
+  # it's optional since it doesn't need to exist and is excluded from version control
+  - optional launcher: local.yaml
+  # - optional launcher: slurm.yaml
+
+  # debugging config (enable through command line, e.g. `python train.py debug=default)
+  - debug: null
+
+# task name, determines output directory path
+task_name: "train"
+
+# tags to help you identify your experiments
+# you can overwrite this in experiment configs
+# overwrite from command line with `python train.py tags="[first_tag, second_tag]"`
+# appending lists from command line is currently not supported :(
+# https://github.com/facebookresearch/hydra/issues/1547
+tags: ["dev"]
+
+# set False to skip model training
+train: True
+
+# evaluate on test set, using best model weights achieved during training
+# lightning chooses best weights based on the metric specified in checkpoint callback
+test: False
+
+# simply provide checkpoint path to resume training
+ckpt_path: null
+
+# seed for random number generators in pytorch, numpy and python.random
+seed: null

hamer/configs_hydra/trainer/cpu.yaml

@@ -0,0 +1,6 @@
+defaults:
+  - default.yaml
+  - default_hamer.yaml
+
+accelerator: cpu
+devices: 1

hamer/configs_hydra/trainer/ddp.yaml

@@ -0,0 +1,14 @@
+defaults:
+  - default.yaml
+  - default_hamer.yaml
+
+# use "ddp_spawn" instead of "ddp",
+# it's slower but normal "ddp" currently doesn't work ideally with hydra
+# https://github.com/facebookresearch/hydra/issues/2070
+# https://pytorch-lightning.readthedocs.io/en/latest/accelerators/gpu_intermediate.html#distributed-data-parallel-spawn
+strategy: ddp
+
+accelerator: gpu
+devices: 8
+num_nodes: 1
+sync_batchnorm: True

hamer/configs_hydra/trainer/default.yaml

@@ -0,0 +1,10 @@
+_target_: pytorch_lightning.Trainer
+
+default_root_dir: ${paths.output_dir}
+
+accelerator: cpu
+devices: 1
+
+# set True to to ensure deterministic results
+# makes training slower but gives more reproducibility than just setting seeds
+deterministic: False

hamer/configs_hydra/trainer/default_hamer.yaml

@@ -0,0 +1,8 @@
+num_sanity_val_steps: 0
+log_every_n_steps: ${GENERAL.LOG_STEPS}
+val_check_interval: ${GENERAL.VAL_STEPS}
+precision: 16
+max_steps: ${GENERAL.TOTAL_STEPS}
+# move_metrics_to_cpu: True
+limit_val_batches: 1
+# track_grad_norm: -1

hamer/configs_hydra/trainer/gpu.yaml

@@ -0,0 +1,6 @@
+defaults:
+  - default.yaml
+  - default_hamer.yaml
+
+accelerator: gpu
+devices: 1

hamer/configs_hydra/trainer/mps.yaml

@@ -0,0 +1,6 @@
+defaults:
+  - default.yaml
+  - default_hamer.yaml
+
+accelerator: mps
+devices: 1

hamer/datasets/__init__.py

@@ -0,0 +1,56 @@
+from typing import Dict, Optional
+
+import torch
+import numpy as np
+import pytorch_lightning as pl
+from yacs.config import CfgNode
+
+from ..configs import to_lower
+from .dataset import Dataset
+
+class HAMERDataModule(pl.LightningDataModule):
+
+    def __init__(self, cfg: CfgNode, dataset_cfg: CfgNode) -> None:
+        """
+        Initialize LightningDataModule for HAMER training
+        Args:
+            cfg (CfgNode): Config file as a yacs CfgNode containing necessary dataset info.
+            dataset_cfg (CfgNode): Dataset configuration file
+        """
+        super().__init__()
+        self.cfg = cfg
+        self.dataset_cfg = dataset_cfg
+        self.train_dataset = None
+        self.val_dataset = None
+        self.test_dataset = None
+        self.mocap_dataset = None
+
+    def setup(self, stage: Optional[str] = None) -> None:
+        """
+        Load datasets necessary for training
+        Args:
+            cfg (CfgNode): Config file as a yacs CfgNode containing necessary dataset info.
+        """
+        if self.train_dataset == None:
+            self.train_dataset = MixedWebDataset(self.cfg, self.dataset_cfg, train=True).with_epoch(100_000).shuffle(4000)
+            self.val_dataset = MixedWebDataset(self.cfg, self.dataset_cfg, train=False).shuffle(4000)
+            self.mocap_dataset = MoCapDataset(**to_lower(self.dataset_cfg[self.cfg.DATASETS.MOCAP]))
+
+    def train_dataloader(self) -> Dict:
+        """
+        Setup training data loader.
+        Returns:
+            Dict: Dictionary containing image and mocap data dataloaders
+        """
+        train_dataloader = torch.utils.data.DataLoader(self.train_dataset, self.cfg.TRAIN.BATCH_SIZE, drop_last=True, num_workers=self.cfg.GENERAL.NUM_WORKERS, prefetch_factor=self.cfg.GENERAL.PREFETCH_FACTOR)
+        mocap_dataloader = torch.utils.data.DataLoader(self.mocap_dataset, self.cfg.TRAIN.NUM_TRAIN_SAMPLES * self.cfg.TRAIN.BATCH_SIZE, shuffle=True, drop_last=True, num_workers=1)
+        return {'img': train_dataloader, 'mocap': mocap_dataloader}
+
+    def val_dataloader(self) -> torch.utils.data.DataLoader:
+        """
+        Setup val data loader.
+        Returns:
+            torch.utils.data.DataLoader: Validation dataloader  
+        """
+        val_dataloader = torch.utils.data.DataLoader(self.val_dataset, self.cfg.TRAIN.BATCH_SIZE, drop_last=True, num_workers=self.cfg.GENERAL.NUM_WORKERS)
+        return val_dataloader

hamer/datasets/dataset.py

@@ -0,0 +1,27 @@
+"""
+This file contains the defition of the base Dataset class.
+"""
+
+class DatasetRegistration(type):
+    """
+    Metaclass for registering different datasets
+    """
+    def __init__(cls, name, bases, nmspc):
+        super().__init__(name, bases, nmspc)
+        if not hasattr(cls, 'registry'):
+            cls.registry = dict()
+        cls.registry[name] = cls
+
+    # Metamethods, called on class objects:
+    def __iter__(cls):
+        return iter(cls.registry)
+
+    def __str__(cls):
+        return str(cls.registry)
+
+class Dataset(metaclass=DatasetRegistration):
+    """
+    Base Dataset class
+    """
+    def __init__(self, *args, **kwargs):
+        pass

hamer/datasets/image_dataset.py

@@ -0,0 +1,275 @@
+import copy
+import os
+import numpy as np
+import torch
+from typing import List
+from yacs.config import CfgNode
+import braceexpand
+import cv2
+
+from .dataset import Dataset
+from .utils import get_example, expand_to_aspect_ratio
+
+def expand(s):
+    return os.path.expanduser(os.path.expandvars(s))
+def expand_urls(urls: str|List[str]):
+    if isinstance(urls, str):
+        urls = [urls]
+    urls = [u for url in urls for u in braceexpand.braceexpand(expand(url))]
+    return urls
+
+FLIP_KEYPOINT_PERMUTATION = [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20]
+
+DEFAULT_MEAN = 255. * np.array([0.485, 0.456, 0.406])
+DEFAULT_STD = 255. * np.array([0.229, 0.224, 0.225])
+DEFAULT_IMG_SIZE = 256
+
+class ImageDataset(Dataset):
+
+    @staticmethod
+    def load_tars_as_webdataset(cfg: CfgNode, urls: str|List[str], train: bool,
+            resampled=False,
+            epoch_size=None,
+            cache_dir=None,
+            **kwargs) -> Dataset:
+        """
+        Loads the dataset from a webdataset tar file.
+        """
+
+        IMG_SIZE = cfg.MODEL.IMAGE_SIZE
+        BBOX_SHAPE = cfg.MODEL.get('BBOX_SHAPE', None)
+        MEAN = 255. * np.array(cfg.MODEL.IMAGE_MEAN)
+        STD = 255. * np.array(cfg.MODEL.IMAGE_STD)
+
+        def split_data(source):
+            for item in source:
+                datas = item['data.pyd']
+                for data in datas:
+                    if 'detection.npz' in item:
+                        det_idx = data['extra_info']['detection_npz_idx']
+                        mask = item['detection.npz']['masks'][det_idx]
+                    else:
+                        mask = np.ones_like(item['jpg'][:,:,0], dtype=bool)
+                    yield {
+                        '__key__': item['__key__'],
+                        'jpg': item['jpg'],
+                        'data.pyd': data,
+                        'mask': mask,
+                    }
+
+        def suppress_bad_kps(item, thresh=0.0):
+            if thresh > 0:
+                kp2d = item['data.pyd']['keypoints_2d']
+                kp2d_conf = np.where(kp2d[:, 2] < thresh, 0.0, kp2d[:, 2])
+                item['data.pyd']['keypoints_2d'] = np.concatenate([kp2d[:,:2], kp2d_conf[:,None]], axis=1)
+            return item
+
+        def filter_numkp(item, numkp=4, thresh=0.0):
+            kp_conf = item['data.pyd']['keypoints_2d'][:, 2]
+            return (kp_conf > thresh).sum() > numkp
+
+        def filter_reproj_error(item, thresh=10**4.5):
+            losses = item['data.pyd'].get('extra_info', {}).get('fitting_loss', np.array({})).item()
+            reproj_loss = losses.get('reprojection_loss', None)
+            return reproj_loss is None or reproj_loss < thresh
+
+        def filter_bbox_size(item, thresh=1):
+            bbox_size_min = item['data.pyd']['scale'].min().item() * 200.
+            return bbox_size_min > thresh
+
+        def filter_no_poses(item):
+            return (item['data.pyd']['has_hand_pose'] > 0)
+
+        def supress_bad_betas(item, thresh=3):
+            has_betas = item['data.pyd']['has_betas']
+            if thresh > 0 and has_betas:
+                betas_abs = np.abs(item['data.pyd']['betas'])
+                if (betas_abs > thresh).any():
+                    item['data.pyd']['has_betas'] = False
+            return item
+
+        def supress_bad_poses(item):
+            has_hand_pose = item['data.pyd']['has_hand_pose']
+            if has_hand_pose:
+                hand_pose = item['data.pyd']['hand_pose']
+                pose_is_probable = poses_check_probable(torch.from_numpy(hand_pose)[None, 3:], amass_poses_hist100_smooth).item()
+                if not pose_is_probable:
+                    item['data.pyd']['has_hand_pose'] = False
+            return item
+
+        def poses_betas_simultaneous(item):
+            # We either have both hand_pose and betas, or neither
+            has_betas = item['data.pyd']['has_betas']
+            has_hand_pose = item['data.pyd']['has_hand_pose']
+            item['data.pyd']['has_betas'] = item['data.pyd']['has_hand_pose'] = np.array(float((has_hand_pose>0) and (has_betas>0)))
+            return item
+
+        def set_betas_for_reg(item):
+            # Always have betas set to true
+            has_betas = item['data.pyd']['has_betas']
+            betas = item['data.pyd']['betas']
+
+            if not (has_betas>0):
+                item['data.pyd']['has_betas'] = np.array(float((True)))
+                item['data.pyd']['betas'] = betas * 0
+            return item
+
+        # Load the dataset
+        if epoch_size is not None:
+            resampled = True
+        #corrupt_filter = lambda sample: (sample['__key__'] not in CORRUPT_KEYS)
+        import webdataset as wds
+        dataset = wds.WebDataset(expand_urls(urls),
+                                nodesplitter=wds.split_by_node,
+                                shardshuffle=True,
+                                resampled=resampled,
+                                cache_dir=cache_dir,
+                              ) #.select(corrupt_filter)
+        if train:
+            dataset = dataset.shuffle(100)
+        dataset = dataset.decode('rgb8').rename(jpg='jpg;jpeg;png')
+
+        # Process the dataset
+        dataset = dataset.compose(split_data)
+
+        # Filter/clean the dataset
+        SUPPRESS_KP_CONF_THRESH = cfg.DATASETS.get('SUPPRESS_KP_CONF_THRESH', 0.0)
+        SUPPRESS_BETAS_THRESH = cfg.DATASETS.get('SUPPRESS_BETAS_THRESH', 0.0)
+        SUPPRESS_BAD_POSES = cfg.DATASETS.get('SUPPRESS_BAD_POSES', False)
+        POSES_BETAS_SIMULTANEOUS = cfg.DATASETS.get('POSES_BETAS_SIMULTANEOUS', False)
+        BETAS_REG = cfg.DATASETS.get('BETAS_REG', False)
+        FILTER_NO_POSES = cfg.DATASETS.get('FILTER_NO_POSES', False)
+        FILTER_NUM_KP = cfg.DATASETS.get('FILTER_NUM_KP', 4)
+        FILTER_NUM_KP_THRESH = cfg.DATASETS.get('FILTER_NUM_KP_THRESH', 0.0)
+        FILTER_REPROJ_THRESH = cfg.DATASETS.get('FILTER_REPROJ_THRESH', 0.0)
+        FILTER_MIN_BBOX_SIZE = cfg.DATASETS.get('FILTER_MIN_BBOX_SIZE', 0.0)
+        if SUPPRESS_KP_CONF_THRESH > 0:
+            dataset = dataset.map(lambda x: suppress_bad_kps(x, thresh=SUPPRESS_KP_CONF_THRESH))
+        if SUPPRESS_BETAS_THRESH > 0:
+            dataset = dataset.map(lambda x: supress_bad_betas(x, thresh=SUPPRESS_BETAS_THRESH))
+        if SUPPRESS_BAD_POSES:
+            dataset = dataset.map(lambda x: supress_bad_poses(x))
+        if POSES_BETAS_SIMULTANEOUS:
+            dataset = dataset.map(lambda x: poses_betas_simultaneous(x))
+        if FILTER_NO_POSES:
+            dataset = dataset.select(lambda x: filter_no_poses(x))
+        if FILTER_NUM_KP > 0:
+            dataset = dataset.select(lambda x: filter_numkp(x, numkp=FILTER_NUM_KP, thresh=FILTER_NUM_KP_THRESH))
+        if FILTER_REPROJ_THRESH > 0:
+            dataset = dataset.select(lambda x: filter_reproj_error(x, thresh=FILTER_REPROJ_THRESH))
+        if FILTER_MIN_BBOX_SIZE > 0:
+            dataset = dataset.select(lambda x: filter_bbox_size(x, thresh=FILTER_MIN_BBOX_SIZE))
+        if BETAS_REG:
+            dataset = dataset.map(lambda x: set_betas_for_reg(x))       # NOTE: Must be at the end
+
+        use_skimage_antialias = cfg.DATASETS.get('USE_SKIMAGE_ANTIALIAS', False)
+        border_mode = {
+            'constant': cv2.BORDER_CONSTANT,
+            'replicate': cv2.BORDER_REPLICATE,
+        }[cfg.DATASETS.get('BORDER_MODE', 'constant')]
+
+        # Process the dataset further
+        dataset = dataset.map(lambda x: ImageDataset.process_webdataset_tar_item(x, train,
+                                                        augm_config=cfg.DATASETS.CONFIG,
+                                                        MEAN=MEAN, STD=STD, IMG_SIZE=IMG_SIZE,
+                                                        BBOX_SHAPE=BBOX_SHAPE,
+                                                        use_skimage_antialias=use_skimage_antialias,
+                                                        border_mode=border_mode,
+                                                        ))
+        if epoch_size is not None:
+            dataset = dataset.with_epoch(epoch_size)
+
+        return dataset
+
+    @staticmethod
+    def process_webdataset_tar_item(item, train, 
+                                    augm_config=None, 
+                                    MEAN=DEFAULT_MEAN, 
+                                    STD=DEFAULT_STD, 
+                                    IMG_SIZE=DEFAULT_IMG_SIZE,
+                                    BBOX_SHAPE=None,
+                                    use_skimage_antialias=False,
+                                    border_mode=cv2.BORDER_CONSTANT,
+                                    ):
+        # Read data from item
+        key = item['__key__']
+        image = item['jpg']
+        data = item['data.pyd']
+        mask = item['mask']
+
+        keypoints_2d = data['keypoints_2d']
+        keypoints_3d = data['keypoints_3d']
+        center = data['center']
+        scale = data['scale']
+        hand_pose = data['hand_pose']
+        betas = data['betas']
+        right = data['right']
+        #right = True
+        has_hand_pose = data['has_hand_pose']
+        has_betas = data['has_betas']
+        # image_file = data['image_file']
+
+        # Process data
+        orig_keypoints_2d = keypoints_2d.copy()
+        center_x = center[0]
+        center_y = center[1]
+        bbox_size = expand_to_aspect_ratio(scale*200, target_aspect_ratio=BBOX_SHAPE).max()
+        if bbox_size < 1:
+            breakpoint()
+
+
+        mano_params = {'global_orient': hand_pose[:3],
+                    'hand_pose': hand_pose[3:],
+                    'betas': betas
+                    }
+
+        has_mano_params = {'global_orient': has_hand_pose,
+                        'hand_pose': has_hand_pose,
+                        'betas': has_betas
+                        }
+
+        mano_params_is_axis_angle = {'global_orient': True,
+                                    'hand_pose': True,
+                                    'betas': False
+                                    }
+
+        augm_config = copy.deepcopy(augm_config)
+        # Crop image and (possibly) perform data augmentation
+        img_rgba = np.concatenate([image, mask.astype(np.uint8)[:,:,None]*255], axis=2)
+        img_patch_rgba, keypoints_2d, keypoints_3d, mano_params, has_mano_params, img_size, trans = get_example(img_rgba,
+                                                                                                    center_x, center_y,
+                                                                                                    bbox_size, bbox_size,
+                                                                                                    keypoints_2d, keypoints_3d,
+                                                                                                    mano_params, has_mano_params,
+                                                                                                    FLIP_KEYPOINT_PERMUTATION,
+                                                                                                    IMG_SIZE, IMG_SIZE,
+                                                                                                    MEAN, STD, train, right, augm_config,
+                                                                                                    is_bgr=False, return_trans=True,
+                                                                                                    use_skimage_antialias=use_skimage_antialias,
+                                                                                                    border_mode=border_mode,
+                                                                                                    )
+        img_patch = img_patch_rgba[:3,:,:]
+        mask_patch = (img_patch_rgba[3,:,:] / 255.0).clip(0,1)
+        if (mask_patch < 0.5).all():
+            mask_patch = np.ones_like(mask_patch)
+
+        item = {}
+
+        item['img'] = img_patch
+        item['mask'] = mask_patch
+        # item['img_og'] = image
+        # item['mask_og'] = mask
+        item['keypoints_2d'] = keypoints_2d.astype(np.float32)
+        item['keypoints_3d'] = keypoints_3d.astype(np.float32)
+        item['orig_keypoints_2d'] = orig_keypoints_2d
+        item['box_center'] = center.copy()
+        item['box_size'] = bbox_size
+        item['img_size'] = 1.0 * img_size[::-1].copy()
+        item['mano_params'] = mano_params
+        item['has_mano_params'] = has_mano_params
+        item['mano_params_is_axis_angle'] = mano_params_is_axis_angle
+        item['_scale'] = scale
+        item['_trans'] = trans
+        item['imgname'] = key
+        # item['idx'] = idx
+        return item

hamer/datasets/json_dataset.py

@@ -0,0 +1,213 @@
+import copy
+import os
+import json
+import glob
+import numpy as np
+import torch
+from typing import Any, Dict, List
+from yacs.config import CfgNode
+import braceexpand
+import cv2
+
+from .dataset import Dataset
+from .utils import get_example, expand_to_aspect_ratio
+from .smplh_prob_filter import poses_check_probable, load_amass_hist_smooth
+
+def expand(s):
+    return os.path.expanduser(os.path.expandvars(s))
+def expand_urls(urls: str|List[str]):
+    if isinstance(urls, str):
+        urls = [urls]
+    urls = [u for url in urls for u in braceexpand.braceexpand(expand(url))]
+    return urls
+
+AIC_TRAIN_CORRUPT_KEYS = {
+    '0a047f0124ae48f8eee15a9506ce1449ee1ba669',
+    '1a703aa174450c02fbc9cfbf578a5435ef403689',
+    '0394e6dc4df78042929b891dbc24f0fd7ffb6b6d',
+    '5c032b9626e410441544c7669123ecc4ae077058',
+    'ca018a7b4c5f53494006ebeeff9b4c0917a55f07',
+    '4a77adb695bef75a5d34c04d589baf646fe2ba35',
+    'a0689017b1065c664daef4ae2d14ea03d543217e',
+    '39596a45cbd21bed4a5f9c2342505532f8ec5cbb',
+    '3d33283b40610d87db660b62982f797d50a7366b',
+}
+CORRUPT_KEYS = {
+    *{f'aic-train/{k}' for k in AIC_TRAIN_CORRUPT_KEYS},
+    *{f'aic-train-vitpose/{k}' for k in AIC_TRAIN_CORRUPT_KEYS},
+}
+
+FLIP_KEYPOINT_PERMUTATION = [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20]
+
+DEFAULT_MEAN = 255. * np.array([0.485, 0.456, 0.406])
+DEFAULT_STD = 255. * np.array([0.229, 0.224, 0.225])
+DEFAULT_IMG_SIZE = 256
+
+class JsonDataset(Dataset):
+
+    def __init__(self,
+                 cfg: CfgNode,
+                 dataset_file: str,
+                 img_dir: str,
+                 right: bool,
+                 train: bool = False,
+                 prune: Dict[str, Any] = {},
+                 **kwargs):
+        """
+        Dataset class used for loading images and corresponding annotations.
+        Args:
+            cfg (CfgNode): Model config file.
+            dataset_file (str): Path to npz file containing dataset info.
+            img_dir (str): Path to image folder.
+            train (bool): Whether it is for training or not (enables data augmentation).
+        """
+        super(JsonDataset, self).__init__()
+        self.train = train
+        self.cfg = cfg
+
+        self.img_size = cfg.MODEL.IMAGE_SIZE
+        self.mean = 255. * np.array(self.cfg.MODEL.IMAGE_MEAN)
+        self.std = 255. * np.array(self.cfg.MODEL.IMAGE_STD)
+
+        self.img_dir = img_dir
+        boxes = np.array(json.load(open(dataset_file, 'rb')))
+
+        self.imgname = glob.glob(os.path.join(self.img_dir,'*.jpg'))
+        self.imgname.sort()
+
+        self.flip_keypoint_permutation = copy.copy(FLIP_KEYPOINT_PERMUTATION)
+
+        num_pose = 3 * (self.cfg.MANO.NUM_HAND_JOINTS + 1)
+
+        # Bounding boxes are assumed to be in the center and scale format
+        boxes = boxes.astype(np.float32)
+        self.center = (boxes[:, 2:4] + boxes[:, 0:2]) / 2.0
+        self.scale = 2 * (boxes[:, 2:4] - boxes[:, 0:2]) / 200.0
+        self.personid = np.arange(len(boxes), dtype=np.int32)
+        if right:
+            self.right = np.ones(len(self.imgname), dtype=np.float32)
+        else:
+            self.right = np.zeros(len(self.imgname), dtype=np.float32)
+        assert self.scale.shape == (len(self.center), 2)
+
+        # Get gt SMPLX parameters, if available
+        try:
+            self.hand_pose = self.data['hand_pose'].astype(np.float32)
+            self.has_hand_pose = self.data['has_hand_pose'].astype(np.float32)
+        except:
+            self.hand_pose = np.zeros((len(self.imgname), num_pose), dtype=np.float32)
+            self.has_hand_pose = np.zeros(len(self.imgname), dtype=np.float32)
+        try:
+            self.betas = self.data['betas'].astype(np.float32)
+            self.has_betas = self.data['has_betas'].astype(np.float32)
+        except:
+            self.betas = np.zeros((len(self.imgname), 10), dtype=np.float32)
+            self.has_betas = np.zeros(len(self.imgname), dtype=np.float32)
+
+        # Try to get 2d keypoints, if available
+        try:
+            hand_keypoints_2d = self.data['hand_keypoints_2d']
+        except:
+            hand_keypoints_2d = np.zeros((len(self.center), 21, 3))
+        ## Try to get extra 2d keypoints, if available
+        #try:
+        #    extra_keypoints_2d = self.data['extra_keypoints_2d']
+        #except KeyError:
+        #    extra_keypoints_2d = np.zeros((len(self.center), 19, 3))
+
+        #self.keypoints_2d = np.concatenate((hand_keypoints_2d, extra_keypoints_2d), axis=1).astype(np.float32)
+        self.keypoints_2d = hand_keypoints_2d
+
+        # Try to get 3d keypoints, if available
+        try:
+            hand_keypoints_3d = self.data['hand_keypoints_3d'].astype(np.float32)
+        except:
+            hand_keypoints_3d = np.zeros((len(self.center), 21, 4), dtype=np.float32)
+        ## Try to get extra 3d keypoints, if available
+        #try:
+        #    extra_keypoints_3d = self.data['extra_keypoints_3d'].astype(np.float32)
+        #except KeyError:
+        #    extra_keypoints_3d = np.zeros((len(self.center), 19, 4), dtype=np.float32)
+
+        self.keypoints_3d = hand_keypoints_3d
+
+        #body_keypoints_3d[:, [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14], -1] = 0
+
+        #self.keypoints_3d = np.concatenate((body_keypoints_3d, extra_keypoints_3d), axis=1).astype(np.float32)
+
+    def __len__(self) -> int:
+        return len(self.scale)
+
+    def __getitem__(self, idx: int) -> Dict:
+        """
+        Returns an example from the dataset.
+        """
+        try:
+            image_file = self.imgname[idx].decode('utf-8')
+        except AttributeError:
+            image_file = self.imgname[idx]
+        keypoints_2d = self.keypoints_2d[idx].copy()
+        keypoints_3d = self.keypoints_3d[idx].copy()
+
+        center = self.center[idx].copy()
+        center_x = center[0]
+        center_y = center[1]
+        scale = self.scale[idx]
+        right = self.right[idx].copy()
+        BBOX_SHAPE = self.cfg.MODEL.get('BBOX_SHAPE', None)
+        #bbox_size = expand_to_aspect_ratio(scale*200, target_aspect_ratio=BBOX_SHAPE).max()
+        bbox_size = ((scale*200).max())
+        bbox_expand_factor = bbox_size / ((scale*200).max())
+        hand_pose = self.hand_pose[idx].copy().astype(np.float32)
+        betas = self.betas[idx].copy().astype(np.float32)
+
+        has_hand_pose = self.has_hand_pose[idx].copy()
+        has_betas = self.has_betas[idx].copy()
+
+        mano_params = {'global_orient': hand_pose[:3],
+                       'hand_pose': hand_pose[3:],
+                       'betas': betas
+                      }
+
+        has_mano_params = {'global_orient': has_hand_pose,
+                           'hand_pose': has_hand_pose,
+                           'betas': has_betas
+                           }
+
+        mano_params_is_axis_angle = {'global_orient': True,
+                                     'hand_pose': True,
+                                     'betas': False
+                                    }
+
+        augm_config = self.cfg.DATASETS.CONFIG
+        # Crop image and (possibly) perform data augmentation
+        img_patch, keypoints_2d, keypoints_3d, mano_params, has_mano_params, img_size = get_example(image_file,
+                                                                                                    center_x, center_y,
+                                                                                                    bbox_size, bbox_size,
+                                                                                                    keypoints_2d, keypoints_3d,
+                                                                                                    mano_params, has_mano_params,
+                                                                                                    self.flip_keypoint_permutation,
+                                                                                                    self.img_size, self.img_size,
+                                                                                                    self.mean, self.std, self.train, right, augm_config)
+
+        item = {}
+        # These are the keypoints in the original image coordinates (before cropping)
+        orig_keypoints_2d = self.keypoints_2d[idx].copy()
+
+        item['img'] = img_patch
+        item['keypoints_2d'] = keypoints_2d.astype(np.float32)
+        item['keypoints_3d'] = keypoints_3d.astype(np.float32)
+        item['orig_keypoints_2d'] = orig_keypoints_2d
+        item['box_center'] = self.center[idx].copy()
+        item['box_size'] = bbox_size
+        item['bbox_expand_factor'] = bbox_expand_factor
+        item['img_size'] = 1.0 * img_size[::-1].copy()
+        item['mano_params'] = mano_params
+        item['has_mano_params'] = has_mano_params
+        item['mano_params_is_axis_angle'] = mano_params_is_axis_angle
+        item['imgname'] = image_file
+        item['personid'] = int(self.personid[idx])
+        item['idx'] = idx
+        item['_scale'] = scale
+        item['right'] = self.right[idx].copy()
+        return item

hamer/datasets/mocap_dataset.py

@@ -0,0 +1,25 @@
+import numpy as np
+from typing import Dict
+
+class MoCapDataset:
+
+    def __init__(self, dataset_file: str):
+        """
+        Dataset class used for loading a dataset of unpaired MANO parameter annotations
+        Args:
+            cfg (CfgNode): Model config file.
+            dataset_file (str): Path to npz file containing dataset info.
+        """
+        data = np.load(dataset_file)
+        self.pose = data['hand_pose'].astype(np.float32)[:, 3:]
+        self.betas = data['betas'].astype(np.float32)
+        self.length = len(self.pose)
+
+    def __getitem__(self, idx: int) -> Dict:
+        pose = self.pose[idx].copy()
+        betas = self.betas[idx].copy()
+        item = {'hand_pose': pose, 'betas': betas}
+        return item
+
+    def __len__(self) -> int:
+        return self.length

hamer/datasets/utils.py

@@ -0,0 +1,993 @@
+"""
+Parts of the code are taken or adapted from
+https://github.com/mkocabas/EpipolarPose/blob/master/lib/utils/img_utils.py
+"""
+import torch
+import numpy as np
+from skimage.transform import rotate, resize
+from skimage.filters import gaussian
+import random
+import cv2
+from typing import List, Dict, Tuple
+from yacs.config import CfgNode
+
+def expand_to_aspect_ratio(input_shape, target_aspect_ratio=None):
+    """Increase the size of the bounding box to match the target shape."""
+    if target_aspect_ratio is None:
+        return input_shape
+
+    try:
+        w , h = input_shape
+    except (ValueError, TypeError):
+        return input_shape
+
+    w_t, h_t = target_aspect_ratio
+    if h / w < h_t / w_t:
+        h_new = max(w * h_t / w_t, h)
+        w_new = w
+    else:
+        h_new = h
+        w_new = max(h * w_t / h_t, w)
+    if h_new < h or w_new < w:
+        breakpoint()
+    return np.array([w_new, h_new])
+
+def do_augmentation(aug_config: CfgNode) -> Tuple:
+    """
+    Compute random augmentation parameters.
+    Args:
+        aug_config (CfgNode): Config containing augmentation parameters.
+    Returns:
+        scale (float): Box rescaling factor.
+        rot (float): Random image rotation.
+        do_flip (bool): Whether to flip image or not.
+        do_extreme_crop (bool): Whether to apply extreme cropping (as proposed in EFT).
+        color_scale (List): Color rescaling factor
+        tx (float): Random translation along the x axis.
+        ty (float): Random translation along the y axis. 
+    """
+
+    tx = np.clip(np.random.randn(), -1.0, 1.0) * aug_config.TRANS_FACTOR
+    ty = np.clip(np.random.randn(), -1.0, 1.0) * aug_config.TRANS_FACTOR
+    scale = np.clip(np.random.randn(), -1.0, 1.0) * aug_config.SCALE_FACTOR + 1.0
+    rot = np.clip(np.random.randn(), -2.0,
+                  2.0) * aug_config.ROT_FACTOR if random.random() <= aug_config.ROT_AUG_RATE else 0
+    do_flip = aug_config.DO_FLIP and random.random() <= aug_config.FLIP_AUG_RATE
+    do_extreme_crop = random.random() <= aug_config.EXTREME_CROP_AUG_RATE
+    extreme_crop_lvl = aug_config.get('EXTREME_CROP_AUG_LEVEL', 0)
+    # extreme_crop_lvl = 0
+    c_up = 1.0 + aug_config.COLOR_SCALE
+    c_low = 1.0 - aug_config.COLOR_SCALE
+    color_scale = [random.uniform(c_low, c_up), random.uniform(c_low, c_up), random.uniform(c_low, c_up)]
+    return scale, rot, do_flip, do_extreme_crop, extreme_crop_lvl, color_scale, tx, ty
+
+def rotate_2d(pt_2d: np.array, rot_rad: float) -> np.array:
+    """
+    Rotate a 2D point on the x-y plane.
+    Args:
+        pt_2d (np.array): Input 2D point with shape (2,).
+        rot_rad (float): Rotation angle
+    Returns:
+        np.array: Rotated 2D point.
+    """
+    x = pt_2d[0]
+    y = pt_2d[1]
+    sn, cs = np.sin(rot_rad), np.cos(rot_rad)
+    xx = x * cs - y * sn
+    yy = x * sn + y * cs
+    return np.array([xx, yy], dtype=np.float32)
+
+
+def gen_trans_from_patch_cv(c_x: float, c_y: float,
+                            src_width: float, src_height: float,
+                            dst_width: float, dst_height: float,
+                            scale: float, rot: float) -> np.array:
+    """
+    Create transformation matrix for the bounding box crop.
+    Args:
+        c_x (float): Bounding box center x coordinate in the original image.
+        c_y (float): Bounding box center y coordinate in the original image.
+        src_width (float): Bounding box width.
+        src_height (float): Bounding box height.
+        dst_width (float): Output box width.
+        dst_height (float): Output box height.
+        scale (float): Rescaling factor for the bounding box (augmentation).
+        rot (float): Random rotation applied to the box.
+    Returns:
+        trans (np.array): Target geometric transformation.
+    """
+    # augment size with scale
+    src_w = src_width * scale
+    src_h = src_height * scale
+    src_center = np.zeros(2)
+    src_center[0] = c_x
+    src_center[1] = c_y
+    # augment rotation
+    rot_rad = np.pi * rot / 180
+    src_downdir = rotate_2d(np.array([0, src_h * 0.5], dtype=np.float32), rot_rad)
+    src_rightdir = rotate_2d(np.array([src_w * 0.5, 0], dtype=np.float32), rot_rad)
+
+    dst_w = dst_width
+    dst_h = dst_height
+    dst_center = np.array([dst_w * 0.5, dst_h * 0.5], dtype=np.float32)
+    dst_downdir = np.array([0, dst_h * 0.5], dtype=np.float32)
+    dst_rightdir = np.array([dst_w * 0.5, 0], dtype=np.float32)
+
+    src = np.zeros((3, 2), dtype=np.float32)
+    src[0, :] = src_center
+    src[1, :] = src_center + src_downdir
+    src[2, :] = src_center + src_rightdir
+
+    dst = np.zeros((3, 2), dtype=np.float32)
+    dst[0, :] = dst_center
+    dst[1, :] = dst_center + dst_downdir
+    dst[2, :] = dst_center + dst_rightdir
+
+    trans = cv2.getAffineTransform(np.float32(src), np.float32(dst))
+
+    return trans
+
+
+def trans_point2d(pt_2d: np.array, trans: np.array):
+    """
+    Transform a 2D point using translation matrix trans.
+    Args:
+        pt_2d (np.array): Input 2D point with shape (2,).
+        trans (np.array): Transformation matrix.
+    Returns:
+        np.array: Transformed 2D point.
+    """
+    src_pt = np.array([pt_2d[0], pt_2d[1], 1.]).T
+    dst_pt = np.dot(trans, src_pt)
+    return dst_pt[0:2]
+
+def get_transform(center, scale, res, rot=0):
+    """Generate transformation matrix."""
+    """Taken from PARE: https://github.com/mkocabas/PARE/blob/6e0caca86c6ab49ff80014b661350958e5b72fd8/pare/utils/image_utils.py"""
+    h = 200 * scale
+    t = np.zeros((3, 3))
+    t[0, 0] = float(res[1]) / h
+    t[1, 1] = float(res[0]) / h
+    t[0, 2] = res[1] * (-float(center[0]) / h + .5)
+    t[1, 2] = res[0] * (-float(center[1]) / h + .5)
+    t[2, 2] = 1
+    if not rot == 0:
+        rot = -rot  # To match direction of rotation from cropping
+        rot_mat = np.zeros((3, 3))
+        rot_rad = rot * np.pi / 180
+        sn, cs = np.sin(rot_rad), np.cos(rot_rad)
+        rot_mat[0, :2] = [cs, -sn]
+        rot_mat[1, :2] = [sn, cs]
+        rot_mat[2, 2] = 1
+        # Need to rotate around center
+        t_mat = np.eye(3)
+        t_mat[0, 2] = -res[1] / 2
+        t_mat[1, 2] = -res[0] / 2
+        t_inv = t_mat.copy()
+        t_inv[:2, 2] *= -1
+        t = np.dot(t_inv, np.dot(rot_mat, np.dot(t_mat, t)))
+    return t
+
+
+def transform(pt, center, scale, res, invert=0, rot=0, as_int=True):
+    """Transform pixel location to different reference."""
+    """Taken from PARE: https://github.com/mkocabas/PARE/blob/6e0caca86c6ab49ff80014b661350958e5b72fd8/pare/utils/image_utils.py"""
+    t = get_transform(center, scale, res, rot=rot)
+    if invert:
+        t = np.linalg.inv(t)
+    new_pt = np.array([pt[0] - 1, pt[1] - 1, 1.]).T
+    new_pt = np.dot(t, new_pt)
+    if as_int:
+        new_pt = new_pt.astype(int)
+    return new_pt[:2] + 1
+
+def crop_img(img, ul, br, border_mode=cv2.BORDER_CONSTANT, border_value=0):
+    c_x = (ul[0] + br[0])/2
+    c_y = (ul[1] + br[1])/2
+    bb_width = patch_width = br[0] - ul[0]
+    bb_height = patch_height = br[1] - ul[1]
+    trans = gen_trans_from_patch_cv(c_x, c_y, bb_width, bb_height, patch_width, patch_height, 1.0, 0)
+    img_patch = cv2.warpAffine(img, trans, (int(patch_width), int(patch_height)), 
+                                flags=cv2.INTER_LINEAR, 
+                                borderMode=border_mode,
+                                borderValue=border_value
+                        )
+    
+    # Force borderValue=cv2.BORDER_CONSTANT for alpha channel
+    if (img.shape[2] == 4) and (border_mode != cv2.BORDER_CONSTANT):
+        img_patch[:,:,3] = cv2.warpAffine(img[:,:,3], trans, (int(patch_width), int(patch_height)), 
+                                            flags=cv2.INTER_LINEAR, 
+                                            borderMode=cv2.BORDER_CONSTANT,
+                            )
+
+    return img_patch
+
+def generate_image_patch_skimage(img: np.array, c_x: float, c_y: float,
+                                 bb_width: float, bb_height: float,
+                                 patch_width: float, patch_height: float,
+                                 do_flip: bool, scale: float, rot: float,
+                                 border_mode=cv2.BORDER_CONSTANT, border_value=0) -> Tuple[np.array, np.array]:
+    """
+    Crop image according to the supplied bounding box.
+    Args:
+        img (np.array): Input image of shape (H, W, 3)
+        c_x (float): Bounding box center x coordinate in the original image.
+        c_y (float): Bounding box center y coordinate in the original image.
+        bb_width (float): Bounding box width.
+        bb_height (float): Bounding box height.
+        patch_width (float): Output box width.
+        patch_height (float): Output box height.
+        do_flip (bool): Whether to flip image or not.
+        scale (float): Rescaling factor for the bounding box (augmentation).
+        rot (float): Random rotation applied to the box.
+    Returns:
+        img_patch (np.array): Cropped image patch of shape (patch_height, patch_height, 3)
+        trans (np.array): Transformation matrix.
+    """
+    
+    img_height, img_width, img_channels = img.shape
+    if do_flip:
+       img = img[:, ::-1, :]
+       c_x = img_width - c_x - 1
+
+    trans = gen_trans_from_patch_cv(c_x, c_y, bb_width, bb_height, patch_width, patch_height, scale, rot)
+
+    #img_patch = cv2.warpAffine(img, trans, (int(patch_width), int(patch_height)), flags=cv2.INTER_LINEAR)
+
+    # skimage
+    center = np.zeros(2)
+    center[0] = c_x
+    center[1] = c_y
+    res = np.zeros(2)
+    res[0] = patch_width
+    res[1] = patch_height
+    # assumes bb_width = bb_height
+    # assumes patch_width = patch_height
+    assert bb_width == bb_height, f'{bb_width=} != {bb_height=}'
+    assert patch_width == patch_height, f'{patch_width=} != {patch_height=}'
+    scale1 = scale*bb_width/200.
+    
+    # Upper left point
+    ul = np.array(transform([1, 1], center, scale1, res, invert=1, as_int=False)) - 1
+    # Bottom right point
+    br = np.array(transform([res[0] + 1,
+                             res[1] + 1], center, scale1, res, invert=1, as_int=False)) - 1
+
+    # Padding so that when rotated proper amount of context is included
+    try:
+        pad = int(np.linalg.norm(br - ul) / 2 - float(br[1] - ul[1]) / 2) + 1
+    except:
+        breakpoint()
+    if not rot == 0:
+        ul -= pad
+        br += pad
+
+
+    if False:
+        # Old way of cropping image
+        ul_int = ul.astype(int)
+        br_int = br.astype(int)
+        new_shape = [br_int[1] - ul_int[1], br_int[0] - ul_int[0]]
+        if len(img.shape) > 2:
+            new_shape += [img.shape[2]]
+        new_img = np.zeros(new_shape)
+
+        # Range to fill new array
+        new_x = max(0, -ul_int[0]), min(br_int[0], len(img[0])) - ul_int[0]
+        new_y = max(0, -ul_int[1]), min(br_int[1], len(img)) - ul_int[1]
+        # Range to sample from original image
+        old_x = max(0, ul_int[0]), min(len(img[0]), br_int[0])
+        old_y = max(0, ul_int[1]), min(len(img), br_int[1])
+        new_img[new_y[0]:new_y[1], new_x[0]:new_x[1]] = img[old_y[0]:old_y[1],
+                                                        old_x[0]:old_x[1]]
+
+    # New way of cropping image
+    new_img = crop_img(img, ul, br, border_mode=border_mode, border_value=border_value).astype(np.float32)
+
+    # print(f'{new_img.shape=}')
+    # print(f'{new_img1.shape=}')
+    # print(f'{np.allclose(new_img, new_img1)=}')
+    # print(f'{img.dtype=}')
+
+
+    if not rot == 0:
+        # Remove padding
+
+        new_img = rotate(new_img, rot) # scipy.misc.imrotate(new_img, rot)
+        new_img = new_img[pad:-pad, pad:-pad]
+
+    if new_img.shape[0] < 1 or new_img.shape[1] < 1:
+        print(f'{img.shape=}')
+        print(f'{new_img.shape=}')
+        print(f'{ul=}')
+        print(f'{br=}')
+        print(f'{pad=}')
+        print(f'{rot=}')
+
+        breakpoint()
+
+    # resize image
+    new_img = resize(new_img, res) # scipy.misc.imresize(new_img, res)
+    
+    new_img = np.clip(new_img, 0, 255).astype(np.uint8)
+
+    return new_img, trans
+
+
+def generate_image_patch_cv2(img: np.array, c_x: float, c_y: float,
+                             bb_width: float, bb_height: float,
+                             patch_width: float, patch_height: float,
+                             do_flip: bool, scale: float, rot: float,
+                             border_mode=cv2.BORDER_CONSTANT, border_value=0) -> Tuple[np.array, np.array]:
+    """
+    Crop the input image and return the crop and the corresponding transformation matrix.
+    Args:
+        img (np.array): Input image of shape (H, W, 3)
+        c_x (float): Bounding box center x coordinate in the original image.
+        c_y (float): Bounding box center y coordinate in the original image.
+        bb_width (float): Bounding box width.
+        bb_height (float): Bounding box height.
+        patch_width (float): Output box width.
+        patch_height (float): Output box height.
+        do_flip (bool): Whether to flip image or not.
+        scale (float): Rescaling factor for the bounding box (augmentation).
+        rot (float): Random rotation applied to the box.
+    Returns:
+        img_patch (np.array): Cropped image patch of shape (patch_height, patch_height, 3)
+        trans (np.array): Transformation matrix.
+    """
+
+    img_height, img_width, img_channels = img.shape
+    if do_flip:
+        img = img[:, ::-1, :]
+        c_x = img_width - c_x - 1
+
+
+    trans = gen_trans_from_patch_cv(c_x, c_y, bb_width, bb_height, patch_width, patch_height, scale, rot)
+
+    img_patch = cv2.warpAffine(img, trans, (int(patch_width), int(patch_height)), 
+                        flags=cv2.INTER_LINEAR, 
+                        borderMode=border_mode,
+                        borderValue=border_value,
+                )
+    # Force borderValue=cv2.BORDER_CONSTANT for alpha channel
+    if (img.shape[2] == 4) and (border_mode != cv2.BORDER_CONSTANT):
+        img_patch[:,:,3] = cv2.warpAffine(img[:,:,3], trans, (int(patch_width), int(patch_height)), 
+                                            flags=cv2.INTER_LINEAR, 
+                                            borderMode=cv2.BORDER_CONSTANT,
+                            )
+
+    return img_patch, trans
+
+
+def convert_cvimg_to_tensor(cvimg: np.array):
+    """
+    Convert image from HWC to CHW format.
+    Args:
+        cvimg (np.array): Image of shape (H, W, 3) as loaded by OpenCV.
+    Returns:
+        np.array: Output image of shape (3, H, W).
+    """
+    # from h,w,c(OpenCV) to c,h,w
+    img = cvimg.copy()
+    img = np.transpose(img, (2, 0, 1))
+    # from int to float
+    img = img.astype(np.float32)
+    return img
+
+def fliplr_params(mano_params: Dict, has_mano_params: Dict) -> Tuple[Dict, Dict]:
+    """
+    Flip MANO parameters when flipping the image.
+    Args:
+        mano_params (Dict): MANO parameter annotations.
+        has_mano_params (Dict): Whether MANO annotations are valid.
+    Returns:
+        Dict, Dict: Flipped MANO parameters and valid flags.
+    """
+    global_orient = mano_params['global_orient'].copy()
+    hand_pose = mano_params['hand_pose'].copy()
+    betas = mano_params['betas'].copy()
+    has_global_orient = has_mano_params['global_orient'].copy()
+    has_hand_pose = has_mano_params['hand_pose'].copy()
+    has_betas = has_mano_params['betas'].copy()
+
+    global_orient[1::3] *= -1
+    global_orient[2::3] *= -1
+    hand_pose[1::3] *= -1
+    hand_pose[2::3] *= -1
+
+    mano_params = {'global_orient': global_orient.astype(np.float32),
+                   'hand_pose': hand_pose.astype(np.float32),
+                   'betas': betas.astype(np.float32)
+                  }
+
+    has_mano_params = {'global_orient': has_global_orient,
+                       'hand_pose': has_hand_pose,
+                       'betas': has_betas
+                      }
+
+    return mano_params, has_mano_params
+
+
+def fliplr_keypoints(joints: np.array, width: float, flip_permutation: List[int]) -> np.array:
+    """
+    Flip 2D or 3D keypoints.
+    Args:
+        joints (np.array): Array of shape (N, 3) or (N, 4) containing 2D or 3D keypoint locations and confidence.
+        flip_permutation (List): Permutation to apply after flipping.
+    Returns:
+        np.array: Flipped 2D or 3D keypoints with shape (N, 3) or (N, 4) respectively.
+    """
+    joints = joints.copy()
+    # Flip horizontal
+    joints[:, 0] = width - joints[:, 0] - 1
+    joints = joints[flip_permutation, :]
+
+    return joints
+
+def keypoint_3d_processing(keypoints_3d: np.array, flip_permutation: List[int], rot: float, do_flip: float) -> np.array:
+    """
+    Process 3D keypoints (rotation/flipping).
+    Args:
+        keypoints_3d (np.array): Input array of shape (N, 4) containing the 3D keypoints and confidence.
+        flip_permutation (List): Permutation to apply after flipping.
+        rot (float): Random rotation applied to the keypoints.
+        do_flip (bool): Whether to flip keypoints or not.
+    Returns:
+        np.array: Transformed 3D keypoints with shape (N, 4).
+    """
+    if do_flip:
+        keypoints_3d = fliplr_keypoints(keypoints_3d, 1, flip_permutation)
+    # in-plane rotation
+    rot_mat = np.eye(3)
+    if not rot == 0:
+        rot_rad = -rot * np.pi / 180
+        sn,cs = np.sin(rot_rad), np.cos(rot_rad)
+        rot_mat[0,:2] = [cs, -sn]
+        rot_mat[1,:2] = [sn, cs]
+    keypoints_3d[:, :-1] = np.einsum('ij,kj->ki', rot_mat, keypoints_3d[:, :-1])
+    # flip the x coordinates
+    keypoints_3d = keypoints_3d.astype('float32')
+    return keypoints_3d
+
+def rot_aa(aa: np.array, rot: float) -> np.array:
+    """
+    Rotate axis angle parameters.
+    Args:
+        aa (np.array): Axis-angle vector of shape (3,).
+        rot (np.array): Rotation angle in degrees.
+    Returns:
+        np.array: Rotated axis-angle vector.
+    """
+    # pose parameters
+    R = np.array([[np.cos(np.deg2rad(-rot)), -np.sin(np.deg2rad(-rot)), 0],
+                  [np.sin(np.deg2rad(-rot)), np.cos(np.deg2rad(-rot)), 0],
+                  [0, 0, 1]])
+    # find the rotation of the hand in camera frame
+    per_rdg, _ = cv2.Rodrigues(aa)
+    # apply the global rotation to the global orientation
+    resrot, _ = cv2.Rodrigues(np.dot(R,per_rdg))
+    aa = (resrot.T)[0]
+    return aa.astype(np.float32)
+
+def mano_param_processing(mano_params: Dict, has_mano_params: Dict, rot: float, do_flip: bool) -> Tuple[Dict, Dict]:
+    """
+    Apply random augmentations to the MANO parameters.
+    Args:
+        mano_params (Dict): MANO parameter annotations.
+        has_mano_params (Dict): Whether mano annotations are valid.
+        rot (float): Random rotation applied to the keypoints.
+        do_flip (bool): Whether to flip keypoints or not.
+    Returns:
+        Dict, Dict: Transformed MANO parameters and valid flags.
+    """
+    if do_flip:
+        mano_params, has_mano_params = fliplr_params(mano_params, has_mano_params)
+    mano_params['global_orient'] = rot_aa(mano_params['global_orient'], rot)
+    return mano_params, has_mano_params
+
+
+
+def get_example(img_path: str|np.ndarray, center_x: float, center_y: float,
+                width: float, height: float,
+                keypoints_2d: np.array, keypoints_3d: np.array,
+                mano_params: Dict, has_mano_params: Dict,
+                flip_kp_permutation: List[int],
+                patch_width: int, patch_height: int,
+                mean: np.array, std: np.array,
+                do_augment: bool, is_right: bool, augm_config: CfgNode,
+                is_bgr: bool = True,
+                use_skimage_antialias: bool = False,
+                border_mode: int = cv2.BORDER_CONSTANT,
+                return_trans: bool = False) -> Tuple:
+    """
+    Get an example from the dataset and (possibly) apply random augmentations.
+    Args:
+        img_path (str): Image filename
+        center_x (float): Bounding box center x coordinate in the original image.
+        center_y (float): Bounding box center y coordinate in the original image.
+        width (float): Bounding box width.
+        height (float): Bounding box height.
+        keypoints_2d (np.array): Array with shape (N,3) containing the 2D keypoints in the original image coordinates.
+        keypoints_3d (np.array): Array with shape (N,4) containing the 3D keypoints.
+        mano_params (Dict): MANO parameter annotations.
+        has_mano_params (Dict): Whether MANO annotations are valid.
+        flip_kp_permutation (List): Permutation to apply to the keypoints after flipping.
+        patch_width (float): Output box width.
+        patch_height (float): Output box height.
+        mean (np.array): Array of shape (3,) containing the mean for normalizing the input image.
+        std (np.array): Array of shape (3,) containing the std for normalizing the input image.
+        do_augment (bool): Whether to apply data augmentation or not.
+        aug_config (CfgNode): Config containing augmentation parameters.
+    Returns:
+        return img_patch, keypoints_2d, keypoints_3d, mano_params, has_mano_params, img_size
+        img_patch (np.array): Cropped image patch of shape (3, patch_height, patch_height)
+        keypoints_2d (np.array): Array with shape (N,3) containing the transformed 2D keypoints.
+        keypoints_3d (np.array): Array with shape (N,4) containing the transformed 3D keypoints.
+        mano_params (Dict): Transformed MANO parameters.
+        has_mano_params (Dict): Valid flag for transformed MANO parameters.
+        img_size (np.array): Image size of the original image.
+        """
+    if isinstance(img_path, str):
+        # 1. load image
+        cvimg = cv2.imread(img_path, cv2.IMREAD_COLOR | cv2.IMREAD_IGNORE_ORIENTATION)
+        if not isinstance(cvimg, np.ndarray):
+            raise IOError("Fail to read %s" % img_path)
+    elif isinstance(img_path, np.ndarray):
+        cvimg = img_path
+    else:
+        raise TypeError('img_path must be either a string or a numpy array')
+    img_height, img_width, img_channels = cvimg.shape
+
+    img_size = np.array([img_height, img_width])
+
+    # 2. get augmentation params
+    if do_augment:
+        scale, rot, do_flip, do_extreme_crop, extreme_crop_lvl, color_scale, tx, ty = do_augmentation(augm_config)
+    else:
+        scale, rot, do_flip, do_extreme_crop, extreme_crop_lvl, color_scale, tx, ty = 1.0, 0, False, False, 0, [1.0, 1.0, 1.0], 0., 0.
+
+    # if it's a left hand, we flip
+    if not is_right:
+        do_flip = True
+
+    if width < 1 or height < 1:
+        breakpoint()
+
+    if do_extreme_crop:
+        if extreme_crop_lvl == 0:
+            center_x1, center_y1, width1, height1 = extreme_cropping(center_x, center_y, width, height, keypoints_2d)
+        elif extreme_crop_lvl == 1:
+            center_x1, center_y1, width1, height1 = extreme_cropping_aggressive(center_x, center_y, width, height, keypoints_2d)
+
+        THRESH = 4
+        if width1 < THRESH or height1 < THRESH:
+            # print(f'{do_extreme_crop=}')
+            # print(f'width: {width}, height: {height}')
+            # print(f'width1: {width1}, height1: {height1}')
+            # print(f'center_x: {center_x}, center_y: {center_y}')
+            # print(f'center_x1: {center_x1}, center_y1: {center_y1}')
+            # print(f'keypoints_2d: {keypoints_2d}')
+            # print(f'\n\n', flush=True)
+            # breakpoint()
+            pass
+            # print(f'skip ==> width1: {width1}, height1: {height1}, width: {width}, height: {height}')
+        else:
+            center_x, center_y, width, height = center_x1, center_y1, width1, height1
+
+    center_x += width * tx
+    center_y += height * ty
+
+    # Process 3D keypoints
+    keypoints_3d = keypoint_3d_processing(keypoints_3d, flip_kp_permutation, rot, do_flip)
+
+    # 3. generate image patch
+    if use_skimage_antialias:
+        # Blur image to avoid aliasing artifacts
+        downsampling_factor = (patch_width / (width*scale))
+        if downsampling_factor > 1.1:
+            cvimg  = gaussian(cvimg, sigma=(downsampling_factor-1)/2, channel_axis=2, preserve_range=True, truncate=3.0)
+
+    img_patch_cv, trans = generate_image_patch_cv2(cvimg,
+                                                    center_x, center_y,
+                                                    width, height,
+                                                    patch_width, patch_height,
+                                                    do_flip, scale, rot, 
+                                                    border_mode=border_mode)
+        # img_patch_cv, trans = generate_image_patch_skimage(cvimg,
+        #                                                 center_x, center_y,
+        #                                                 width, height,
+        #                                                 patch_width, patch_height,
+        #                                                 do_flip, scale, rot, 
+        #                                                 border_mode=border_mode)
+
+    image = img_patch_cv.copy()
+    if is_bgr:
+        image = image[:, :, ::-1]
+    img_patch_cv = image.copy()
+    img_patch = convert_cvimg_to_tensor(image)
+
+
+    mano_params, has_mano_params = mano_param_processing(mano_params, has_mano_params, rot, do_flip)
+
+    # apply normalization
+    for n_c in range(min(img_channels, 3)):
+        img_patch[n_c, :, :] = np.clip(img_patch[n_c, :, :] * color_scale[n_c], 0, 255)
+        if mean is not None and std is not None:
+            img_patch[n_c, :, :] = (img_patch[n_c, :, :] - mean[n_c]) / std[n_c]
+    if do_flip:
+        keypoints_2d = fliplr_keypoints(keypoints_2d, img_width, flip_kp_permutation)
+
+
+    for n_jt in range(len(keypoints_2d)):
+        keypoints_2d[n_jt, 0:2] = trans_point2d(keypoints_2d[n_jt, 0:2], trans)
+    keypoints_2d[:, :-1] = keypoints_2d[:, :-1] / patch_width - 0.5
+
+    if not return_trans:
+        return img_patch, keypoints_2d, keypoints_3d, mano_params, has_mano_params, img_size
+    else:
+        return img_patch, keypoints_2d, keypoints_3d, mano_params, has_mano_params, img_size, trans
+
+def crop_to_hips(center_x: float, center_y: float, width: float, height: float, keypoints_2d: np.array) -> Tuple:
+    """
+    Extreme cropping: Crop the box up to the hip locations.
+    Args:
+        center_x (float): x coordinate of the bounding box center.
+        center_y (float): y coordinate of the bounding box center.
+        width (float): Bounding box width.
+        height (float): Bounding box height.
+        keypoints_2d (np.array): Array of shape (N, 3) containing 2D keypoint locations.
+    Returns:
+        center_x (float): x coordinate of the new bounding box center.
+        center_y (float): y coordinate of the new bounding box center.
+        width (float): New bounding box width.
+        height (float): New bounding box height.
+    """
+    keypoints_2d = keypoints_2d.copy()
+    lower_body_keypoints = [10, 11, 13, 14, 19, 20, 21, 22, 23, 24, 25+0, 25+1, 25+4, 25+5]
+    keypoints_2d[lower_body_keypoints, :] = 0
+    if keypoints_2d[:, -1].sum() > 1:
+        center, scale = get_bbox(keypoints_2d)
+        center_x = center[0]
+        center_y = center[1]
+        width = 1.1 * scale[0]
+        height = 1.1 * scale[1]
+    return center_x, center_y, width, height
+
+
+def crop_to_shoulders(center_x: float, center_y: float, width: float, height: float, keypoints_2d: np.array):
+    """
+    Extreme cropping: Crop the box up to the shoulder locations.
+    Args:
+        center_x (float): x coordinate of the bounding box center.
+        center_y (float): y coordinate of the bounding box center.
+        width (float): Bounding box width.
+        height (float): Bounding box height.
+        keypoints_2d (np.array): Array of shape (N, 3) containing 2D keypoint locations.
+    Returns:
+        center_x (float): x coordinate of the new bounding box center.
+        center_y (float): y coordinate of the new bounding box center.
+        width (float): New bounding box width.
+        height (float): New bounding box height.
+    """
+    keypoints_2d = keypoints_2d.copy()
+    lower_body_keypoints = [3, 4, 6, 7, 8, 9, 10, 11, 12, 13, 14, 19, 20, 21, 22, 23, 24] + [25 + i for i in [0, 1, 2, 3, 4, 5, 6, 7, 10, 11, 14, 15, 16]]
+    keypoints_2d[lower_body_keypoints, :] = 0
+    center, scale = get_bbox(keypoints_2d)
+    if keypoints_2d[:, -1].sum() > 1:
+        center, scale = get_bbox(keypoints_2d)
+        center_x = center[0]
+        center_y = center[1]
+        width = 1.2 * scale[0]
+        height = 1.2 * scale[1]
+    return center_x, center_y, width, height
+
+def crop_to_head(center_x: float, center_y: float, width: float, height: float, keypoints_2d: np.array):
+    """
+    Extreme cropping: Crop the box and keep on only the head.
+    Args:
+        center_x (float): x coordinate of the bounding box center.
+        center_y (float): y coordinate of the bounding box center.
+        width (float): Bounding box width.
+        height (float): Bounding box height.
+        keypoints_2d (np.array): Array of shape (N, 3) containing 2D keypoint locations.
+    Returns:
+        center_x (float): x coordinate of the new bounding box center.
+        center_y (float): y coordinate of the new bounding box center.
+        width (float): New bounding box width.
+        height (float): New bounding box height.
+    """
+    keypoints_2d = keypoints_2d.copy()
+    lower_body_keypoints = [3, 4, 6, 7, 8, 9, 10, 11, 12, 13, 14, 19, 20, 21, 22, 23, 24] + [25 + i for i in [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 14, 15, 16]]
+    keypoints_2d[lower_body_keypoints, :] = 0
+    if keypoints_2d[:, -1].sum() > 1:
+        center, scale = get_bbox(keypoints_2d)
+        center_x = center[0]
+        center_y = center[1]
+        width = 1.3 * scale[0]
+        height = 1.3 * scale[1]
+    return center_x, center_y, width, height
+
+def crop_torso_only(center_x: float, center_y: float, width: float, height: float, keypoints_2d: np.array):
+    """
+    Extreme cropping: Crop the box and keep on only the torso.
+    Args:
+        center_x (float): x coordinate of the bounding box center.
+        center_y (float): y coordinate of the bounding box center.
+        width (float): Bounding box width.
+        height (float): Bounding box height.
+        keypoints_2d (np.array): Array of shape (N, 3) containing 2D keypoint locations.
+    Returns:
+        center_x (float): x coordinate of the new bounding box center.
+        center_y (float): y coordinate of the new bounding box center.
+        width (float): New bounding box width.
+        height (float): New bounding box height.
+    """
+    keypoints_2d = keypoints_2d.copy()
+    nontorso_body_keypoints = [0, 3, 4, 6, 7, 10, 11, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24] + [25 + i for i in [0, 1, 4, 5, 6, 7, 10, 11, 13, 17, 18]]
+    keypoints_2d[nontorso_body_keypoints, :] = 0
+    if keypoints_2d[:, -1].sum() > 1:
+        center, scale = get_bbox(keypoints_2d)
+        center_x = center[0]
+        center_y = center[1]
+        width = 1.1 * scale[0]
+        height = 1.1 * scale[1]
+    return center_x, center_y, width, height
+
+def crop_rightarm_only(center_x: float, center_y: float, width: float, height: float, keypoints_2d: np.array):
+    """
+    Extreme cropping: Crop the box and keep on only the right arm.
+    Args:
+        center_x (float): x coordinate of the bounding box center.
+        center_y (float): y coordinate of the bounding box center.
+        width (float): Bounding box width.
+        height (float): Bounding box height.
+        keypoints_2d (np.array): Array of shape (N, 3) containing 2D keypoint locations.
+    Returns:
+        center_x (float): x coordinate of the new bounding box center.
+        center_y (float): y coordinate of the new bounding box center.
+        width (float): New bounding box width.
+        height (float): New bounding box height.
+    """
+    keypoints_2d = keypoints_2d.copy()
+    nonrightarm_body_keypoints = [0, 1, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24] + [25 + i for i in [0, 1, 2, 3, 4, 5, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18]]
+    keypoints_2d[nonrightarm_body_keypoints, :] = 0
+    if keypoints_2d[:, -1].sum() > 1:
+        center, scale = get_bbox(keypoints_2d)
+        center_x = center[0]
+        center_y = center[1]
+        width = 1.1 * scale[0]
+        height = 1.1 * scale[1]
+    return center_x, center_y, width, height
+
+def crop_leftarm_only(center_x: float, center_y: float, width: float, height: float, keypoints_2d: np.array):
+    """
+    Extreme cropping: Crop the box and keep on only the left arm.
+    Args:
+        center_x (float): x coordinate of the bounding box center.
+        center_y (float): y coordinate of the bounding box center.
+        width (float): Bounding box width.
+        height (float): Bounding box height.
+        keypoints_2d (np.array): Array of shape (N, 3) containing 2D keypoint locations.
+    Returns:
+        center_x (float): x coordinate of the new bounding box center.
+        center_y (float): y coordinate of the new bounding box center.
+        width (float): New bounding box width.
+        height (float): New bounding box height.
+    """
+    keypoints_2d = keypoints_2d.copy()
+    nonleftarm_body_keypoints = [0, 1, 2, 3, 4, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24] + [25 + i for i in [0, 1, 2, 3, 4, 5, 6, 7, 8, 12, 13, 14, 15, 16, 17, 18]]
+    keypoints_2d[nonleftarm_body_keypoints, :] = 0
+    if keypoints_2d[:, -1].sum() > 1:
+        center, scale = get_bbox(keypoints_2d)
+        center_x = center[0]
+        center_y = center[1]
+        width = 1.1 * scale[0]
+        height = 1.1 * scale[1]
+    return center_x, center_y, width, height
+
+def crop_legs_only(center_x: float, center_y: float, width: float, height: float, keypoints_2d: np.array):
+    """
+    Extreme cropping: Crop the box and keep on only the legs.
+    Args:
+        center_x (float): x coordinate of the bounding box center.
+        center_y (float): y coordinate of the bounding box center.
+        width (float): Bounding box width.
+        height (float): Bounding box height.
+        keypoints_2d (np.array): Array of shape (N, 3) containing 2D keypoint locations.
+    Returns:
+        center_x (float): x coordinate of the new bounding box center.
+        center_y (float): y coordinate of the new bounding box center.
+        width (float): New bounding box width.
+        height (float): New bounding box height.
+    """
+    keypoints_2d = keypoints_2d.copy()
+    nonlegs_body_keypoints = [0, 1, 2, 3, 4, 5, 6, 7, 15, 16, 17, 18] + [25 + i for i in [6, 7, 8, 9, 10, 11, 12, 13, 15, 16, 17, 18]]
+    keypoints_2d[nonlegs_body_keypoints, :] = 0
+    if keypoints_2d[:, -1].sum() > 1:
+        center, scale = get_bbox(keypoints_2d)
+        center_x = center[0]
+        center_y = center[1]
+        width = 1.1 * scale[0]
+        height = 1.1 * scale[1]
+    return center_x, center_y, width, height
+
+def crop_rightleg_only(center_x: float, center_y: float, width: float, height: float, keypoints_2d: np.array):
+    """
+    Extreme cropping: Crop the box and keep on only the right leg.
+    Args:
+        center_x (float): x coordinate of the bounding box center.
+        center_y (float): y coordinate of the bounding box center.
+        width (float): Bounding box width.
+        height (float): Bounding box height.
+        keypoints_2d (np.array): Array of shape (N, 3) containing 2D keypoint locations.
+    Returns:
+        center_x (float): x coordinate of the new bounding box center.
+        center_y (float): y coordinate of the new bounding box center.
+        width (float): New bounding box width.
+        height (float): New bounding box height.
+    """
+    keypoints_2d = keypoints_2d.copy()
+    nonrightleg_body_keypoints = [0, 1, 2, 3, 4, 5, 6, 7, 8, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21] + [25 + i for i in [3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18]]
+    keypoints_2d[nonrightleg_body_keypoints, :] = 0
+    if keypoints_2d[:, -1].sum() > 1:
+        center, scale = get_bbox(keypoints_2d)
+        center_x = center[0]
+        center_y = center[1]
+        width = 1.1 * scale[0]
+        height = 1.1 * scale[1]
+    return center_x, center_y, width, height
+
+def crop_leftleg_only(center_x: float, center_y: float, width: float, height: float, keypoints_2d: np.array):
+    """
+    Extreme cropping: Crop the box and keep on only the left leg.
+    Args:
+        center_x (float): x coordinate of the bounding box center.
+        center_y (float): y coordinate of the bounding box center.
+        width (float): Bounding box width.
+        height (float): Bounding box height.
+        keypoints_2d (np.array): Array of shape (N, 3) containing 2D keypoint locations.
+    Returns:
+        center_x (float): x coordinate of the new bounding box center.
+        center_y (float): y coordinate of the new bounding box center.
+        width (float): New bounding box width.
+        height (float): New bounding box height.
+    """
+    keypoints_2d = keypoints_2d.copy()
+    nonleftleg_body_keypoints = [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 15, 16, 17, 18, 22, 23, 24] + [25 + i for i in [0, 1, 2, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18]]
+    keypoints_2d[nonleftleg_body_keypoints, :] = 0
+    if keypoints_2d[:, -1].sum() > 1:
+        center, scale = get_bbox(keypoints_2d)
+        center_x = center[0]
+        center_y = center[1]
+        width = 1.1 * scale[0]
+        height = 1.1 * scale[1]
+    return center_x, center_y, width, height
+
+def full_body(keypoints_2d: np.array) -> bool:
+    """
+    Check if all main body joints are visible.
+    Args:
+        keypoints_2d (np.array): Array of shape (N, 3) containing 2D keypoint locations.
+    Returns:
+        bool: True if all main body joints are visible.
+    """
+
+    body_keypoints_openpose = [2, 3, 4, 5, 6, 7, 10, 11, 13, 14]
+    body_keypoints = [25 + i for i in [8, 7, 6, 9, 10, 11, 1, 0, 4, 5]]
+    return (np.maximum(keypoints_2d[body_keypoints, -1], keypoints_2d[body_keypoints_openpose, -1]) > 0).sum() == len(body_keypoints)
+
+def upper_body(keypoints_2d: np.array):
+    """
+    Check if all upper body joints are visible.
+    Args:
+        keypoints_2d (np.array): Array of shape (N, 3) containing 2D keypoint locations.
+    Returns:
+        bool: True if all main body joints are visible.
+    """
+    lower_body_keypoints_openpose = [10, 11, 13, 14]
+    lower_body_keypoints = [25 + i for i in [1, 0, 4, 5]]
+    upper_body_keypoints_openpose = [0, 1, 15, 16, 17, 18]
+    upper_body_keypoints = [25+8, 25+9, 25+12, 25+13, 25+17, 25+18]
+    return ((keypoints_2d[lower_body_keypoints + lower_body_keypoints_openpose, -1] > 0).sum() == 0)\
+       and ((keypoints_2d[upper_body_keypoints + upper_body_keypoints_openpose, -1] > 0).sum() >= 2)
+
+def get_bbox(keypoints_2d: np.array, rescale: float = 1.2) -> Tuple:
+    """
+    Get center and scale for bounding box from openpose detections.
+    Args:
+        keypoints_2d (np.array): Array of shape (N, 3) containing 2D keypoint locations.
+        rescale (float): Scale factor to rescale bounding boxes computed from the keypoints.
+    Returns:
+        center (np.array): Array of shape (2,) containing the new bounding box center.
+        scale (float): New bounding box scale.
+    """
+    valid = keypoints_2d[:,-1] > 0
+    valid_keypoints = keypoints_2d[valid][:,:-1]
+    center = 0.5 * (valid_keypoints.max(axis=0) + valid_keypoints.min(axis=0))
+    bbox_size = (valid_keypoints.max(axis=0) - valid_keypoints.min(axis=0))
+    # adjust bounding box tightness
+    scale = bbox_size
+    scale *= rescale
+    return center, scale
+
+def extreme_cropping(center_x: float, center_y: float, width: float, height: float, keypoints_2d: np.array) -> Tuple:
+    """
+    Perform extreme cropping
+    Args:
+        center_x (float): x coordinate of bounding box center.
+        center_y (float): y coordinate of bounding box center.
+        width (float): bounding box width.
+        height (float): bounding box height.
+        keypoints_2d (np.array): Array of shape (N, 3) containing 2D keypoint locations.
+        rescale (float): Scale factor to rescale bounding boxes computed from the keypoints.
+    Returns:
+        center_x (float): x coordinate of bounding box center.
+        center_y (float): y coordinate of bounding box center.
+        width (float): bounding box width.
+        height (float): bounding box height.
+    """
+    p = torch.rand(1).item()
+    if full_body(keypoints_2d):
+        if p < 0.7:
+            center_x, center_y, width, height = crop_to_hips(center_x, center_y, width, height, keypoints_2d)
+        elif p < 0.9:
+            center_x, center_y, width, height = crop_to_shoulders(center_x, center_y, width, height, keypoints_2d)
+        else:
+            center_x, center_y, width, height = crop_to_head(center_x, center_y, width, height, keypoints_2d)
+    elif upper_body(keypoints_2d):
+        if p < 0.9:
+            center_x, center_y, width, height = crop_to_shoulders(center_x, center_y, width, height, keypoints_2d)
+        else:
+            center_x, center_y, width, height = crop_to_head(center_x, center_y, width, height, keypoints_2d)
+
+    return center_x, center_y, max(width, height), max(width, height)
+
+def extreme_cropping_aggressive(center_x: float, center_y: float, width: float, height: float, keypoints_2d: np.array) -> Tuple:
+    """
+    Perform aggressive extreme cropping
+    Args:
+        center_x (float): x coordinate of bounding box center.
+        center_y (float): y coordinate of bounding box center.
+        width (float): bounding box width.
+        height (float): bounding box height.
+        keypoints_2d (np.array): Array of shape (N, 3) containing 2D keypoint locations.
+        rescale (float): Scale factor to rescale bounding boxes computed from the keypoints.
+    Returns:
+        center_x (float): x coordinate of bounding box center.
+        center_y (float): y coordinate of bounding box center.
+        width (float): bounding box width.
+        height (float): bounding box height.
+    """
+    p = torch.rand(1).item()
+    if full_body(keypoints_2d):
+        if p < 0.2:
+            center_x, center_y, width, height = crop_to_hips(center_x, center_y, width, height, keypoints_2d)
+        elif p < 0.3:
+            center_x, center_y, width, height = crop_to_shoulders(center_x, center_y, width, height, keypoints_2d)
+        elif p < 0.4:
+            center_x, center_y, width, height = crop_to_head(center_x, center_y, width, height, keypoints_2d)
+        elif p < 0.5:
+            center_x, center_y, width, height = crop_torso_only(center_x, center_y, width, height, keypoints_2d)
+        elif p < 0.6:
+            center_x, center_y, width, height = crop_rightarm_only(center_x, center_y, width, height, keypoints_2d)
+        elif p < 0.7:
+            center_x, center_y, width, height = crop_leftarm_only(center_x, center_y, width, height, keypoints_2d)
+        elif p < 0.8:
+            center_x, center_y, width, height = crop_legs_only(center_x, center_y, width, height, keypoints_2d)
+        elif p < 0.9:
+            center_x, center_y, width, height = crop_rightleg_only(center_x, center_y, width, height, keypoints_2d)
+        else:
+            center_x, center_y, width, height = crop_leftleg_only(center_x, center_y, width, height, keypoints_2d)
+    elif upper_body(keypoints_2d):
+        if p < 0.2:
+            center_x, center_y, width, height = crop_to_shoulders(center_x, center_y, width, height, keypoints_2d)
+        elif p < 0.4:
+            center_x, center_y, width, height = crop_to_head(center_x, center_y, width, height, keypoints_2d)
+        elif p < 0.6:
+            center_x, center_y, width, height = crop_torso_only(center_x, center_y, width, height, keypoints_2d)
+        elif p < 0.8:
+            center_x, center_y, width, height = crop_rightarm_only(center_x, center_y, width, height, keypoints_2d)
+        else:
+            center_x, center_y, width, height = crop_leftarm_only(center_x, center_y, width, height, keypoints_2d)
+    return center_x, center_y, max(width, height), max(width, height)

hamer/datasets/vitdet_dataset.py

@@ -0,0 +1,97 @@
+from typing import Dict
+
+import cv2
+import numpy as np
+from skimage.filters import gaussian
+from yacs.config import CfgNode
+import torch
+
+from .utils import (convert_cvimg_to_tensor,
+                    expand_to_aspect_ratio,
+                    generate_image_patch_cv2)
+
+DEFAULT_MEAN = 255. * np.array([0.485, 0.456, 0.406])
+DEFAULT_STD = 255. * np.array([0.229, 0.224, 0.225])
+
+class ViTDetDataset(torch.utils.data.Dataset):
+
+    def __init__(self,
+                 cfg: CfgNode,
+                 img_cv2: np.array,
+                 boxes: np.array,
+                 right: np.array,
+                 rescale_factor=2.5,
+                 train: bool = False,
+                 **kwargs):
+        super().__init__()
+        self.cfg = cfg
+        self.img_cv2 = img_cv2
+        # self.boxes = boxes
+
+        assert train == False, "ViTDetDataset is only for inference"
+        self.train = train
+        self.img_size = cfg.MODEL.IMAGE_SIZE
+        self.mean = 255. * np.array(self.cfg.MODEL.IMAGE_MEAN)
+        self.std = 255. * np.array(self.cfg.MODEL.IMAGE_STD)
+
+        # Preprocess annotations
+        boxes = boxes.astype(np.float32)
+        self.center = (boxes[:, 2:4] + boxes[:, 0:2]) / 2.0
+        self.scale = rescale_factor * (boxes[:, 2:4] - boxes[:, 0:2]) / 200.0
+        #self.scale = (boxes[:, 2:4] - boxes[:, 0:2]) / 200.0
+        self.personid = np.arange(len(boxes), dtype=np.int32)
+        self.right = right.astype(np.float32)
+
+    def __len__(self) -> int:
+        return len(self.personid)
+
+    def __getitem__(self, idx: int) -> Dict[str, np.array]:
+
+        center = self.center[idx].copy()
+        center_x = center[0]
+        center_y = center[1]
+
+        scale = self.scale[idx]
+        BBOX_SHAPE = self.cfg.MODEL.get('BBOX_SHAPE', None)
+        bbox_size = expand_to_aspect_ratio(scale*200, target_aspect_ratio=BBOX_SHAPE).max()
+        #bbox_size = scale.max()*200
+
+        patch_width = patch_height = self.img_size
+
+        right = self.right[idx].copy()
+        flip = right == 0
+
+        # 3. generate image patch
+        # if use_skimage_antialias:
+        cvimg = self.img_cv2.copy()
+        if True:
+            # Blur image to avoid aliasing artifacts
+            downsampling_factor = ((bbox_size*1.0) / patch_width)
+            print(f'{downsampling_factor=}')
+            downsampling_factor = downsampling_factor / 2.0
+            if downsampling_factor > 1.1:
+                cvimg  = gaussian(cvimg, sigma=(downsampling_factor-1)/2, channel_axis=2, preserve_range=True)
+
+
+        img_patch_cv, trans = generate_image_patch_cv2(cvimg,
+                                                    center_x, center_y,
+                                                    bbox_size, bbox_size,
+                                                    patch_width, patch_height,
+                                                    flip, 1.0, 0,
+                                                    border_mode=cv2.BORDER_CONSTANT)
+        img_patch_cv = img_patch_cv[:, :, ::-1]
+        img_patch = convert_cvimg_to_tensor(img_patch_cv)
+
+        # apply normalization
+        for n_c in range(min(self.img_cv2.shape[2], 3)):
+            img_patch[n_c, :, :] = (img_patch[n_c, :, :] - self.mean[n_c]) / self.std[n_c]
+
+        item = {
+            'img': img_patch,
+            'personid': int(self.personid[idx]),
+        }
+        item['box_center'] = self.center[idx].copy()
+        item['box_size'] = bbox_size
+        item['img_size'] = 1.0 * np.array([cvimg.shape[1], cvimg.shape[0]])
+        item['right'] = self.right[idx].copy()
+        return item

hamer/models/__init__.py

@@ -0,0 +1,46 @@
+from .mano_wrapper import MANO
+from .hamer import HAMER
+from .discriminator import Discriminator
+
+from ..utils.download import cache_url
+from ..configs import CACHE_DIR_HAMER
+
+
+def download_models(folder=CACHE_DIR_HAMER):
+    """Download checkpoints and files for running inference.
+    """
+    import os
+    os.makedirs(folder, exist_ok=True)
+    download_files = {
+        "hamer_data.tar.gz"      : ["https://people.eecs.berkeley.edu/~jathushan/projects/4dhumans/hamer_data.tar.gz", folder],
+    }
+    
+    for file_name, url in download_files.items():
+        output_path = os.path.join(url[1], file_name)
+        if not os.path.exists(output_path):
+            print("Downloading file: " + file_name)
+            # output = gdown.cached_download(url[0], output_path, fuzzy=True)
+            output = cache_url(url[0], output_path)
+            assert os.path.exists(output_path), f"{output} does not exist"
+
+            # if ends with tar.gz, tar -xzf
+            if file_name.endswith(".tar.gz"):
+                print("Extracting file: " + file_name)
+                os.system("tar -xvf " + output_path + " -C " + url[1])
+
+DEFAULT_CHECKPOINT=f'{CACHE_DIR_HAMER}/hamer_ckpts/checkpoints/hamer.ckpt'
+def load_hamer(checkpoint_path=DEFAULT_CHECKPOINT):
+    from pathlib import Path
+    from ..configs import get_config
+    model_cfg = str(Path(checkpoint_path).parent.parent / 'model_config.yaml')
+    model_cfg = get_config(model_cfg, update_cachedir=True)
+
+    # Override some config values, to crop bbox correctly
+    if (model_cfg.MODEL.BACKBONE.TYPE == 'vit') and ('BBOX_SHAPE' not in model_cfg.MODEL):
+        model_cfg.defrost()
+        assert model_cfg.MODEL.IMAGE_SIZE == 256, f"MODEL.IMAGE_SIZE ({model_cfg.MODEL.IMAGE_SIZE}) should be 256 for ViT backbone"
+        model_cfg.MODEL.BBOX_SHAPE = [192,256]
+        model_cfg.freeze()
+
+    model = HAMER.load_from_checkpoint(checkpoint_path, strict=False, cfg=model_cfg)
+    return model, model_cfg

hamer/models/backbones/__init__.py

@@ -0,0 +1,7 @@
+from .vit import vit
+
+def create_backbone(cfg):
+    if cfg.MODEL.BACKBONE.TYPE == 'vit':
+        return vit(cfg)
+    else:
+        raise NotImplementedError('Backbone type is not implemented')

hamer/models/backbones/vit.py

@@ -0,0 +1,348 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+import math
+
+import torch
+from functools import partial
+import torch.nn as nn
+import torch.nn.functional as F
+import torch.utils.checkpoint as checkpoint
+
+from timm.models.layers import drop_path, to_2tuple, trunc_normal_
+
+def vit(cfg):
+    return ViT(
+                img_size=(256, 192),
+                patch_size=16,
+                embed_dim=1280,
+                depth=32,
+                num_heads=16,
+                ratio=1,
+                use_checkpoint=False,
+                mlp_ratio=4,
+                qkv_bias=True,
+                drop_path_rate=0.55,
+            )
+
+def get_abs_pos(abs_pos, h, w, ori_h, ori_w, has_cls_token=True):
+    """
+    Calculate absolute positional embeddings. If needed, resize embeddings and remove cls_token
+        dimension for the original embeddings.
+    Args:
+        abs_pos (Tensor): absolute positional embeddings with (1, num_position, C).
+        has_cls_token (bool): If true, has 1 embedding in abs_pos for cls token.
+        hw (Tuple): size of input image tokens.
+
+    Returns:
+        Absolute positional embeddings after processing with shape (1, H, W, C)
+    """
+    cls_token = None
+    B, L, C = abs_pos.shape
+    if has_cls_token:
+        cls_token = abs_pos[:, 0:1]
+        abs_pos = abs_pos[:, 1:]
+
+    if ori_h != h or ori_w != w:
+        new_abs_pos = F.interpolate(
+            abs_pos.reshape(1, ori_h, ori_w, -1).permute(0, 3, 1, 2),
+            size=(h, w),
+            mode="bicubic",
+            align_corners=False,
+        ).permute(0, 2, 3, 1).reshape(B, -1, C)
+
+    else:
+        new_abs_pos = abs_pos
+    
+    if cls_token is not None:
+        new_abs_pos = torch.cat([cls_token, new_abs_pos], dim=1)
+    return new_abs_pos
+
+class DropPath(nn.Module):
+    """Drop paths (Stochastic Depth) per sample  (when applied in main path of residual blocks).
+    """
+    def __init__(self, drop_prob=None):
+        super(DropPath, self).__init__()
+        self.drop_prob = drop_prob
+
+    def forward(self, x):
+        return drop_path(x, self.drop_prob, self.training)
+    
+    def extra_repr(self):
+        return 'p={}'.format(self.drop_prob)
+
+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 = nn.Linear(in_features, hidden_features)
+        self.act = act_layer()
+        self.fc2 = nn.Linear(hidden_features, out_features)
+        self.drop = nn.Dropout(drop)
+
+    def forward(self, x):
+        x = self.fc1(x)
+        x = self.act(x)
+        x = self.fc2(x)
+        x = self.drop(x)
+        return x
+
+class Attention(nn.Module):
+    def __init__(
+            self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0.,
+            proj_drop=0., attn_head_dim=None,):
+        super().__init__()
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+        self.dim = dim
+
+        if attn_head_dim is not None:
+            head_dim = attn_head_dim
+        all_head_dim = head_dim * self.num_heads
+
+        self.scale = qk_scale or head_dim ** -0.5
+
+        self.qkv = nn.Linear(dim, all_head_dim * 3, bias=qkv_bias)
+
+        self.attn_drop = nn.Dropout(attn_drop)
+        self.proj = nn.Linear(all_head_dim, dim)
+        self.proj_drop = nn.Dropout(proj_drop)
+
+    def forward(self, x):
+        B, N, C = x.shape
+        qkv = self.qkv(x)
+        qkv = qkv.reshape(B, N, 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
+        q, k, v = qkv[0], qkv[1], qkv[2]   # make torchscript happy (cannot use tensor as tuple)
+
+        q = q * self.scale
+        attn = (q @ k.transpose(-2, -1))
+
+        attn = attn.softmax(dim=-1)
+        attn = self.attn_drop(attn)
+
+        x = (attn @ v).transpose(1, 2).reshape(B, N, -1)
+        x = self.proj(x)
+        x = self.proj_drop(x)
+
+        return x
+
+class Block(nn.Module):
+
+    def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, 
+                 drop=0., attn_drop=0., drop_path=0., act_layer=nn.GELU, 
+                 norm_layer=nn.LayerNorm, attn_head_dim=None
+                 ):
+        super().__init__()
+        
+        self.norm1 = norm_layer(dim)
+        self.attn = Attention(
+            dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale,
+            attn_drop=attn_drop, proj_drop=drop, attn_head_dim=attn_head_dim
+            )
+
+        # NOTE: drop path for stochastic depth, we shall see if this is better than dropout here
+        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
+        self.norm2 = norm_layer(dim)
+        mlp_hidden_dim = int(dim * mlp_ratio)
+        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
+
+    def forward(self, x):
+        x = x + self.drop_path(self.attn(self.norm1(x)))
+        x = x + self.drop_path(self.mlp(self.norm2(x)))
+        return x
+
+
+class PatchEmbed(nn.Module):
+    """ Image to Patch Embedding
+    """
+    def __init__(self, img_size=224, patch_size=16, in_chans=3, embed_dim=768, ratio=1):
+        super().__init__()
+        img_size = to_2tuple(img_size)
+        patch_size = to_2tuple(patch_size)
+        num_patches = (img_size[1] // patch_size[1]) * (img_size[0] // patch_size[0]) * (ratio ** 2)
+        self.patch_shape = (int(img_size[0] // patch_size[0] * ratio), int(img_size[1] // patch_size[1] * ratio))
+        self.origin_patch_shape = (int(img_size[0] // patch_size[0]), int(img_size[1] // patch_size[1]))
+        self.img_size = img_size
+        self.patch_size = patch_size
+        self.num_patches = num_patches
+
+        self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=(patch_size[0] // ratio), padding=4 + 2 * (ratio//2-1))
+
+    def forward(self, x, **kwargs):
+        B, C, H, W = x.shape
+        x = self.proj(x)
+        Hp, Wp = x.shape[2], x.shape[3]
+
+        x = x.flatten(2).transpose(1, 2)
+        return x, (Hp, Wp)
+
+
+class HybridEmbed(nn.Module):
+    """ CNN Feature Map Embedding
+    Extract feature map from CNN, flatten, project to embedding dim.
+    """
+    def __init__(self, backbone, img_size=224, feature_size=None, in_chans=3, embed_dim=768):
+        super().__init__()
+        assert isinstance(backbone, nn.Module)
+        img_size = to_2tuple(img_size)
+        self.img_size = img_size
+        self.backbone = backbone
+        if feature_size is None:
+            with torch.no_grad():
+                training = backbone.training
+                if training:
+                    backbone.eval()
+                o = self.backbone(torch.zeros(1, in_chans, img_size[0], img_size[1]))[-1]
+                feature_size = o.shape[-2:]
+                feature_dim = o.shape[1]
+                backbone.train(training)
+        else:
+            feature_size = to_2tuple(feature_size)
+            feature_dim = self.backbone.feature_info.channels()[-1]
+        self.num_patches = feature_size[0] * feature_size[1]
+        self.proj = nn.Linear(feature_dim, embed_dim)
+
+    def forward(self, x):
+        x = self.backbone(x)[-1]
+        x = x.flatten(2).transpose(1, 2)
+        x = self.proj(x)
+        return x
+
+
+class ViT(nn.Module):
+
+    def __init__(self,
+                 img_size=224, patch_size=16, in_chans=3, num_classes=80, embed_dim=768, depth=12,
+                 num_heads=12, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop_rate=0., attn_drop_rate=0.,
+                 drop_path_rate=0., hybrid_backbone=None, norm_layer=None, use_checkpoint=False, 
+                 frozen_stages=-1, ratio=1, last_norm=True,
+                 patch_padding='pad', freeze_attn=False, freeze_ffn=False,
+                 ):
+        # Protect mutable default arguments
+        super(ViT, self).__init__()
+        norm_layer = norm_layer or partial(nn.LayerNorm, eps=1e-6)
+        self.num_classes = num_classes
+        self.num_features = self.embed_dim = embed_dim  # num_features for consistency with other models
+        self.frozen_stages = frozen_stages
+        self.use_checkpoint = use_checkpoint
+        self.patch_padding = patch_padding
+        self.freeze_attn = freeze_attn
+        self.freeze_ffn = freeze_ffn
+        self.depth = depth
+
+        if hybrid_backbone is not None:
+            self.patch_embed = HybridEmbed(
+                hybrid_backbone, img_size=img_size, in_chans=in_chans, embed_dim=embed_dim)
+        else:
+            self.patch_embed = PatchEmbed(
+                img_size=img_size, patch_size=patch_size, in_chans=in_chans, embed_dim=embed_dim, ratio=ratio)
+        num_patches = self.patch_embed.num_patches
+
+        # since the pretraining model has class token
+        self.pos_embed = nn.Parameter(torch.zeros(1, num_patches + 1, embed_dim))
+
+        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, depth)]  # stochastic depth decay rule
+
+        self.blocks = nn.ModuleList([
+            Block(
+                dim=embed_dim, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale,
+                drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[i], norm_layer=norm_layer,
+                )
+            for i in range(depth)])
+
+        self.last_norm = norm_layer(embed_dim) if last_norm else nn.Identity()
+
+        if self.pos_embed is not None:
+            trunc_normal_(self.pos_embed, std=.02)
+
+        self._freeze_stages()
+
+    def _freeze_stages(self):
+        """Freeze parameters."""
+        if self.frozen_stages >= 0:
+            self.patch_embed.eval()
+            for param in self.patch_embed.parameters():
+                param.requires_grad = False
+
+        for i in range(1, self.frozen_stages + 1):
+            m = self.blocks[i]
+            m.eval()
+            for param in m.parameters():
+                param.requires_grad = False
+
+        if self.freeze_attn:
+            for i in range(0, self.depth):
+                m = self.blocks[i]
+                m.attn.eval()
+                m.norm1.eval()
+                for param in m.attn.parameters():
+                    param.requires_grad = False
+                for param in m.norm1.parameters():
+                    param.requires_grad = False
+
+        if self.freeze_ffn:
+            self.pos_embed.requires_grad = False
+            self.patch_embed.eval()
+            for param in self.patch_embed.parameters():
+                param.requires_grad = False
+            for i in range(0, self.depth):
+                m = self.blocks[i]
+                m.mlp.eval()
+                m.norm2.eval()
+                for param in m.mlp.parameters():
+                    param.requires_grad = False
+                for param in m.norm2.parameters():
+                    param.requires_grad = False
+
+    def init_weights(self):
+        """Initialize the weights in backbone.
+        Args:
+            pretrained (str, optional): Path to pre-trained weights.
+                Defaults to None.
+        """
+        def _init_weights(m):
+            if isinstance(m, nn.Linear):
+                trunc_normal_(m.weight, std=.02)
+                if isinstance(m, nn.Linear) and m.bias is not None:
+                    nn.init.constant_(m.bias, 0)
+            elif isinstance(m, nn.LayerNorm):
+                nn.init.constant_(m.bias, 0)
+                nn.init.constant_(m.weight, 1.0)
+
+        self.apply(_init_weights)
+
+    def get_num_layers(self):
+        return len(self.blocks)
+
+    @torch.jit.ignore
+    def no_weight_decay(self):
+        return {'pos_embed', 'cls_token'}
+
+    def forward_features(self, x):
+        B, C, H, W = x.shape
+        x, (Hp, Wp) = self.patch_embed(x)
+
+        if self.pos_embed is not None:
+            # fit for multiple GPU training
+            # since the first element for pos embed (sin-cos manner) is zero, it will cause no difference
+            x = x + self.pos_embed[:, 1:] + self.pos_embed[:, :1]
+
+        for blk in self.blocks:
+            if self.use_checkpoint:
+                x = checkpoint.checkpoint(blk, x)
+            else:
+                x = blk(x)
+
+        x = self.last_norm(x)
+
+        xp = x.permute(0, 2, 1).reshape(B, -1, Hp, Wp).contiguous()
+
+        return xp
+
+    def forward(self, x):
+        x = self.forward_features(x)
+        return x
+
+    def train(self, mode=True):
+        """Convert the model into training mode."""
+        super().train(mode)
+        self._freeze_stages()

hamer/models/components/pose_transformer.py

@@ -0,0 +1,358 @@
+from inspect import isfunction
+from typing import Callable, Optional
+
+import torch
+from einops import rearrange
+from einops.layers.torch import Rearrange
+from torch import nn
+
+from .t_cond_mlp import (
+    AdaptiveLayerNorm1D,
+    FrequencyEmbedder,
+    normalization_layer,
+)
+# from .vit import Attention, FeedForward
+
+
+def exists(val):
+    return val is not None
+
+
+def default(val, d):
+    if exists(val):
+        return val
+    return d() if isfunction(d) else d
+
+
+class PreNorm(nn.Module):
+    def __init__(self, dim: int, fn: Callable, norm: str = "layer", norm_cond_dim: int = -1):
+        super().__init__()
+        self.norm = normalization_layer(norm, dim, norm_cond_dim)
+        self.fn = fn
+
+    def forward(self, x: torch.Tensor, *args, **kwargs):
+        if isinstance(self.norm, AdaptiveLayerNorm1D):
+            return self.fn(self.norm(x, *args), **kwargs)
+        else:
+            return self.fn(self.norm(x), **kwargs)
+
+
+class FeedForward(nn.Module):
+    def __init__(self, dim, hidden_dim, dropout=0.0):
+        super().__init__()
+        self.net = nn.Sequential(
+            nn.Linear(dim, hidden_dim),
+            nn.GELU(),
+            nn.Dropout(dropout),
+            nn.Linear(hidden_dim, dim),
+            nn.Dropout(dropout),
+        )
+
+    def forward(self, x):
+        return self.net(x)
+
+
+class Attention(nn.Module):
+    def __init__(self, dim, heads=8, dim_head=64, dropout=0.0):
+        super().__init__()
+        inner_dim = dim_head * heads
+        project_out = not (heads == 1 and dim_head == dim)
+
+        self.heads = heads
+        self.scale = dim_head**-0.5
+
+        self.attend = nn.Softmax(dim=-1)
+        self.dropout = nn.Dropout(dropout)
+
+        self.to_qkv = nn.Linear(dim, inner_dim * 3, bias=False)
+
+        self.to_out = (
+            nn.Sequential(nn.Linear(inner_dim, dim), nn.Dropout(dropout))
+            if project_out
+            else nn.Identity()
+        )
+
+    def forward(self, x):
+        qkv = self.to_qkv(x).chunk(3, dim=-1)
+        q, k, v = map(lambda t: rearrange(t, "b n (h d) -> b h n d", h=self.heads), qkv)
+
+        dots = torch.matmul(q, k.transpose(-1, -2)) * self.scale
+
+        attn = self.attend(dots)
+        attn = self.dropout(attn)
+
+        out = torch.matmul(attn, v)
+        out = rearrange(out, "b h n d -> b n (h d)")
+        return self.to_out(out)
+
+
+class CrossAttention(nn.Module):
+    def __init__(self, dim, context_dim=None, heads=8, dim_head=64, dropout=0.0):
+        super().__init__()
+        inner_dim = dim_head * heads
+        project_out = not (heads == 1 and dim_head == dim)
+
+        self.heads = heads
+        self.scale = dim_head**-0.5
+
+        self.attend = nn.Softmax(dim=-1)
+        self.dropout = nn.Dropout(dropout)
+
+        context_dim = default(context_dim, dim)
+        self.to_kv = nn.Linear(context_dim, inner_dim * 2, bias=False)
+        self.to_q = nn.Linear(dim, inner_dim, bias=False)
+
+        self.to_out = (
+            nn.Sequential(nn.Linear(inner_dim, dim), nn.Dropout(dropout))
+            if project_out
+            else nn.Identity()
+        )
+
+    def forward(self, x, context=None):
+        context = default(context, x)
+        k, v = self.to_kv(context).chunk(2, dim=-1)
+        q = self.to_q(x)
+        q, k, v = map(lambda t: rearrange(t, "b n (h d) -> b h n d", h=self.heads), [q, k, v])
+
+        dots = torch.matmul(q, k.transpose(-1, -2)) * self.scale
+
+        attn = self.attend(dots)
+        attn = self.dropout(attn)
+
+        out = torch.matmul(attn, v)
+        out = rearrange(out, "b h n d -> b n (h d)")
+        return self.to_out(out)
+
+
+class Transformer(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        depth: int,
+        heads: int,
+        dim_head: int,
+        mlp_dim: int,
+        dropout: float = 0.0,
+        norm: str = "layer",
+        norm_cond_dim: int = -1,
+    ):
+        super().__init__()
+        self.layers = nn.ModuleList([])
+        for _ in range(depth):
+            sa = Attention(dim, heads=heads, dim_head=dim_head, dropout=dropout)
+            ff = FeedForward(dim, mlp_dim, dropout=dropout)
+            self.layers.append(
+                nn.ModuleList(
+                    [
+                        PreNorm(dim, sa, norm=norm, norm_cond_dim=norm_cond_dim),
+                        PreNorm(dim, ff, norm=norm, norm_cond_dim=norm_cond_dim),
+                    ]
+                )
+            )
+
+    def forward(self, x: torch.Tensor, *args):
+        for attn, ff in self.layers:
+            x = attn(x, *args) + x
+            x = ff(x, *args) + x
+        return x
+
+
+class TransformerCrossAttn(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        depth: int,
+        heads: int,
+        dim_head: int,
+        mlp_dim: int,
+        dropout: float = 0.0,
+        norm: str = "layer",
+        norm_cond_dim: int = -1,
+        context_dim: Optional[int] = None,
+    ):
+        super().__init__()
+        self.layers = nn.ModuleList([])
+        for _ in range(depth):
+            sa = Attention(dim, heads=heads, dim_head=dim_head, dropout=dropout)
+            ca = CrossAttention(
+                dim, context_dim=context_dim, heads=heads, dim_head=dim_head, dropout=dropout
+            )
+            ff = FeedForward(dim, mlp_dim, dropout=dropout)
+            self.layers.append(
+                nn.ModuleList(
+                    [
+                        PreNorm(dim, sa, norm=norm, norm_cond_dim=norm_cond_dim),
+                        PreNorm(dim, ca, norm=norm, norm_cond_dim=norm_cond_dim),
+                        PreNorm(dim, ff, norm=norm, norm_cond_dim=norm_cond_dim),
+                    ]
+                )
+            )
+
+    def forward(self, x: torch.Tensor, *args, context=None, context_list=None):
+        if context_list is None:
+            context_list = [context] * len(self.layers)
+        if len(context_list) != len(self.layers):
+            raise ValueError(f"len(context_list) != len(self.layers) ({len(context_list)} != {len(self.layers)})")
+
+        for i, (self_attn, cross_attn, ff) in enumerate(self.layers):
+            x = self_attn(x, *args) + x
+            x = cross_attn(x, *args, context=context_list[i]) + x
+            x = ff(x, *args) + x
+        return x
+
+
+class DropTokenDropout(nn.Module):
+    def __init__(self, p: float = 0.1):
+        super().__init__()
+        if p < 0 or p > 1:
+            raise ValueError(
+                "dropout probability has to be between 0 and 1, " "but got {}".format(p)
+            )
+        self.p = p
+
+    def forward(self, x: torch.Tensor):
+        # x: (batch_size, seq_len, dim)
+        if self.training and self.p > 0:
+            zero_mask = torch.full_like(x[0, :, 0], self.p).bernoulli().bool()
+            # TODO: permutation idx for each batch using torch.argsort
+            if zero_mask.any():
+                x = x[:, ~zero_mask, :]
+        return x
+
+
+class ZeroTokenDropout(nn.Module):
+    def __init__(self, p: float = 0.1):
+        super().__init__()
+        if p < 0 or p > 1:
+            raise ValueError(
+                "dropout probability has to be between 0 and 1, " "but got {}".format(p)
+            )
+        self.p = p
+
+    def forward(self, x: torch.Tensor):
+        # x: (batch_size, seq_len, dim)
+        if self.training and self.p > 0:
+            zero_mask = torch.full_like(x[:, :, 0], self.p).bernoulli().bool()
+            # Zero-out the masked tokens
+            x[zero_mask, :] = 0
+        return x
+
+
+class TransformerEncoder(nn.Module):
+    def __init__(
+        self,
+        num_tokens: int,
+        token_dim: int,
+        dim: int,
+        depth: int,
+        heads: int,
+        mlp_dim: int,
+        dim_head: int = 64,
+        dropout: float = 0.0,
+        emb_dropout: float = 0.0,
+        emb_dropout_type: str = "drop",
+        emb_dropout_loc: str = "token",
+        norm: str = "layer",
+        norm_cond_dim: int = -1,
+        token_pe_numfreq: int = -1,
+    ):
+        super().__init__()
+        if token_pe_numfreq > 0:
+            token_dim_new = token_dim * (2 * token_pe_numfreq + 1)
+            self.to_token_embedding = nn.Sequential(
+                Rearrange("b n d -> (b n) d", n=num_tokens, d=token_dim),
+                FrequencyEmbedder(token_pe_numfreq, token_pe_numfreq - 1),
+                Rearrange("(b n) d -> b n d", n=num_tokens, d=token_dim_new),
+                nn.Linear(token_dim_new, dim),
+            )
+        else:
+            self.to_token_embedding = nn.Linear(token_dim, dim)
+        self.pos_embedding = nn.Parameter(torch.randn(1, num_tokens, dim))
+        if emb_dropout_type == "drop":
+            self.dropout = DropTokenDropout(emb_dropout)
+        elif emb_dropout_type == "zero":
+            self.dropout = ZeroTokenDropout(emb_dropout)
+        else:
+            raise ValueError(f"Unknown emb_dropout_type: {emb_dropout_type}")
+        self.emb_dropout_loc = emb_dropout_loc
+
+        self.transformer = Transformer(
+            dim, depth, heads, dim_head, mlp_dim, dropout, norm=norm, norm_cond_dim=norm_cond_dim
+        )
+
+    def forward(self, inp: torch.Tensor, *args, **kwargs):
+        x = inp
+
+        if self.emb_dropout_loc == "input":
+            x = self.dropout(x)
+        x = self.to_token_embedding(x)
+
+        if self.emb_dropout_loc == "token":
+            x = self.dropout(x)
+        b, n, _ = x.shape
+        x += self.pos_embedding[:, :n]
+
+        if self.emb_dropout_loc == "token_afterpos":
+            x = self.dropout(x)
+        x = self.transformer(x, *args)
+        return x
+
+
+class TransformerDecoder(nn.Module):
+    def __init__(
+        self,
+        num_tokens: int,
+        token_dim: int,
+        dim: int,
+        depth: int,
+        heads: int,
+        mlp_dim: int,
+        dim_head: int = 64,
+        dropout: float = 0.0,
+        emb_dropout: float = 0.0,
+        emb_dropout_type: str = 'drop',
+        norm: str = "layer",
+        norm_cond_dim: int = -1,
+        context_dim: Optional[int] = None,
+        skip_token_embedding: bool = False,
+    ):
+        super().__init__()
+        if not skip_token_embedding:
+            self.to_token_embedding = nn.Linear(token_dim, dim)
+        else:
+            self.to_token_embedding = nn.Identity()
+            if token_dim != dim:
+                raise ValueError(
+                    f"token_dim ({token_dim}) != dim ({dim}) when skip_token_embedding is True"
+                )
+
+        self.pos_embedding = nn.Parameter(torch.randn(1, num_tokens, dim))
+        if emb_dropout_type == "drop":
+            self.dropout = DropTokenDropout(emb_dropout)
+        elif emb_dropout_type == "zero":
+            self.dropout = ZeroTokenDropout(emb_dropout)
+        elif emb_dropout_type == "normal":
+            self.dropout = nn.Dropout(emb_dropout)
+
+        self.transformer = TransformerCrossAttn(
+            dim,
+            depth,
+            heads,
+            dim_head,
+            mlp_dim,
+            dropout,
+            norm=norm,
+            norm_cond_dim=norm_cond_dim,
+            context_dim=context_dim,
+        )
+
+    def forward(self, inp: torch.Tensor, *args, context=None, context_list=None):
+        x = self.to_token_embedding(inp)
+        b, n, _ = x.shape
+
+        x = self.dropout(x)
+        x += self.pos_embedding[:, :n]
+
+        x = self.transformer(x, *args, context=context, context_list=context_list)
+        return x
+

hamer/models/components/t_cond_mlp.py

@@ -0,0 +1,199 @@
+import copy
+from typing import List, Optional
+
+import torch
+
+
+class AdaptiveLayerNorm1D(torch.nn.Module):
+    def __init__(self, data_dim: int, norm_cond_dim: int):
+        super().__init__()
+        if data_dim <= 0:
+            raise ValueError(f"data_dim must be positive, but got {data_dim}")
+        if norm_cond_dim <= 0:
+            raise ValueError(f"norm_cond_dim must be positive, but got {norm_cond_dim}")
+        self.norm = torch.nn.LayerNorm(
+            data_dim
+        )  # TODO: Check if elementwise_affine=True is correct
+        self.linear = torch.nn.Linear(norm_cond_dim, 2 * data_dim)
+        torch.nn.init.zeros_(self.linear.weight)
+        torch.nn.init.zeros_(self.linear.bias)
+
+    def forward(self, x: torch.Tensor, t: torch.Tensor) -> torch.Tensor:
+        # x: (batch, ..., data_dim)
+        # t: (batch, norm_cond_dim)
+        # return: (batch, data_dim)
+        x = self.norm(x)
+        alpha, beta = self.linear(t).chunk(2, dim=-1)
+
+        # Add singleton dimensions to alpha and beta
+        if x.dim() > 2:
+            alpha = alpha.view(alpha.shape[0], *([1] * (x.dim() - 2)), alpha.shape[1])
+            beta = beta.view(beta.shape[0], *([1] * (x.dim() - 2)), beta.shape[1])
+
+        return x * (1 + alpha) + beta
+
+
+class SequentialCond(torch.nn.Sequential):
+    def forward(self, input, *args, **kwargs):
+        for module in self:
+            if isinstance(module, (AdaptiveLayerNorm1D, SequentialCond, ResidualMLPBlock)):
+                # print(f'Passing on args to {module}', [a.shape for a in args])
+                input = module(input, *args, **kwargs)
+            else:
+                # print(f'Skipping passing args to {module}', [a.shape for a in args])
+                input = module(input)
+        return input
+
+
+def normalization_layer(norm: Optional[str], dim: int, norm_cond_dim: int = -1):
+    if norm == "batch":
+        return torch.nn.BatchNorm1d(dim)
+    elif norm == "layer":
+        return torch.nn.LayerNorm(dim)
+    elif norm == "ada":
+        assert norm_cond_dim > 0, f"norm_cond_dim must be positive, got {norm_cond_dim}"
+        return AdaptiveLayerNorm1D(dim, norm_cond_dim)
+    elif norm is None:
+        return torch.nn.Identity()
+    else:
+        raise ValueError(f"Unknown norm: {norm}")
+
+
+def linear_norm_activ_dropout(
+    input_dim: int,
+    output_dim: int,
+    activation: torch.nn.Module = torch.nn.ReLU(),
+    bias: bool = True,
+    norm: Optional[str] = "layer",  # Options: ada/batch/layer
+    dropout: float = 0.0,
+    norm_cond_dim: int = -1,
+) -> SequentialCond:
+    layers = []
+    layers.append(torch.nn.Linear(input_dim, output_dim, bias=bias))
+    if norm is not None:
+        layers.append(normalization_layer(norm, output_dim, norm_cond_dim))
+    layers.append(copy.deepcopy(activation))
+    if dropout > 0.0:
+        layers.append(torch.nn.Dropout(dropout))
+    return SequentialCond(*layers)
+
+
+def create_simple_mlp(
+    input_dim: int,
+    hidden_dims: List[int],
+    output_dim: int,
+    activation: torch.nn.Module = torch.nn.ReLU(),
+    bias: bool = True,
+    norm: Optional[str] = "layer",  # Options: ada/batch/layer
+    dropout: float = 0.0,
+    norm_cond_dim: int = -1,
+) -> SequentialCond:
+    layers = []
+    prev_dim = input_dim
+    for hidden_dim in hidden_dims:
+        layers.extend(
+            linear_norm_activ_dropout(
+                prev_dim, hidden_dim, activation, bias, norm, dropout, norm_cond_dim
+            )
+        )
+        prev_dim = hidden_dim
+    layers.append(torch.nn.Linear(prev_dim, output_dim, bias=bias))
+    return SequentialCond(*layers)
+
+
+class ResidualMLPBlock(torch.nn.Module):
+    def __init__(
+        self,
+        input_dim: int,
+        hidden_dim: int,
+        num_hidden_layers: int,
+        output_dim: int,
+        activation: torch.nn.Module = torch.nn.ReLU(),
+        bias: bool = True,
+        norm: Optional[str] = "layer",  # Options: ada/batch/layer
+        dropout: float = 0.0,
+        norm_cond_dim: int = -1,
+    ):
+        super().__init__()
+        if not (input_dim == output_dim == hidden_dim):
+            raise NotImplementedError(
+                f"input_dim {input_dim} != output_dim {output_dim} is not implemented"
+            )
+
+        layers = []
+        prev_dim = input_dim
+        for i in range(num_hidden_layers):
+            layers.append(
+                linear_norm_activ_dropout(
+                    prev_dim, hidden_dim, activation, bias, norm, dropout, norm_cond_dim
+                )
+            )
+            prev_dim = hidden_dim
+        self.model = SequentialCond(*layers)
+        self.skip = torch.nn.Identity()
+
+    def forward(self, x: torch.Tensor, *args, **kwargs) -> torch.Tensor:
+        return x + self.model(x, *args, **kwargs)
+
+
+class ResidualMLP(torch.nn.Module):
+    def __init__(
+        self,
+        input_dim: int,
+        hidden_dim: int,
+        num_hidden_layers: int,
+        output_dim: int,
+        activation: torch.nn.Module = torch.nn.ReLU(),
+        bias: bool = True,
+        norm: Optional[str] = "layer",  # Options: ada/batch/layer
+        dropout: float = 0.0,
+        num_blocks: int = 1,
+        norm_cond_dim: int = -1,
+    ):
+        super().__init__()
+        self.input_dim = input_dim
+        self.model = SequentialCond(
+            linear_norm_activ_dropout(
+                input_dim, hidden_dim, activation, bias, norm, dropout, norm_cond_dim
+            ),
+            *[
+                ResidualMLPBlock(
+                    hidden_dim,
+                    hidden_dim,
+                    num_hidden_layers,
+                    hidden_dim,
+                    activation,
+                    bias,
+                    norm,
+                    dropout,
+                    norm_cond_dim,
+                )
+                for _ in range(num_blocks)
+            ],
+            torch.nn.Linear(hidden_dim, output_dim, bias=bias),
+        )
+
+    def forward(self, x: torch.Tensor, *args, **kwargs) -> torch.Tensor:
+        return self.model(x, *args, **kwargs)
+
+
+class FrequencyEmbedder(torch.nn.Module):
+    def __init__(self, num_frequencies, max_freq_log2):
+        super().__init__()
+        frequencies = 2 ** torch.linspace(0, max_freq_log2, steps=num_frequencies)
+        self.register_buffer("frequencies", frequencies)
+
+    def forward(self, x):
+        # x should be of size (N,) or (N, D)
+        N = x.size(0)
+        if x.dim() == 1:  # (N,)
+            x = x.unsqueeze(1)  # (N, D) where D=1
+        x_unsqueezed = x.unsqueeze(-1)  # (N, D, 1)
+        scaled = self.frequencies.view(1, 1, -1) * x_unsqueezed  # (N, D, num_frequencies)
+        s = torch.sin(scaled)
+        c = torch.cos(scaled)
+        embedded = torch.cat([s, c, x_unsqueezed], dim=-1).view(
+            N, -1
+        )  # (N, D * 2 * num_frequencies + D)
+        return embedded
+

hamer/models/discriminator.py

@@ -0,0 +1,99 @@
+import torch
+import torch.nn as nn
+
+class Discriminator(nn.Module):
+
+    def __init__(self):
+        """
+        Pose + Shape discriminator proposed in HMR
+        """
+        super(Discriminator, self).__init__()
+
+        self.num_joints = 15
+        # poses_alone
+        self.D_conv1 = nn.Conv2d(9, 32, kernel_size=1)
+        nn.init.xavier_uniform_(self.D_conv1.weight)
+        nn.init.zeros_(self.D_conv1.bias)
+        self.relu = nn.ReLU(inplace=True)
+        self.D_conv2 = nn.Conv2d(32, 32, kernel_size=1)
+        nn.init.xavier_uniform_(self.D_conv2.weight)
+        nn.init.zeros_(self.D_conv2.bias)
+        pose_out = []
+        for i in range(self.num_joints):
+            pose_out_temp = nn.Linear(32, 1)
+            nn.init.xavier_uniform_(pose_out_temp.weight)
+            nn.init.zeros_(pose_out_temp.bias)
+            pose_out.append(pose_out_temp)
+        self.pose_out = nn.ModuleList(pose_out)
+
+        # betas
+        self.betas_fc1 = nn.Linear(10, 10)
+        nn.init.xavier_uniform_(self.betas_fc1.weight)
+        nn.init.zeros_(self.betas_fc1.bias)
+        self.betas_fc2 = nn.Linear(10, 5)
+        nn.init.xavier_uniform_(self.betas_fc2.weight)
+        nn.init.zeros_(self.betas_fc2.bias)
+        self.betas_out = nn.Linear(5, 1)
+        nn.init.xavier_uniform_(self.betas_out.weight)
+        nn.init.zeros_(self.betas_out.bias)
+
+        # poses_joint
+        self.D_alljoints_fc1 = nn.Linear(32*self.num_joints, 1024)
+        nn.init.xavier_uniform_(self.D_alljoints_fc1.weight)
+        nn.init.zeros_(self.D_alljoints_fc1.bias)
+        self.D_alljoints_fc2 = nn.Linear(1024, 1024)
+        nn.init.xavier_uniform_(self.D_alljoints_fc2.weight)
+        nn.init.zeros_(self.D_alljoints_fc2.bias)
+        self.D_alljoints_out = nn.Linear(1024, 1)
+        nn.init.xavier_uniform_(self.D_alljoints_out.weight)
+        nn.init.zeros_(self.D_alljoints_out.bias)
+
+
+    def forward(self, poses: torch.Tensor, betas: torch.Tensor) -> torch.Tensor:
+        """
+        Forward pass of the discriminator.
+        Args:
+            poses (torch.Tensor): Tensor of shape (B, 23, 3, 3) containing a batch of MANO hand poses (excluding the global orientation).
+            betas (torch.Tensor): Tensor of shape (B, 10) containign a batch of MANO beta coefficients.
+        Returns:
+            torch.Tensor: Discriminator output with shape (B, 25)
+        """
+        #import ipdb; ipdb.set_trace()
+        #bn = poses.shape[0]
+        # poses B x 207
+        #poses = poses.reshape(bn, -1)
+        # poses B x num_joints x 1 x 9
+        poses = poses.reshape(-1, self.num_joints, 1, 9)
+        bn = poses.shape[0]
+        # poses B x 9 x num_joints x 1
+        poses = poses.permute(0, 3, 1, 2).contiguous()
+
+        # poses_alone
+        poses = self.D_conv1(poses)
+        poses = self.relu(poses)
+        poses = self.D_conv2(poses)
+        poses = self.relu(poses)
+
+        poses_out = []
+        for i in range(self.num_joints):
+            poses_out_ = self.pose_out[i](poses[:, :, i, 0])
+            poses_out.append(poses_out_)
+        poses_out = torch.cat(poses_out, dim=1)
+
+        # betas
+        betas = self.betas_fc1(betas)
+        betas = self.relu(betas)
+        betas = self.betas_fc2(betas)
+        betas = self.relu(betas)
+        betas_out = self.betas_out(betas)
+
+        # poses_joint
+        poses = poses.reshape(bn,-1)
+        poses_all = self.D_alljoints_fc1(poses)
+        poses_all = self.relu(poses_all)
+        poses_all = self.D_alljoints_fc2(poses_all)
+        poses_all = self.relu(poses_all)
+        poses_all_out = self.D_alljoints_out(poses_all)
+
+        disc_out = torch.cat((poses_out, betas_out, poses_all_out), 1)
+        return disc_out

hamer/models/hamer.py

@@ -0,0 +1,363 @@
+import torch
+import pytorch_lightning as pl
+from typing import Any, Dict, Mapping, Tuple
+
+from yacs.config import CfgNode
+
+from ..utils import SkeletonRenderer, MeshRenderer
+from ..utils.geometry import aa_to_rotmat, perspective_projection
+from ..utils.pylogger import get_pylogger
+from .backbones import create_backbone
+from .heads import build_mano_head
+from .discriminator import Discriminator
+from .losses import Keypoint3DLoss, Keypoint2DLoss, ParameterLoss
+from . import MANO
+
+log = get_pylogger(__name__)
+
+class HAMER(pl.LightningModule):
+
+    def __init__(self, cfg: CfgNode, init_renderer: bool = False):
+        """
+        Setup HAMER model
+        Args:
+            cfg (CfgNode): Config file as a yacs CfgNode
+        """
+        super().__init__()
+
+        # Save hyperparameters
+        self.save_hyperparameters(logger=False, ignore=['init_renderer'])
+
+        self.cfg = cfg
+        # Create backbone feature extractor
+        self.backbone = create_backbone(cfg)
+        #if cfg.MODEL.BACKBONE.get('PRETRAINED_WEIGHTS', None):
+        #    log.info(f'Loading backbone weights from {cfg.MODEL.BACKBONE.PRETRAINED_WEIGHTS}')
+        #    self.backbone.load_state_dict(torch.load(cfg.MODEL.BACKBONE.PRETRAINED_WEIGHTS, map_location='cpu')['state_dict'])
+
+        # Create MANO head
+        self.mano_head = build_mano_head(cfg)
+
+        # Create discriminator
+        if self.cfg.LOSS_WEIGHTS.ADVERSARIAL > 0:
+            self.discriminator = Discriminator()
+
+        # Define loss functions
+        self.keypoint_3d_loss = Keypoint3DLoss(loss_type='l1')
+        self.keypoint_2d_loss = Keypoint2DLoss(loss_type='l1')
+        self.mano_parameter_loss = ParameterLoss()
+
+        # Instantiate MANO model
+        mano_cfg = {k.lower(): v for k,v in dict(cfg.MANO).items()}
+        self.mano = MANO(**mano_cfg)
+
+        # Buffer that shows whetheer we need to initialize ActNorm layers
+        self.register_buffer('initialized', torch.tensor(False))
+        # Setup renderer for visualization
+        if init_renderer:
+            self.renderer = SkeletonRenderer(self.cfg)
+            self.mesh_renderer = MeshRenderer(self.cfg, faces=self.mano.faces)
+        else:
+            self.renderer = None
+            self.mesh_renderer = None
+
+        # Disable automatic optimization since we use adversarial training
+        self.automatic_optimization = False
+
+    def on_after_backward(self):
+        for name, param in self.named_parameters():
+            if param.grad is None:
+                print(param.shape)
+                print(name)
+
+    def get_parameters(self):
+        all_params = list(self.mano_head.parameters())
+        all_params += list(self.backbone.parameters())
+        return all_params
+
+    def configure_optimizers(self) -> Tuple[torch.optim.Optimizer, torch.optim.Optimizer]:
+        """
+        Setup model and distriminator Optimizers
+        Returns:
+            Tuple[torch.optim.Optimizer, torch.optim.Optimizer]: Model and discriminator optimizers
+        """
+        param_groups = [{'params': filter(lambda p: p.requires_grad, self.get_parameters()), 'lr': self.cfg.TRAIN.LR}]
+
+        optimizer = torch.optim.AdamW(params=param_groups,
+                                        # lr=self.cfg.TRAIN.LR,
+                                        weight_decay=self.cfg.TRAIN.WEIGHT_DECAY)
+        optimizer_disc = torch.optim.AdamW(params=self.discriminator.parameters(),
+                                            lr=self.cfg.TRAIN.LR,
+                                            weight_decay=self.cfg.TRAIN.WEIGHT_DECAY)
+
+        return optimizer, optimizer_disc
+
+    def forward_step(self, batch: Dict, train: bool = False) -> Dict:
+        """
+        Run a forward step of the network
+        Args:
+            batch (Dict): Dictionary containing batch data
+            train (bool): Flag indicating whether it is training or validation mode
+        Returns:
+            Dict: Dictionary containing the regression output
+        """
+
+        # Use RGB image as input
+        x = batch['img']
+        batch_size = x.shape[0]
+
+        # Compute conditioning features using the backbone
+        # if using ViT backbone, we need to use a different aspect ratio
+        conditioning_feats = self.backbone(x[:,:,:,32:-32])
+
+        pred_mano_params, pred_cam, _ = self.mano_head(conditioning_feats)
+
+        # Store useful regression outputs to the output dict
+        output = {}
+        output['pred_cam'] = pred_cam
+        output['pred_mano_params'] = {k: v.clone() for k,v in pred_mano_params.items()}
+
+        # Compute camera translation
+        device = pred_mano_params['hand_pose'].device
+        dtype = pred_mano_params['hand_pose'].dtype
+        focal_length = self.cfg.EXTRA.FOCAL_LENGTH * torch.ones(batch_size, 2, device=device, dtype=dtype)
+        pred_cam_t = torch.stack([pred_cam[:, 1],
+                                  pred_cam[:, 2],
+                                  2*focal_length[:, 0]/(self.cfg.MODEL.IMAGE_SIZE * pred_cam[:, 0] +1e-9)],dim=-1)
+        output['pred_cam_t'] = pred_cam_t
+        output['focal_length'] = focal_length
+
+        # Compute model vertices, joints and the projected joints
+        pred_mano_params['global_orient'] = pred_mano_params['global_orient'].reshape(batch_size, -1, 3, 3)
+        pred_mano_params['hand_pose'] = pred_mano_params['hand_pose'].reshape(batch_size, -1, 3, 3)
+        pred_mano_params['betas'] = pred_mano_params['betas'].reshape(batch_size, -1)
+        mano_output = self.mano(**{k: v.float() for k,v in pred_mano_params.items()}, pose2rot=False)
+        pred_keypoints_3d = mano_output.joints
+        pred_vertices = mano_output.vertices
+        output['pred_keypoints_3d'] = pred_keypoints_3d.reshape(batch_size, -1, 3)
+        output['pred_vertices'] = pred_vertices.reshape(batch_size, -1, 3)
+        pred_cam_t = pred_cam_t.reshape(-1, 3)
+        focal_length = focal_length.reshape(-1, 2)
+        pred_keypoints_2d = perspective_projection(pred_keypoints_3d,
+                                                   translation=pred_cam_t,
+                                                   focal_length=focal_length / self.cfg.MODEL.IMAGE_SIZE)
+
+        output['pred_keypoints_2d'] = pred_keypoints_2d.reshape(batch_size, -1, 2)
+        return output
+
+    def compute_loss(self, batch: Dict, output: Dict, train: bool = True) -> torch.Tensor:
+        """
+        Compute losses given the input batch and the regression output
+        Args:
+            batch (Dict): Dictionary containing batch data
+            output (Dict): Dictionary containing the regression output
+            train (bool): Flag indicating whether it is training or validation mode
+        Returns:
+            torch.Tensor : Total loss for current batch
+        """
+
+        pred_mano_params = output['pred_mano_params']
+        pred_keypoints_2d = output['pred_keypoints_2d']
+        pred_keypoints_3d = output['pred_keypoints_3d']
+
+
+        batch_size = pred_mano_params['hand_pose'].shape[0]
+        device = pred_mano_params['hand_pose'].device
+        dtype = pred_mano_params['hand_pose'].dtype
+
+        # Get annotations
+        gt_keypoints_2d = batch['keypoints_2d']
+        gt_keypoints_3d = batch['keypoints_3d']
+        gt_mano_params = batch['mano_params']
+        has_mano_params = batch['has_mano_params']
+        is_axis_angle = batch['mano_params_is_axis_angle']
+
+        # Compute 3D keypoint loss
+        loss_keypoints_2d = self.keypoint_2d_loss(pred_keypoints_2d, gt_keypoints_2d)
+        loss_keypoints_3d = self.keypoint_3d_loss(pred_keypoints_3d, gt_keypoints_3d, pelvis_id=0)
+
+        # Compute loss on MANO parameters
+        loss_mano_params = {}
+        for k, pred in pred_mano_params.items():
+            gt = gt_mano_params[k].view(batch_size, -1)
+            if is_axis_angle[k].all():
+                gt = aa_to_rotmat(gt.reshape(-1, 3)).view(batch_size, -1, 3, 3)
+            has_gt = has_mano_params[k]
+            loss_mano_params[k] = self.mano_parameter_loss(pred.reshape(batch_size, -1), gt.reshape(batch_size, -1), has_gt)
+
+        loss = self.cfg.LOSS_WEIGHTS['KEYPOINTS_3D'] * loss_keypoints_3d+\
+               self.cfg.LOSS_WEIGHTS['KEYPOINTS_2D'] * loss_keypoints_2d+\
+               sum([loss_mano_params[k] * self.cfg.LOSS_WEIGHTS[k.upper()] for k in loss_mano_params])
+
+        #loss = loss + 0*self.mano.body_pose.mean()
+
+        losses = dict(loss=loss.detach(),
+                      loss_keypoints_2d=loss_keypoints_2d.detach(),
+                      loss_keypoints_3d=loss_keypoints_3d.detach())
+
+        for k, v in loss_mano_params.items():
+            losses['loss_' + k] = v.detach()
+
+        output['losses'] = losses
+
+        return loss
+
+    # Tensoroboard logging should run from first rank only
+    @pl.utilities.rank_zero.rank_zero_only
+    def tensorboard_logging(self, batch: Dict, output: Dict, step_count: int, train: bool = True, write_to_summary_writer: bool = True) -> None:
+        """
+        Log results to Tensorboard
+        Args:
+            batch (Dict): Dictionary containing batch data
+            output (Dict): Dictionary containing the regression output
+            step_count (int): Global training step count
+            train (bool): Flag indicating whether it is training or validation mode
+        """
+
+        mode = 'train' if train else 'val'
+        batch_size = batch['keypoints_2d'].shape[0]
+        images = batch['img']
+        images = images * torch.tensor([0.229, 0.224, 0.225], device=images.device).reshape(1,3,1,1)
+        images = images + torch.tensor([0.485, 0.456, 0.406], device=images.device).reshape(1,3,1,1)
+        #images = 255*images.permute(0, 2, 3, 1).cpu().numpy()
+
+        pred_keypoints_3d = output['pred_keypoints_3d'].detach().reshape(batch_size, -1, 3)
+        pred_vertices = output['pred_vertices'].detach().reshape(batch_size, -1, 3)
+        focal_length = output['focal_length'].detach().reshape(batch_size, 2)
+        gt_keypoints_3d = batch['keypoints_3d']
+        gt_keypoints_2d = batch['keypoints_2d']
+        losses = output['losses']
+        pred_cam_t = output['pred_cam_t'].detach().reshape(batch_size, 3)
+        pred_keypoints_2d = output['pred_keypoints_2d'].detach().reshape(batch_size, -1, 2)
+
+        if write_to_summary_writer:
+            summary_writer = self.logger.experiment
+            for loss_name, val in losses.items():
+                summary_writer.add_scalar(mode +'/' + loss_name, val.detach().item(), step_count)
+        num_images = min(batch_size, self.cfg.EXTRA.NUM_LOG_IMAGES)
+
+        gt_keypoints_3d = batch['keypoints_3d']
+        pred_keypoints_3d = output['pred_keypoints_3d'].detach().reshape(batch_size, -1, 3)
+
+        # We render the skeletons instead of the full mesh because rendering a lot of meshes will make the training slow.
+        #predictions = self.renderer(pred_keypoints_3d[:num_images],
+        #                            gt_keypoints_3d[:num_images],
+        #                            2 * gt_keypoints_2d[:num_images],
+        #                            images=images[:num_images],
+        #                            camera_translation=pred_cam_t[:num_images])
+        predictions = self.mesh_renderer.visualize_tensorboard(pred_vertices[:num_images].cpu().numpy(),
+                                                               pred_cam_t[:num_images].cpu().numpy(),
+                                                               images[:num_images].cpu().numpy(),
+                                                               pred_keypoints_2d[:num_images].cpu().numpy(),
+                                                               gt_keypoints_2d[:num_images].cpu().numpy(),
+                                                               focal_length=focal_length[:num_images].cpu().numpy())
+        if write_to_summary_writer:
+            summary_writer.add_image('%s/predictions' % mode, predictions, step_count)
+
+        return predictions
+
+    def forward(self, batch: Dict) -> Dict:
+        """
+        Run a forward step of the network in val mode
+        Args:
+            batch (Dict): Dictionary containing batch data
+        Returns:
+            Dict: Dictionary containing the regression output
+        """
+        return self.forward_step(batch, train=False)
+
+    def training_step_discriminator(self, batch: Dict,
+                                    hand_pose: torch.Tensor,
+                                    betas: torch.Tensor,
+                                    optimizer: torch.optim.Optimizer) -> torch.Tensor:
+        """
+        Run a discriminator training step
+        Args:
+            batch (Dict): Dictionary containing mocap batch data
+            hand_pose (torch.Tensor): Regressed hand pose from current step
+            betas (torch.Tensor): Regressed betas from current step
+            optimizer (torch.optim.Optimizer): Discriminator optimizer
+        Returns:
+            torch.Tensor: Discriminator loss
+        """
+        batch_size = hand_pose.shape[0]
+        gt_hand_pose = batch['hand_pose']
+        gt_betas = batch['betas']
+        gt_rotmat = aa_to_rotmat(gt_hand_pose.view(-1,3)).view(batch_size, -1, 3, 3)
+        disc_fake_out = self.discriminator(hand_pose.detach(), betas.detach())
+        loss_fake = ((disc_fake_out - 0.0) ** 2).sum() / batch_size
+        disc_real_out = self.discriminator(gt_rotmat, gt_betas)
+        loss_real = ((disc_real_out - 1.0) ** 2).sum() / batch_size
+        loss_disc = loss_fake + loss_real
+        loss = self.cfg.LOSS_WEIGHTS.ADVERSARIAL * loss_disc
+        optimizer.zero_grad()
+        self.manual_backward(loss)
+        optimizer.step()
+        return loss_disc.detach()
+
+    def training_step(self, joint_batch: Dict, batch_idx: int) -> Dict:
+        """
+        Run a full training step
+        Args:
+            joint_batch (Dict): Dictionary containing image and mocap batch data
+            batch_idx (int): Unused.
+            batch_idx (torch.Tensor): Unused.
+        Returns:
+            Dict: Dictionary containing regression output.
+        """
+        batch = joint_batch['img']
+        mocap_batch = joint_batch['mocap']
+        optimizer = self.optimizers(use_pl_optimizer=True)
+        if self.cfg.LOSS_WEIGHTS.ADVERSARIAL > 0:
+            optimizer, optimizer_disc = optimizer
+
+        batch_size = batch['img'].shape[0]
+        output = self.forward_step(batch, train=True)
+        pred_mano_params = output['pred_mano_params']
+        if self.cfg.get('UPDATE_GT_SPIN', False):
+            self.update_batch_gt_spin(batch, output)
+        loss = self.compute_loss(batch, output, train=True)
+        if self.cfg.LOSS_WEIGHTS.ADVERSARIAL > 0:
+            disc_out = self.discriminator(pred_mano_params['hand_pose'].reshape(batch_size, -1), pred_mano_params['betas'].reshape(batch_size, -1))
+            loss_adv = ((disc_out - 1.0) ** 2).sum() / batch_size
+            loss = loss + self.cfg.LOSS_WEIGHTS.ADVERSARIAL * loss_adv
+
+        # Error if Nan
+        if torch.isnan(loss):
+            raise ValueError('Loss is NaN')
+
+        optimizer.zero_grad()
+        self.manual_backward(loss)
+        # Clip gradient
+        if self.cfg.TRAIN.get('GRAD_CLIP_VAL', 0) > 0:
+            gn = torch.nn.utils.clip_grad_norm_(self.get_parameters(), self.cfg.TRAIN.GRAD_CLIP_VAL, error_if_nonfinite=True)
+            self.log('train/grad_norm', gn, on_step=True, on_epoch=True, prog_bar=True, logger=True)
+        optimizer.step()
+        if self.cfg.LOSS_WEIGHTS.ADVERSARIAL > 0:
+            loss_disc = self.training_step_discriminator(mocap_batch, pred_mano_params['hand_pose'].reshape(batch_size, -1), pred_mano_params['betas'].reshape(batch_size, -1), optimizer_disc)
+            output['losses']['loss_gen'] = loss_adv
+            output['losses']['loss_disc'] = loss_disc
+
+        if self.global_step > 0 and self.global_step % self.cfg.GENERAL.LOG_STEPS == 0:
+            self.tensorboard_logging(batch, output, self.global_step, train=True)
+
+        self.log('train/loss', output['losses']['loss'], on_step=True, on_epoch=True, prog_bar=True, logger=False)
+
+        return output
+
+    def validation_step(self, batch: Dict, batch_idx: int, dataloader_idx=0) -> Dict:
+        """
+        Run a validation step and log to Tensorboard
+        Args:
+            batch (Dict): Dictionary containing batch data
+            batch_idx (int): Unused.
+        Returns:
+            Dict: Dictionary containing regression output.
+        """
+        # batch_size = batch['img'].shape[0]
+        output = self.forward_step(batch, train=False)
+        loss = self.compute_loss(batch, output, train=False)
+        output['loss'] = loss
+        self.tensorboard_logging(batch, output, self.global_step, train=False)
+
+        return output

hamer/models/heads/__init__.py

@@ -0,0 +1 @@
+from .mano_head import build_mano_head