Image Audio Alingment Train

.gitattributes

@@ -142,3 +142,4 @@ Vaani/audio_urls[[:space:]]copy.txt filter=lfs diff=lfs merge=lfs -text
 Vaani/imageBY.csv filter=lfs diff=lfs merge=lfs -text
 Vaani/imageBY2.csv filter=lfs diff=lfs merge=lfs -text
 Vaani/imageBY3.csv filter=lfs diff=lfs merge=lfs -text
+Vaani/Image-Audio-Hindi.csv filter=lfs diff=lfs merge=lfs -text

Vaani/23m1521.code-workspace

@@ -0,0 +1,13 @@
+{
+	"folders": [
+		{
+			"path": "../../.."
+		},
+		{
+			"path": ".."
+		}
+	],
+	"settings": {
+		"terminal.integrated.mouseWheelZoom": true
+	}
+}

Vaani/Image-Audio-Hindi.csv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a44517e6935e21d887dbe1fe076e9b14607a22c5a709bb33e504b353c6fbb99d
+size 1561128197

Vaani/Img_Audio_Alignment/_2.1_Train.py

@@ -0,0 +1,1788 @@
+# ==================================================================
+#             L A T E N T   D I F F U S I O N   M O D E L
+# ==================================================================
+# Author    : Ashish Kumar Uchadiya
+# Created   : May 11, 2025
+# Description: This script implements the training of a VQ-VAE model for
+# image reconstruction, integrated with Latent Diffusion Models (LDMs) and
+# audio conditioning. The VQ-VAE maps images to a discrete latent space, 
+# which is then modeled by the LDM for learning a diffusion process over the 
+# compressed representation. Audio features are used as conditioning inputs 
+# to guide the generation process. The training minimizes a combination of 
+# LPIPS (Learned Perceptual Image Patch Similarity) loss for perceptual 
+# fidelity and PatchGAN loss to enforce local realism. This setup enables 
+# efficient and semantically-aware generation of high-quality images driven 
+# by audio cues.
+# ==================================================================
+#                         I M P O R T S
+# ==================================================================
+from __future__ import annotations
+import warnings
+warnings.filterwarnings("ignore")
+
+
+import os
+import io
+import sys
+import math
+import random
+import collections
+import collections.abc
+import re
+from itertools import repeat
+from pathlib import Path
+from typing import Optional, Tuple, Union, List, Dict
+
+import csv
+import numpy as np
+import pandas as pd
+from PIL import Image
+import seaborn as sns
+import matplotlib.pyplot as plt
+from tqdm import trange, tqdm
+
+import torch
+import torch.nn.functional as F
+from torch import nn
+from torch.nn.init import _calculate_fan_in_and_fan_out
+import torch.utils.checkpoint as checkpoint
+
+import torchvision as tv
+from torchvision.transforms import v2
+from torch.utils.tensorboard import SummaryWriter
+
+os.environ["CUDA_VISIBLE_DEVICES"] = "1"
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+print(f"Using device: {device}")
+
+import torchaudio
+import torchaudio.transforms as T
+from torchlibrosa.stft import Spectrogram, LogmelFilterBank
+from torchlibrosa.augmentation import SpecAugmentation
+
+from transformers import AutoModel, AutoTokenizer, logging
+from huggingface_hub.file_download import hf_hub_download
+from huggingface_hub.file_download import hf_hub_download
+from peft import get_peft_config, get_peft_model
+from transformers import CLIPVisionModel, AutoProcessor
+
+
+# ==================================================================
+#                        H T S - A T
+# ==================================================================
+class HTSATConfig:
+    # Ke Chen
+    # [email protected]
+    # HTS-AT: A HIERARCHICAL TOKEN-SEMANTIC AUDIO TRANSFORMER FOR SOUND CLASSIFICATION AND DETECTION
+    # The configuration for training the model
+
+    exp_name = "exp_htsat_pretrain" # the saved ckpt prefix name of the model 
+    workspace = "/home/kechen/Research/HTSAT" # the folder of your code
+    dataset_path = "/home/Research/audioset" # the dataset path
+    desed_folder = "/home/Research/DESED" # the desed file
+
+    dataset_type = "audioset" # "audioset" "esc-50" "scv2"
+    index_type = "full_train" # only works for audioset
+    balanced_data = True # only works for audioset
+
+    loss_type = "clip_bce" # 
+    # AudioSet & SCV2: "clip_bce" |  ESC-50: "clip_ce" 
+
+    # trained from a checkpoint, or evaluate a single model 
+    resume_checkpoint = None 
+    # "/home/Research/model_backup/AudioSet/HTSAT_AudioSet_Saved_1.ckpt"
+    
+    esc_fold = 0 # just for esc dataset, select the fold you need for evaluation and (+1) validation
+
+
+    debug = False
+
+    random_seed = 970131 # 19970318 970131 12412 127777 1009 34047
+    batch_size = 32 * 4 # batch size per GPU x GPU number , default is 32 x 4 = 128
+    learning_rate = 1e-3 # 1e-4 also workable 
+    max_epoch = 100
+    num_workers = 3
+
+    lr_scheduler_epoch = [10,20,30]
+    lr_rate = [0.02, 0.05, 0.1]
+
+    # these data preparation optimizations do not bring many improvements, so deprecated
+    enable_token_label = False # token label
+    class_map_path = "class_hier_map.npy"
+    class_filter = None 
+    retrieval_index = [15382, 9202, 130, 17618, 17157, 17516, 16356, 6165, 13992, 9238, 5550, 5733, 1914, 1600, 3450, 13735, 11108, 3762, 
+        9840, 11318, 8131, 4429, 16748, 4992, 16783, 12691, 4945, 8779, 2805, 9418, 2797, 14357, 5603, 212, 3852, 12666, 1338, 10269, 2388, 8260, 4293, 14454, 7677, 11253, 5060, 14938, 8840, 4542, 2627, 16336, 8992, 15496, 11140, 446, 6126, 10691, 8624, 10127, 9068, 16710, 10155, 14358, 7567, 5695, 2354, 8057, 17635, 133, 16183, 14535, 7248, 4560, 14429, 2463, 10773, 113, 2462, 9223, 4929, 14274, 4716, 17307, 4617, 2132, 11083, 1039, 1403, 9621, 13936, 2229, 2875, 17840, 9359, 13311, 9790, 13288, 4750, 17052, 8260, 14900]
+    token_label_range = [0.2,0.6]
+    enable_time_shift = False # shift time
+    enable_label_enhance = False # enhance hierarchical label
+    enable_repeat_mode = False # repeat the spectrogram / reshape the spectrogram
+
+
+
+    # for model's design
+    enable_tscam = True # enbale the token-semantic layer
+
+    # for signal processing
+    sample_rate = 32000 # 16000 for scv2, 32000 for audioset and esc-50
+    clip_samples = sample_rate * 10 # audio_set 10-sec clip
+    window_size = 1024
+    hop_size = 320 # 160 for scv2, 320 for audioset and esc-50
+    mel_bins = 64
+    fmin = 50
+    fmax = 14000
+    shift_max = int(clip_samples * 0.5)
+
+    # for data collection
+    classes_num = 527 # esc: 50 | audioset: 527 | scv2: 35
+    patch_size = (25, 4) # deprecated
+    crop_size = None # int(clip_samples * 0.5) deprecated
+
+    # for htsat hyperparamater
+    htsat_window_size = 8
+    htsat_spec_size =  256
+    htsat_patch_size = 4 
+    htsat_stride = (4, 4)
+    htsat_num_head = [4,8,16,32]
+    htsat_dim = 96 
+    htsat_depth = [2,2,6,2]
+
+    swin_pretrain_path = None
+    # "/home/Research/model_backup/pretrain/swin_tiny_c24_patch4_window8_256.pth"
+
+    # Some Deprecated Optimization in the model design, check the model code for details
+    htsat_attn_heatmap = False
+    htsat_hier_output = False 
+    htsat_use_max = False
+
+
+    # for ensemble test 
+
+    ensemble_checkpoints = []
+    ensemble_strides = []
+
+
+    # weight average folder
+    wa_folder = "/home/version_0/checkpoints/"
+    # weight average output filename
+    wa_model_path = "HTSAT_AudioSet_Saved_x.ckpt"
+
+    esm_model_pathes = [
+        "/home/Research/model_backup/AudioSet/HTSAT_AudioSet_Saved_1.ckpt",
+        "/home/Research/model_backup/AudioSet/HTSAT_AudioSet_Saved_2.ckpt",
+        "/home/Research/model_backup/AudioSet/HTSAT_AudioSet_Saved_3.ckpt",
+        "/home/Research/model_backup/AudioSet/HTSAT_AudioSet_Saved_4.ckpt",
+        "/home/Research/model_backup/AudioSet/HTSAT_AudioSet_Saved_5.ckpt",
+        "/home/Research/model_backup/AudioSet/HTSAT_AudioSet_Saved_6.ckpt"
+    ]
+
+    # for framewise localization
+    heatmap_dir = "/home/Research/heatmap_output"
+    test_file = "htsat-test-ensemble"
+    fl_local = False # indicate if we need to use this dataset for the framewise detection
+    fl_dataset = "/home/Research/desed/desedim_embval.npy"  
+    fl_class_num = [
+        "Speech", "Frying", "Dishes", "Running_water",
+        "Blender", "Electric_shaver_toothbrush", "Alarm_bell_ringing",
+        "Cat", "Dog", "Vacuum_cleaner"
+    ]
+
+    # map 527 classes into 10 classes
+    fl_audioset_mapping = [
+        [0,1,2,3,4,5,6,7],
+        [366, 367, 368],
+        [364],
+        [288, 289, 290, 291, 292, 293, 294, 295, 296, 297],
+        [369],
+        [382],
+        [310, 388, 389, 390, 391, 392, 393, 394, 395, 396, 397, 398, 399, 400, 401, 402],
+        [81, 82, 83, 84, 85],
+        [74, 75, 76, 77, 78, 79],
+        [377]
+    ]
+
+
+
+def _ntuple(n):
+    def parse(x):
+        if isinstance(x, collections.abc.Iterable):
+            return x
+        return tuple(repeat(x, n))
+    return parse
+
+to_1tuple = _ntuple(1)
+to_2tuple = _ntuple(2)
+to_3tuple = _ntuple(3)
+to_4tuple = _ntuple(4)
+to_ntuple = _ntuple
+
+def do_mixup(x, mixup_lambda):
+    """Mixup x of even indexes (0, 2, 4, ...) with x of odd indexes 
+    (1, 3, 5, ...).
+    Args:
+      x: (batch_size * 2, ...)
+      mixup_lambda: (batch_size * 2,)
+    Returns:
+      out: (batch_size, ...)
+    """
+    out = (x[0 :: 2].transpose(0, -1) * mixup_lambda[0 :: 2] + \
+        x[1 :: 2].transpose(0, -1) * mixup_lambda[1 :: 2]).transpose(0, -1)
+    return out
+
+def interpolate(x, ratio):
+    """Interpolate data in time domain. This is used to compensate the 
+    resolution reduction in downsampling of a CNN.
+    
+    Args:
+      x: (batch_size, time_steps, classes_num)
+      ratio: int, ratio to interpolate
+    Returns:
+      upsampled: (batch_size, time_steps * ratio, classes_num)
+    """
+    (batch_size, time_steps, classes_num) = x.shape
+    upsampled = x[:, :, None, :].repeat(1, 1, ratio, 1)
+    upsampled = upsampled.reshape(batch_size, time_steps * ratio, classes_num)
+    return upsampled
+
+
+def drop_path(x, drop_prob: float = 0., training: bool = False):
+    """Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).
+    This is the same as the DropConnect impl I created for EfficientNet, etc networks, however,
+    the original name is misleading as 'Drop Connect' is a different form of dropout in a separate paper...
+    See discussion: https://github.com/tensorflow/tpu/issues/494#issuecomment-532968956 ... I've opted for
+    changing the layer and argument names to 'drop path' rather than mix DropConnect as a layer name and use
+    'survival rate' as the argument.
+    """
+    if drop_prob == 0. or not training:
+        return x
+    keep_prob = 1 - drop_prob
+    shape = (x.shape[0],) + (1,) * (x.ndim - 1)  # work with diff dim tensors, not just 2D ConvNets
+    random_tensor = keep_prob + torch.rand(shape, dtype=x.dtype, device=x.device)
+    random_tensor.floor_()  # binarize
+    output = x.div(keep_prob) * random_tensor
+    return output
+
+
+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)
+
+class PatchEmbed(nn.Module):
+    """ 2D Image to Patch Embedding
+    """
+    def __init__(self, img_size=224, patch_size=16, in_chans=3, embed_dim=768, norm_layer=None, flatten=True, patch_stride = 16):
+        super().__init__()
+        img_size = to_2tuple(img_size)
+        patch_size = to_2tuple(patch_size)
+        patch_stride = to_2tuple(patch_stride)
+        self.img_size = img_size
+        self.patch_size = patch_size
+        self.patch_stride = patch_stride
+        self.grid_size = (img_size[0] // patch_stride[0], img_size[1] // patch_stride[1])
+        self.num_patches = self.grid_size[0] * self.grid_size[1]
+        self.flatten = flatten
+        self.in_chans = in_chans
+        self.embed_dim = embed_dim
+        
+        padding = ((patch_size[0] - patch_stride[0]) // 2, (patch_size[1] - patch_stride[1]) // 2)
+
+        self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_stride, padding=padding)
+        self.norm = norm_layer(embed_dim) if norm_layer else nn.Identity()
+
+    def forward(self, x):
+        B, C, H, W = x.shape
+        assert H == self.img_size[0] and W == self.img_size[1], \
+            f"Input image size ({H}*{W}) doesn't match model ({self.img_size[0]}*{self.img_size[1]})."
+        x = self.proj(x)
+        if self.flatten:
+            x = x.flatten(2).transpose(1, 2)  # BCHW -> BNC
+        x = self.norm(x)
+        return x
+
+class Mlp(nn.Module):
+    """ MLP as used in Vision Transformer, MLP-Mixer and related networks
+    """
+    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.drop(x)
+        x = self.fc2(x)
+        x = self.drop(x)
+        return x
+
+def _no_gradim_audiorunc_normal_(tensor, mean, std, a, b):
+    # Cut & paste from PyTorch official master until it's in a few official releases - RW
+    # Method based on https://people.sc.fsu.edu/~jburkardt/presentations/truncated_normal.pdf
+    def norm_cdf(x):
+        # Computes standard normal cumulative distribution function
+        return (1. + math.erf(x / math.sqrt(2.))) / 2.
+
+    if (mean < a - 2 * std) or (mean > b + 2 * std):
+        warnings.warn("mean is more than 2 std from [a, b] in nn.init.trunc_normal_. "
+                      "The distribution of values may be incorrect.",
+                      stacklevel=2)
+
+    with torch.no_grad():
+        # Values are generated by using a truncated uniform distribution and
+        # then using the inverse CDF for the normal distribution.
+        # Get upper and lower cdf values
+        l = norm_cdf((a - mean) / std)
+        u = norm_cdf((b - mean) / std)
+
+        # Uniformly fill tensor with values from [l, u], then translate to
+        # [2l-1, 2u-1].
+        tensor.uniform_(2 * l - 1, 2 * u - 1)
+
+        # Use inverse cdf transform for normal distribution to get truncated
+        # standard normal
+        tensor.erfinv_()
+
+        # Transform to proper mean, std
+        tensor.mul_(std * math.sqrt(2.))
+        tensor.add_(mean)
+
+        # Clamp to ensure it's in the proper range
+        tensor.clamp_(min=a, max=b)
+        return tensor
+
+
+def trunc_normal_(tensor, mean=0., std=1., a=-2., b=2.):
+    # type: (Tensor, float, float, float, float) -> Tensor
+    r"""Fills the input Tensor with values drawn from a truncated
+    normal distribution. The values are effectively drawn from the
+    normal distribution :math:`\mathcal{N}(\text{mean}, \text{std}^2)`
+    with values outside :math:`[a, b]` redrawn until they are within
+    the bounds. The method used for generating the random values works
+    best when :math:`a \leq \text{mean} \leq b`.
+    Args:
+        tensor: an n-dimensional `torch.Tensor`
+        mean: the mean of the normal distribution
+        std: the standard deviation of the normal distribution
+        a: the minimum cutoff value
+        b: the maximum cutoff value
+    Examples:
+        >>> w = torch.empty(3, 5)
+        >>> nn.init.trunc_normal_(w)
+    """
+    return _no_gradim_audiorunc_normal_(tensor, mean, std, a, b)
+
+
+def variance_scaling_(tensor, scale=1.0, mode='fan_in', distribution='normal'):
+    fan_in, fan_out = _calculate_fan_in_and_fan_out(tensor)
+    if mode == 'fan_in':
+        denom = fan_in
+    elif mode == 'fan_out':
+        denom = fan_out
+    elif mode == 'fan_avg':
+        denom = (fan_in + fan_out) / 2
+
+    variance = scale / denom
+
+    if distribution == "truncated_normal":
+        # constant is stddev of standard normal truncated to (-2, 2)
+        trunc_normal_(tensor, std=math.sqrt(variance) / .87962566103423978)
+    elif distribution == "normal":
+        tensor.normal_(std=math.sqrt(variance))
+    elif distribution == "uniform":
+        bound = math.sqrt(3 * variance)
+        tensor.uniform_(-bound, bound)
+    else:
+        raise ValueError(f"invalid distribution {distribution}")
+
+
+def lecun_normal_(tensor):
+    variance_scaling_(tensor, mode='fan_in', distribution='truncated_normal')
+
+
+# below codes are based and referred from https://github.com/microsoft/Swin-Transformer
+# Swin Transformer for Computer Vision: https://arxiv.org/pdf/2103.14030.pdf
+
+def window_partition(x, window_size):
+    """
+    Args:
+        x: (B, H, W, C)
+        window_size (int): window size
+    Returns:
+        windows: (num_windows*B, window_size, window_size, C)
+    """
+    B, H, W, C = x.shape
+    x = x.view(B, H // window_size, window_size, W // window_size, window_size, C)
+    windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C)
+    return windows
+
+
+def window_reverse(windows, window_size, H, W):
+    """
+    Args:
+        windows: (num_windows*B, window_size, window_size, C)
+        window_size (int): Window size
+        H (int): Height of image
+        W (int): Width of image
+    Returns:
+        x: (B, H, W, C)
+    """
+    B = int(windows.shape[0] / (H * W / window_size / window_size))
+    x = windows.view(B, H // window_size, W // window_size, window_size, window_size, -1)
+    x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, -1)
+    return x
+
+
+class WindowAttention(nn.Module):
+    r""" Window based multi-head self attention (W-MSA) module with relative position bias.
+    It supports both of shifted and non-shifted window.
+    Args:
+        dim (int): Number of input channels.
+        window_size (tuple[int]): The height and width of the window.
+        num_heads (int): Number of attention heads.
+        qkv_bias (bool, optional):  If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set
+        attn_drop (float, optional): Dropout ratio of attention weight. Default: 0.0
+        proj_drop (float, optional): Dropout ratio of output. Default: 0.0
+    """
+
+    def __init__(self, dim, window_size, num_heads, qkv_bias=True, qk_scale=None, attn_drop=0., proj_drop=0.):
+
+        super().__init__()
+        self.dim = dim
+        self.window_size = window_size  # Wh, Ww
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+        self.scale = qk_scale or head_dim ** -0.5
+
+        # define a parameter table of relative position bias
+        self.relative_position_bias_table = nn.Parameter(
+            torch.zeros((2 * window_size[0] - 1) * (2 * window_size[1] - 1), num_heads))  # 2*Wh-1 * 2*Ww-1, nH
+
+        # get pair-wise relative position index for each token inside the window
+        coords_h = torch.arange(self.window_size[0])
+        coords_w = torch.arange(self.window_size[1])
+        coords = torch.stack(torch.meshgrid([coords_h, coords_w]))  # 2, Wh, Ww
+        coords_flatten = torch.flatten(coords, 1)  # 2, Wh*Ww
+        relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :]  # 2, Wh*Ww, Wh*Ww
+        relative_coords = relative_coords.permute(1, 2, 0).contiguous()  # Wh*Ww, Wh*Ww, 2
+        relative_coords[:, :, 0] += self.window_size[0] - 1  # shift to start from 0
+        relative_coords[:, :, 1] += self.window_size[1] - 1
+        relative_coords[:, :, 0] *= 2 * self.window_size[1] - 1
+        relative_position_index = relative_coords.sum(-1)  # Wh*Ww, Wh*Ww
+        self.register_buffer("relative_position_index", relative_position_index)
+
+        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
+        self.attn_drop = nn.Dropout(attn_drop)
+        self.proj = nn.Linear(dim, dim)
+        self.proj_drop = nn.Dropout(proj_drop)
+
+        trunc_normal_(self.relative_position_bias_table, std=.02)
+        self.softmax = nn.Softmax(dim=-1)
+
+    def forward(self, x, mask=None):
+        """
+        Args:
+            x: input features with shape of (num_windows*B, N, C)
+            mask: (0/-inf) mask with shape of (num_windows, Wh*Ww, Wh*Ww) or None
+        """
+        B_, N, C = x.shape
+        qkv = self.qkv(x).reshape(B_, N, 3, self.num_heads, C // self.num_heads).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))
+
+        relative_position_bias = self.relative_position_bias_table[self.relative_position_index.view(-1)].view(
+            self.window_size[0] * self.window_size[1], self.window_size[0] * self.window_size[1], -1)  # Wh*Ww,Wh*Ww,nH
+        relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous()  # nH, Wh*Ww, Wh*Ww
+        attn = attn + relative_position_bias.unsqueeze(0)
+
+        if mask is not None:
+            nW = mask.shape[0]
+            attn = attn.view(B_ // nW, nW, self.num_heads, N, N) + mask.unsqueeze(1).unsqueeze(0)
+            attn = attn.view(-1, self.num_heads, N, N)
+            attn = self.softmax(attn)
+        else:
+            attn = self.softmax(attn)
+
+        attn = self.attn_drop(attn)
+
+        x = (attn @ v).transpose(1, 2).reshape(B_, N, C)
+        x = self.proj(x)
+        x = self.proj_drop(x)
+        return x, attn
+
+    def extra_repr(self):
+        return f'dim={self.dim}, window_size={self.window_size}, num_heads={self.num_heads}'
+
+
+# We use the model based on Swintransformer Block, therefore we can use the swin-transformer pretrained model
+class SwinTransformerBlock(nn.Module):
+    r""" Swin Transformer Block.
+    Args:
+        dim (int): Number of input channels.
+        input_resolution (tuple[int]): Input resulotion.
+        num_heads (int): Number of attention heads.
+        window_size (int): Window size.
+        shift_size (int): Shift size for SW-MSA.
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
+        qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
+        drop (float, optional): Dropout rate. Default: 0.0
+        attn_drop (float, optional): Attention dropout rate. Default: 0.0
+        drop_path (float, optional): Stochastic depth rate. Default: 0.0
+        act_layer (nn.Module, optional): Activation layer. Default: nn.GELU
+        norm_layer (nn.Module, optional): Normalization layer.  Default: nn.LayerNorm
+    """
+
+    def __init__(self, dim, input_resolution, num_heads, window_size=7, shift_size=0,
+                 mlp_ratio=4., qkv_bias=True, qk_scale=None, drop=0., attn_drop=0., drop_path=0.,
+                 act_layer=nn.GELU, norm_layer=nn.LayerNorm, norm_before_mlp='ln'):
+        super().__init__()
+        self.dim = dim
+        self.input_resolution = input_resolution
+        self.num_heads = num_heads
+        self.window_size = window_size
+        self.shift_size = shift_size
+        self.mlp_ratio = mlp_ratio
+        self.norm_before_mlp = norm_before_mlp
+        if min(self.input_resolution) <= self.window_size:
+            # if window size is larger than input resolution, we don't partition windows
+            self.shift_size = 0
+            self.window_size = min(self.input_resolution)
+        assert 0 <= self.shift_size < self.window_size, "shift_size must in 0-window_size"
+
+        self.norm1 = norm_layer(dim)
+        self.attn = WindowAttention(
+            dim, window_size=to_2tuple(self.window_size), num_heads=num_heads,
+            qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop)
+
+        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
+        if self.norm_before_mlp == 'ln':
+            self.norm2 = nn.LayerNorm(dim)
+        elif self.norm_before_mlp == 'bn':
+            self.norm2 = lambda x: nn.BatchNorm1d(dim)(x.transpose(1, 2)).transpose(1, 2)
+        else:
+            raise NotImplementedError
+        mlp_hidden_dim = int(dim * mlp_ratio)
+        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
+
+        if self.shift_size > 0:
+            # calculate attention mask for SW-MSA
+            H, W = self.input_resolution
+            img_mask = torch.zeros((1, H, W, 1))  # 1 H W 1
+            h_slices = (slice(0, -self.window_size),
+                        slice(-self.window_size, -self.shift_size),
+                        slice(-self.shift_size, None))
+            w_slices = (slice(0, -self.window_size),
+                        slice(-self.window_size, -self.shift_size),
+                        slice(-self.shift_size, None))
+            cnt = 0
+            for h in h_slices:
+                for w in w_slices:
+                    img_mask[:, h, w, :] = cnt
+                    cnt += 1
+
+            mask_windows = window_partition(img_mask, self.window_size)  # nW, window_size, window_size, 1
+            mask_windows = mask_windows.view(-1, self.window_size * self.window_size)
+            attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2)
+            attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(attn_mask == 0, float(0.0))
+        else:
+            attn_mask = None
+
+        self.register_buffer("attn_mask", attn_mask)
+
+    def forward(self, x):
+        # pdb.set_trace()
+        H, W = self.input_resolution
+        # print("H: ", H)
+        # print("W: ", W)
+        # pdb.set_trace()
+        B, L, C = x.shape
+        # assert L == H * W, "input feature has wrong size"
+
+        shortcut = x
+        x = self.norm1(x)
+        x = x.view(B, H, W, C)
+
+        # cyclic shift
+        if self.shift_size > 0:
+            shifted_x = torch.roll(x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2))
+        else:
+            shifted_x = x
+
+        # partition windows
+        x_windows = window_partition(shifted_x, self.window_size)  # nW*B, window_size, window_size, C
+        x_windows = x_windows.view(-1, self.window_size * self.window_size, C)  # nW*B, window_size*window_size, C
+
+        # W-MSA/SW-MSA
+        attn_windows, attn = self.attn(x_windows, mask=self.attn_mask)  # nW*B, window_size*window_size, C
+
+        # merge windows
+        attn_windows = attn_windows.view(-1, self.window_size, self.window_size, C)
+        shifted_x = window_reverse(attn_windows, self.window_size, H, W)  # B H' W' C
+
+        # reverse cyclic shift
+        if self.shift_size > 0:
+            x = torch.roll(shifted_x, shifts=(self.shift_size, self.shift_size), dims=(1, 2))
+        else:
+            x = shifted_x
+        x = x.view(B, H * W, C)
+
+        # FFN
+        x = shortcut + self.drop_path(x)
+        x = x + self.drop_path(self.mlp(self.norm2(x)))
+
+        return x, attn
+
+    def extra_repr(self):
+        return f"dim={self.dim}, input_resolution={self.input_resolution}, num_heads={self.num_heads}, " \
+               f"window_size={self.window_size}, shift_size={self.shift_size}, mlp_ratio={self.mlp_ratio}"
+
+
+
+class PatchMerging(nn.Module):
+    r""" Patch Merging Layer.
+    Args:
+        input_resolution (tuple[int]): Resolution of input feature.
+        dim (int): Number of input channels.
+        norm_layer (nn.Module, optional): Normalization layer.  Default: nn.LayerNorm
+    """
+
+    def __init__(self, input_resolution, dim, norm_layer=nn.LayerNorm):
+        super().__init__()
+        self.input_resolution = input_resolution
+        self.dim = dim
+        self.reduction = nn.Linear(4 * dim, 2 * dim, bias=False)
+        self.norm = norm_layer(4 * dim)
+
+    def forward(self, x):
+        """
+        x: B, H*W, C
+        """
+        H, W = self.input_resolution
+        B, L, C = x.shape
+        assert L == H * W, "input feature has wrong size"
+        assert H % 2 == 0 and W % 2 == 0, f"x size ({H}*{W}) are not even."
+
+        x = x.view(B, H, W, C)
+
+        x0 = x[:, 0::2, 0::2, :]  # B H/2 W/2 C
+        x1 = x[:, 1::2, 0::2, :]  # B H/2 W/2 C
+        x2 = x[:, 0::2, 1::2, :]  # B H/2 W/2 C
+        x3 = x[:, 1::2, 1::2, :]  # B H/2 W/2 C
+        x = torch.cat([x0, x1, x2, x3], -1)  # B H/2 W/2 4*C
+        x = x.view(B, -1, 4 * C)  # B H/2*W/2 4*C
+
+        x = self.norm(x)
+        x = self.reduction(x)
+
+        return x
+
+    def extra_repr(self):
+        return f"input_resolution={self.input_resolution}, dim={self.dim}"
+
+
+class BasicLayer(nn.Module):
+    """ A basic Swin Transformer layer for one stage.
+    Args:
+        dim (int): Number of input channels.
+        input_resolution (tuple[int]): Input resolution.
+        depth (int): Number of blocks.
+        num_heads (int): Number of attention heads.
+        window_size (int): Local window size.
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
+        qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
+        drop (float, optional): Dropout rate. Default: 0.0
+        attn_drop (float, optional): Attention dropout rate. Default: 0.0
+        drop_path (float | tuple[float], optional): Stochastic depth rate. Default: 0.0
+        norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
+        downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default: None
+        use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False.
+    """
+
+    def __init__(self, dim, input_resolution, depth, num_heads, window_size,
+                 mlp_ratio=4., qkv_bias=True, qk_scale=None, drop=0., attn_drop=0.,
+                 drop_path=0., norm_layer=nn.LayerNorm, downsample=None, use_checkpoint=False,
+                 norm_before_mlp='ln'):
+
+        super().__init__()
+        self.dim = dim
+        self.input_resolution = input_resolution
+        self.depth = depth
+        self.use_checkpoint = use_checkpoint
+
+        # build blocks
+        self.blocks = nn.ModuleList([
+            SwinTransformerBlock(dim=dim, input_resolution=input_resolution,
+                                 num_heads=num_heads, window_size=window_size,
+                                 shift_size=0 if (i % 2 == 0) else window_size // 2,
+                                 mlp_ratio=mlp_ratio,
+                                 qkv_bias=qkv_bias, qk_scale=qk_scale,
+                                 drop=drop, attn_drop=attn_drop,
+                                 drop_path=drop_path[i] if isinstance(drop_path, list) else drop_path,
+                                 norm_layer=norm_layer, norm_before_mlp=norm_before_mlp)
+            for i in range(depth)])
+
+        # patch merging layer
+        if downsample is not None:
+            self.downsample = downsample(input_resolution, dim=dim, norm_layer=norm_layer)
+        else:
+            self.downsample = None
+
+    def forward(self, x):
+        attns = []
+        for blk in self.blocks:
+            if self.use_checkpoint:
+                x = checkpoint.checkpoint(blk, x)
+            else:
+                x, attn = blk(x)
+                if not self.training:
+                    attns.append(attn.unsqueeze(0))
+        if self.downsample is not None:
+            x = self.downsample(x)
+        if not self.training:
+            attn = torch.cat(attns, dim = 0)
+            attn = torch.mean(attn, dim = 0)
+        return x, attn
+
+    def extra_repr(self):
+        return f"dim={self.dim}, input_resolution={self.input_resolution}, depth={self.depth}"
+
+
+# The Core of HTSAT
+class HTSAT_Swin_Transformer(nn.Module):
+    r"""HTSAT based on the Swin Transformer
+    Args:
+        spec_size (int | tuple(int)): Input Spectrogram size. Default 256
+        patch_size (int | tuple(int)): Patch size. Default: 4
+        path_stride (iot | tuple(int)): Patch Stride for Frequency and Time Axis. Default: 4
+        in_chans (int): Number of input image channels. Default: 1 (mono)
+        num_classes (int): Number of classes for classification head. Default: 527
+        embed_dim (int): Patch embedding dimension. Default: 96
+        depths (tuple(int)): Depth of each HTSAT-Swin Transformer layer.
+        num_heads (tuple(int)): Number of attention heads in different layers.
+        window_size (int): Window size. Default: 8
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4
+        qkv_bias (bool): If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float): Override default qk scale of head_dim ** -0.5 if set. Default: None
+        drop_rate (float): Dropout rate. Default: 0
+        attn_drop_rate (float): Attention dropout rate. Default: 0
+        drop_path_rate (float): Stochastic depth rate. Default: 0.1
+        norm_layer (nn.Module): Normalization layer. Default: nn.LayerNorm.
+        ape (bool): If True, add absolute position embedding to the patch embedding. Default: False
+        patch_norm (bool): If True, add normalization after patch embedding. Default: True
+        use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False
+        config (module): The configuration Module from config.py (HTSATConfig Class)
+    """
+
+    def __init__(self, spec_size=256, patch_size=4, patch_stride=(4,4), 
+                in_chans=1, num_classes=527,
+                 embed_dim=96, depths=[2, 2, 6, 2], num_heads=[4, 8, 16, 32],
+                 window_size=8, mlp_ratio=4., qkv_bias=True, qk_scale=None,
+                 drop_rate=0., attn_drop_rate=0., drop_path_rate=0.1,
+                 norm_layer=nn.LayerNorm, 
+                 ape=False, patch_norm=True,
+                 use_checkpoint=False, norm_before_mlp='ln', config = None, **kwargs):
+        super(HTSAT_Swin_Transformer, self).__init__()
+
+        self.config = config
+        self.spec_size = spec_size 
+        self.patch_stride = patch_stride
+        self.patch_size = patch_size
+        self.window_size = window_size
+        self.embed_dim = embed_dim
+        self.depths = depths
+        self.ape = ape
+        self.in_chans = in_chans
+        self.num_classes = num_classes
+        self.num_heads = num_heads
+        self.num_layers = len(self.depths)
+        self.num_features = int(self.embed_dim * 2 ** (self.num_layers - 1))
+        
+        self.drop_rate = drop_rate
+        self.attn_drop_rate = attn_drop_rate
+        self.drop_path_rate = drop_path_rate
+
+        self.qkv_bias = qkv_bias
+        self.qk_scale = None
+
+        self.patch_norm = patch_norm
+        self.norm_layer = norm_layer if self.patch_norm else None
+        self.norm_before_mlp = norm_before_mlp
+        self.mlp_ratio = mlp_ratio
+
+        self.use_checkpoint = use_checkpoint
+
+        #  process mel-spec ; used only once
+        self.freq_ratio = self.spec_size // self.config.mel_bins
+        window = 'hann'
+        center = True
+        pad_mode = 'reflect'
+        ref = 1.0
+        amin = 1e-10
+        top_db = None
+        self.interpolate_ratio = 32     # Downsampled ratio
+        # Spectrogram extractor
+        self.spectrogram_extractor = Spectrogram(n_fft=config.window_size, hop_length=config.hop_size, 
+            win_length=config.window_size, window=window, center=center, pad_mode=pad_mode, 
+            freeze_parameters=True)
+        # Logmel feature extractor
+        self.logmel_extractor = LogmelFilterBank(sr=config.sample_rate, n_fft=config.window_size, 
+            n_mels=config.mel_bins, fmin=config.fmin, fmax=config.fmax, ref=ref, amin=amin, top_db=top_db, 
+            freeze_parameters=True)
+        # Spec augmenter
+        self.spec_augmenter = SpecAugmentation(time_drop_width=64, time_stripes_num=2, 
+            freq_drop_width=8, freq_stripes_num=2) # 2 2
+        self.bn0 = nn.BatchNorm2d(self.config.mel_bins)
+
+
+        # split spctrogram into non-overlapping patches
+        self.patch_embed = PatchEmbed(
+            img_size=self.spec_size, patch_size=self.patch_size, in_chans=self.in_chans, 
+            embed_dim=self.embed_dim, norm_layer=self.norm_layer, patch_stride = patch_stride)
+
+        num_patches = self.patch_embed.num_patches
+        patches_resolution = self.patch_embed.grid_size
+        self.patches_resolution = patches_resolution
+
+        # absolute position embedding
+        if self.ape:
+            self.absolute_pos_embed = nn.Parameter(torch.zeros(1, num_patches, self.embed_dim))
+            trunc_normal_(self.absolute_pos_embed, std=.02)
+
+        self.pos_drop = nn.Dropout(p=self.drop_rate)
+
+        # stochastic depth
+        dpr = [x.item() for x in torch.linspace(0, self.drop_path_rate, sum(self.depths))]  # stochastic depth decay rule
+
+        # build layers
+        self.layers = nn.ModuleList()
+        for i_layer in range(self.num_layers):
+            layer = BasicLayer(dim=int(self.embed_dim * 2 ** i_layer),
+                input_resolution=(patches_resolution[0] // (2 ** i_layer),
+                                    patches_resolution[1] // (2 ** i_layer)),
+                depth=self.depths[i_layer],
+                num_heads=self.num_heads[i_layer],
+                window_size=self.window_size,
+                mlp_ratio=self.mlp_ratio,
+                qkv_bias=self.qkv_bias, qk_scale=self.qk_scale,
+                drop=self.drop_rate, attn_drop=self.attn_drop_rate,
+                drop_path=dpr[sum(self.depths[:i_layer]):sum(self.depths[:i_layer + 1])],
+                norm_layer=self.norm_layer,
+                downsample=PatchMerging if (i_layer < self.num_layers - 1) else None,
+                use_checkpoint=use_checkpoint,
+                norm_before_mlp=self.norm_before_mlp)
+            self.layers.append(layer)
+
+        self.norm = self.norm_layer(self.num_features)
+        self.avgpool = nn.AdaptiveAvgPool1d(1)
+        self.maxpool = nn.AdaptiveMaxPool1d(1)
+
+        if self.config.enable_tscam:
+            SF = self.spec_size // (2 ** (len(self.depths) - 1)) // self.patch_stride[0] // self.freq_ratio
+            self.tscam_conv = nn.Conv2d(
+                in_channels = self.num_features,
+                out_channels = self.num_classes,
+                kernel_size = (SF,3),
+                padding = (0,1)
+            )
+            self.head = nn.Linear(num_classes, num_classes)
+        else:
+            self.head = nn.Linear(self.num_features, num_classes) if num_classes > 0 else nn.Identity()
+
+        self.apply(self._init_weights)
+
+    def _init_weights(self, 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)
+
+    @torch.jit.ignore
+    def no_weight_decay(self):
+        return {'absolute_pos_embed'}
+
+    @torch.jit.ignore
+    def no_weight_decay_keywords(self):
+        return {'relative_position_bias_table'}
+
+    def forward_features(self, x):
+        frames_num = x.shape[2]        
+        x = self.patch_embed(x)
+        if self.ape:
+            x = x + self.absolute_pos_embed
+        x = self.pos_drop(x)
+        for i, layer in enumerate(self.layers):
+            x, attn = layer(x)
+
+        if self.config.enable_tscam:
+            # for x
+            x = self.norm(x)
+            B, N, C = x.shape
+            SF = frames_num // (2 ** (len(self.depths) - 1)) // self.patch_stride[0]
+            ST = frames_num // (2 ** (len(self.depths) - 1)) // self.patch_stride[1]
+            x = x.permute(0,2,1).contiguous().reshape(B, C, SF, ST)
+            B, C, F, T = x.shape
+            # group 2D CNN
+            c_freq_bin = F // self.freq_ratio
+            x = x.reshape(B, C, F // c_freq_bin, c_freq_bin, T)
+            x = x.permute(0,1,3,2,4).contiguous().reshape(B, C, c_freq_bin, -1)
+
+            # get latent_output
+            latent_output = self.avgpool(torch.flatten(x,2))
+            latent_output = torch.flatten(latent_output, 1)
+
+            # display the attention map, if needed
+            if self.config.htsat_attn_heatmap:
+                # for attn
+                attn = torch.mean(attn, dim = 1)
+                attn = torch.mean(attn, dim = 1)
+                attn = attn.reshape(B, SF, ST)
+                c_freq_bin = SF // self.freq_ratio
+                attn = attn.reshape(B, SF // c_freq_bin, c_freq_bin, ST) 
+                attn = attn.permute(0,2,1,3).contiguous().reshape(B, c_freq_bin, -1)
+                attn = attn.mean(dim = 1)
+                attn_max = torch.max(attn, dim = 1, keepdim = True)[0]
+                attn_min = torch.min(attn, dim = 1, keepdim = True)[0]
+                attn = ((attn * 0.15) + (attn_max * 0.85 - attn_min)) / (attn_max - attn_min)
+                attn = attn.unsqueeze(dim = 2)
+
+            x = self.tscam_conv(x)
+            x = torch.flatten(x, 2) # B, C, T
+
+            if self.config.htsat_attn_heatmap:
+                fpx = interpolate(torch.sigmoid(x).permute(0,2,1).contiguous() * attn, 8 * self.patch_stride[1]) 
+            else: 
+                fpx = interpolate(torch.sigmoid(x).permute(0,2,1).contiguous(), 8 * self.patch_stride[1]) 
+            
+            x = self.avgpool(x)
+            x = torch.flatten(x, 1)
+
+            if self.config.loss_type == "clip_ce":
+                output_dict = {
+                    'framewise_output': fpx, # already sigmoided
+                    'clipwise_output': x,
+                    'latent_output': latent_output
+                }
+            else:
+                output_dict = {
+                    'framewise_output': fpx, # already sigmoided
+                    'clipwise_output': torch.sigmoid(x),
+                    'latent_output': latent_output
+                }
+           
+        else:
+            x = self.norm(x)  # B N C
+            B, N, C = x.shape
+            
+            fpx = x.permute(0,2,1).contiguous().reshape(B, C, frames_num // (2 ** (len(self.depths) + 1)), frames_num // (2 ** (len(self.depths) + 1)) )
+            B, C, F, T = fpx.shape
+            c_freq_bin = F // self.freq_ratio
+            fpx = fpx.reshape(B, C, F // c_freq_bin, c_freq_bin, T)
+            fpx = fpx.permute(0,1,3,2,4).contiguous().reshape(B, C, c_freq_bin, -1)
+            fpx = torch.sum(fpx, dim = 2)
+            fpx = interpolate(fpx.permute(0,2,1).contiguous(), 8 * self.patch_stride[1]) 
+            x = self.avgpool(x.transpose(1, 2))  # B C 1
+            x = torch.flatten(x, 1)
+            if self.num_classes > 0:
+                x = self.head(x)
+                fpx = self.head(fpx)
+            output_dict = {'framewise_output': torch.sigmoid(fpx), 
+                'clipwise_output': torch.sigmoid(x)}
+        return output_dict
+
+    def crop_wav(self, x, crop_size, spe_pos = None):
+        time_steps = x.shape[2]
+        tx = torch.zeros(x.shape[0], x.shape[1], crop_size, x.shape[3]).to(x.device)
+        for i in range(len(x)):
+            if spe_pos is None:
+                crop_pos = random.randint(0, time_steps - crop_size - 1)
+            else:
+                crop_pos = spe_pos
+            tx[i][0] = x[i, 0, crop_pos:crop_pos + crop_size,:]
+        return tx
+
+    # Reshape the wavform to a img size, if you want to use the pretrained swin transformer model
+    def reshape_wav2img(self, x):
+        B, C, T, F = x.shape
+        target_T = int(self.spec_size * self.freq_ratio)
+        target_F = self.spec_size // self.freq_ratio
+        assert T <= target_T and F <= target_F, "the wav size should less than or equal to the swin input size"
+        # to avoid bicubic zero error
+        if T < target_T:
+            x = nn.functional.interpolate(x, (target_T, x.shape[3]), mode="bicubic", align_corners=True)
+        if F < target_F:
+            x = nn.functional.interpolate(x, (x.shape[2], target_F), mode="bicubic", align_corners=True)
+        x = x.permute(0,1,3,2).contiguous()
+        x = x.reshape(x.shape[0], x.shape[1], x.shape[2], self.freq_ratio, x.shape[3] // self.freq_ratio)
+        # print(x.shape)
+        x = x.permute(0,1,3,2,4).contiguous()
+        x = x.reshape(x.shape[0], x.shape[1], x.shape[2] * x.shape[3], x.shape[4])
+        return x
+    
+    # Repeat the wavform to a img size, if you want to use the pretrained swin transformer model
+    def repeat_wat2img(self, x, cur_pos):
+        B, C, T, F = x.shape
+        target_T = int(self.spec_size * self.freq_ratio)
+        target_F = self.spec_size // self.freq_ratio
+        assert T <= target_T and F <= target_F, "the wav size should less than or equal to the swin input size"
+        # to avoid bicubic zero error
+        if T < target_T:
+            x = nn.functional.interpolate(x, (target_T, x.shape[3]), mode="bicubic", align_corners=True)
+        if F < target_F:
+            x = nn.functional.interpolate(x, (x.shape[2], target_F), mode="bicubic", align_corners=True)  
+        x = x.permute(0,1,3,2).contiguous() # B C F T
+        x = x[:,:,:,cur_pos:cur_pos + self.spec_size]
+        x = x.repeat(repeats = (1,1,4,1))
+        return x
+
+    def forward(self, x: torch.Tensor, mixup_lambda = None, infer_mode = False):# out_feat_keys: List[str] = None):
+        x = self.spectrogram_extractor(x)   # (batch_size, 1, time_steps, freq_bins)
+        x = self.logmel_extractor(x)    # (batch_size, 1, time_steps, mel_bins)
+        
+        
+        x = x.transpose(1, 3)
+        x = self.bn0(x)
+        x = x.transpose(1, 3)
+        if self.training:
+            x = self.spec_augmenter(x)
+        if self.training and mixup_lambda is not None:
+            x = do_mixup(x, mixup_lambda)
+        
+        if infer_mode:
+            # in infer mode. we need to handle different length audio input
+            frame_num = x.shape[2]
+            target_T = int(self.spec_size * self.freq_ratio)
+            repeat_ratio = math.floor(target_T / frame_num)
+            x = x.repeat(repeats=(1,1,repeat_ratio,1))
+            x = self.reshape_wav2img(x)
+            output_dict = self.forward_features(x)
+        elif self.config.enable_repeat_mode:
+            if self.training:
+                cur_pos = random.randint(0, (self.freq_ratio - 1) * self.spec_size - 1)
+                x = self.repeat_wat2img(x, cur_pos)
+                output_dict = self.forward_features(x)
+            else:
+                output_dicts = []
+                for cur_pos in range(0, (self.freq_ratio - 1) * self.spec_size + 1, self.spec_size):
+                    tx = x.clone()
+                    tx = self.repeat_wat2img(tx, cur_pos)
+                    output_dicts.append(self.forward_features(tx))
+                clipwise_output = torch.zeros_like(output_dicts[0]["clipwise_output"]).float().to(x.device)
+                framewise_output = torch.zeros_like(output_dicts[0]["framewise_output"]).float().to(x.device)
+                for d in output_dicts:
+                    clipwise_output += d["clipwise_output"]
+                    framewise_output += d["framewise_output"]
+                clipwise_output  = clipwise_output / len(output_dicts)
+                framewise_output = framewise_output / len(output_dicts)
+
+                output_dict = {
+                    'framewise_output': framewise_output, 
+                    'clipwise_output': clipwise_output
+                }
+        else:
+            if x.shape[2] > self.freq_ratio * self.spec_size:
+                if self.training:
+                    x = self.crop_wav(x, crop_size=self.freq_ratio * self.spec_size)
+                    x = self.reshape_wav2img(x)
+                    output_dict = self.forward_features(x)
+                else:
+                    # Change: Hard code here
+                    overlap_size = 344 #(x.shape[2] - 1) // 4
+                    output_dicts = []
+                    crop_size = 689 #(x.shape[2] - 1) // 2
+                    for cur_pos in range(0, x.shape[2] - crop_size - 1, overlap_size):
+                        tx = self.crop_wav(x, crop_size = crop_size, spe_pos = cur_pos)
+                        tx = self.reshape_wav2img(tx)
+                        output_dicts.append(self.forward_features(tx))
+                    clipwise_output = torch.zeros_like(output_dicts[0]["clipwise_output"]).float().to(x.device)
+                    framewise_output = torch.zeros_like(output_dicts[0]["framewise_output"]).float().to(x.device)
+                    latent_output = torch.zeros_like(output_dicts[0]["latent_output"]).float().to(x.device)
+                    for d in output_dicts:
+                        clipwise_output += d["clipwise_output"]
+                        framewise_output += d["framewise_output"]
+                        latent_output += d["latent_output"]
+                    clipwise_output  = clipwise_output / len(output_dicts)
+                    framewise_output = framewise_output / len(output_dicts)
+                    latent_output = latent_output / len(output_dicts)
+                    output_dict = {
+                        'framewise_output': framewise_output, 
+                        'clipwise_output': clipwise_output,
+                        'latent_output': latent_output,
+                    }
+            else: # this part is typically used, and most easy one
+                x = self.reshape_wav2img(x)
+                output_dict = self.forward_features(x)
+        # x = self.head(x)
+        return output_dict
+
+class HTSATWrapper(nn.Module):
+    def __init__(self, sample_rate, window_size, hop_size, mel_bins, fmin, 
+        fmax, classes_num, out_emb):
+        super().__init__()
+
+        # print("parameters are being overidden when using HTSAT")
+        # print("HTSAT only support loading a pretrained model on AudioSet")
+        # @TODO later look at what parameters are same and can be merged
+
+        self.htsat = HTSAT_Swin_Transformer(config=HTSATConfig())
+
+    def forward(self, x):
+        out_dict = self.htsat(x)
+        out_dict['embedding'] = out_dict['latent_output']
+        return out_dict
+
+
+def get_audio_encoder(name: str):
+    if name == "HTSAT":
+        return HTSATWrapper
+    else:
+        raise Exception('The audio encoder name {} is incorrect or not supported'.format(name))
+
+class Projection(nn.Module):
+    def __init__(self, dim_imgn: int, d_out: int, p: float=0.5) -> None:
+        super().__init__()
+        self.linear1 = nn.Linear(dim_imgn, d_out, bias=False)
+        self.linear2 = nn.Linear(d_out, d_out, bias=False)
+        self.layer_norm = nn.LayerNorm(d_out)
+        self.drop = nn.Dropout(p)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        embed1 = self.linear1(x)
+        embed2 = self.drop(self.linear2(F.gelu(embed1)))
+        embeds = self.layer_norm(embed1 + embed2)
+        return embeds
+
+class AudioEncoder(nn.Module):
+    def __init__(self, audioenc_name:str, dim_imgn: int, d_out: int, sample_rate: int, window_size: int,
+            hop_size: int, mel_bins: int, fmin: int, fmax: int, classes_num: int) -> None:
+        super().__init__()
+
+        audio_encoder = get_audio_encoder(audioenc_name)
+
+        self.base = audio_encoder(
+            sample_rate, window_size,
+            hop_size, mel_bins, fmin, fmax,
+            classes_num, dim_imgn)
+
+        self.projection = Projection(dim_imgn, d_out)
+
+    def forward(self, x):
+        out_dict = self.base(x)
+        audio_features, audio_classification_output = out_dict['embedding'], out_dict['clipwise_output']
+        projected_vec = self.projection(audio_features)
+        return projected_vec, audio_classification_output
+
+class TextEncoder(nn.Module):
+    def __init__(self, d_out: int, text_model: str, transformer_embed_dim: int) -> None:
+        super().__init__()
+        self.text_model = text_model
+        self.base = AutoModel.from_pretrained(text_model)
+
+        if 'clip' in text_model:
+            self.clip_text_projection = self.base.text_projection
+            self.base = self.base.text_model
+            if 'base' in text_model:
+                transformer_embed_dim = 512
+        
+        self.projection = Projection(transformer_embed_dim, d_out)
+
+    def forward(self, x):
+        if 'clip' in self.text_model:
+            pooled_output = self.base(**x)[1] # get pooled output
+            out = self.clip_text_projection(pooled_output)  # get CLS token output
+        elif 'gpt' in self.text_model:
+            batch_size = x['input_ids'].shape[0]
+            hidden_states = self.base(**x)[0] # (batch_size=4, seq_len, 768)
+
+            sequence_lengths = torch.ne(x['input_ids'], 0).sum(-1) - 1 # tensor([13, 14, 18, 17])
+            out = hidden_states[torch.arange(batch_size, device=hidden_states.device), sequence_lengths] # [batch_size, 768] = [4, 768]
+        else:
+            out = self.base(**x)[0]
+            out = out[:, 0, :]  # get CLS token output
+        
+        projected_vec = self.projection(out)
+
+        return projected_vec
+
+class CLAP(nn.Module):
+    def __init__(self,
+                # audio
+                audioenc_name: str,
+                sample_rate: int, 
+                window_size: int, 
+                hop_size: int, 
+                mel_bins: int, 
+                fmin: int, 
+                fmax: int, 
+                classes_num: int, 
+                out_emb: int,
+                # text
+                text_model: str,
+                transformer_embed_dim: int,
+                # common
+                d_proj: int,
+                ):
+        super().__init__()
+
+        
+        self.audio_encoder = AudioEncoder(
+            audioenc_name, out_emb, d_proj,
+            sample_rate, window_size, hop_size, mel_bins, fmin, fmax, classes_num)
+
+        self.caption_encoder = TextEncoder(
+            d_proj, text_model, transformer_embed_dim
+        )
+
+        self.logit_scale = nn.Parameter(torch.ones([]) * np.log(1 / 0.07))
+
+    def forward(self, audio, text):
+        audio_embed, _ = self.audio_encoder(audio)
+        caption_embed = self.caption_encoder(text)
+
+        return caption_embed, audio_embed, self.logit_scale.exp()
+    
+    
+    
+# ==================================================================
+#             A U D I O - P R E - P R O C E S S I N G
+# ==================================================================
+def read_audio(audio_path, resample=True):
+    r"""Loads audio file or array and returns a torch tensor"""
+    # Randomly sample a segment of audio_duration from the clip or pad to match duration
+    audio_time_series, sample_rate = torchaudio.load(audio_path)
+
+    resample_rate = clapConfig.sample_rate
+    if resample and resample_rate != sample_rate:
+        resampler = T.Resample(sample_rate, resample_rate)
+        audio_time_series = resampler(audio_time_series)
+    return audio_time_series, resample_rate
+
+def load_audio_into_tensor(audio_path, audio_duration, resample=False):
+    r"""Loads audio file and returns raw audio."""
+    # Randomly sample a segment of audio_duration from the clip or pad to match duration
+    audio_time_series, sample_rate = read_audio(audio_path, resample)
+    audio_time_series = audio_time_series.reshape(-1)
+
+    # audio_time_series is shorter than predefined audio duration,
+    # so audio_time_series is extended
+    if audio_duration*sample_rate >= audio_time_series.shape[0]:
+        repeat_factor = int(np.ceil((audio_duration*sample_rate) /
+                                    audio_time_series.shape[0]))
+        # Repeat audio_time_series by repeat_factor to match audio_duration
+        audio_time_series = audio_time_series.repeat(repeat_factor)
+        # remove excess part of audio_time_series
+        audio_time_series = audio_time_series[0:audio_duration*sample_rate]
+    else:
+        # audio_time_series is longer than predefined audio duration,
+        # so audio_time_series is trimmed
+        start_index = random.randrange(
+            audio_time_series.shape[0] - audio_duration*sample_rate)
+        audio_time_series = audio_time_series[start_index:start_index +
+                                                audio_duration*sample_rate]
+    return torch.FloatTensor(audio_time_series)
+
+np_str_obj_array_pattern = re.compile(r'[SaUO]')
+default_collate_err_msg_format = (
+    "default_collate: batch must contain tensors, numpy arrays, numbers, "
+    "dicts or lists; found {}")
+
+def default_collate(batch):
+    r"""Puts each data field into a tensor with outer dimension batch size"""
+    elem = batch[0]
+    elem_type = type(elem)
+    if isinstance(elem, torch.Tensor):
+        out = None
+        if torch.utils.data.get_worker_info() is not None:
+            # If we're in a background process, concatenate directly into a
+            # shared memory tensor to avoid an extra copy
+            numel = sum([x.numel() for x in batch])
+            storage = elem.storage()._new_shared(numel)
+            out = elem.new(storage)
+        return torch.stack(batch, 0, out=out)
+    elif elem_type.__module__ == 'numpy' and elem_type.__name__ != 'str_' \
+            and elem_type.__name__ != 'string_':
+        if elem_type.__name__ == 'ndarray' or elem_type.__name__ == 'memmap':
+            # array of string classes and object
+            if np_str_obj_array_pattern.search(elem.dtype.str) is not None:
+                raise TypeError(
+                    default_collate_err_msg_format.format(elem.dtype))
+
+            return default_collate([torch.as_tensor(b) for b in batch])
+        elif elem.shape == ():  # scalars
+            return torch.as_tensor(batch)
+    elif isinstance(elem, float):
+        return torch.tensor(batch, dtype=torch.float64)
+    elif isinstance(elem, int):
+        return torch.tensor(batch)
+    elif isinstance(elem, str):
+        return batch
+    elif isinstance(elem, collections.abc.Mapping):
+        return {key: default_collate([d[key] for d in batch]) for key in elem}
+    elif isinstance(elem, tuple) and hasattr(elem, '_fields'):  # namedtuple
+        return elem_type(*(default_collate(samples) for samples in zip(*batch)))
+    elif isinstance(elem, collections.abc.Sequence):
+        # check to make sure that the elements in batch have consistent size
+        it = iter(batch)
+        elem_size = len(next(it))
+        if not all(len(elem) == elem_size for elem in it):
+            raise RuntimeError(
+                'each element in list of batch should be of equal size')
+        transposed = zip(*batch)
+        return [default_collate(samples) for samples in transposed]
+
+    raise TypeError(default_collate_err_msg_format.format(elem_type))
+
+def preprocess_audio(audio_files, resample):
+    r"""Load list of audio files and return raw audio"""
+    audio_tensors = []
+    for audio_file in audio_files:
+        audio_tensor = load_audio_into_tensor(
+            audio_file, clapConfig.duration, resample)
+        audio_tensor = audio_tensor.reshape(1, -1)
+        audio_tensors.append(audio_tensor)
+    return default_collate(audio_tensors)
+
+
+
+# ==================================================================
+#            A U D I O - E M B E D D I N G S - H E L P E R
+# ==================================================================
+def CLAPAudioProcessor(audio_files: List[str], resample=True):
+    preprocessed_audio = preprocess_audio(audio_files, resample)
+    preprocessed_audio = preprocessed_audio.reshape(
+        preprocessed_audio.shape[0], preprocessed_audio.shape[2])
+    preprocessed_audio = preprocessed_audio
+    return preprocessed_audio
+
+def get_audio_embeddings(audio_files: List[str], audio_encoder, resample=True):
+    """Load list of audio files and return audio embeddings"""
+    # preprocessed_audio = preprocess_audio(audio_files, resample)
+    # with torch.no_grad():
+    #     preprocessed_audio = preprocessed_audio.reshape(
+    #         preprocessed_audio.shape[0], preprocessed_audio.shape[2])
+    with torch.no_grad():
+        preprocessed_audio = CLAPAudioProcessor(audio_files, resample)
+        return audio_encoder(preprocessed_audio)[0]
+
+
+# ==================================================================
+#                            C L A P
+# ==================================================================
+class ClapConfig:
+    # TEXT ENCODER CONFIG
+    text_model = 'gpt2'
+    text_len = 77
+    transformer_embed_dim = 768
+    freeze_text_encoder_weights = True
+
+    # AUDIO ENCODER CONFIG
+    audioenc_name = 'HTSAT'
+    out_emb = 768
+    sample_rate = 44100
+    duration = 7
+    fmin = 50
+    fmax = 8000  # 14000
+    n_fft = 1024  # 1028
+    hop_size = 320
+    mel_bins = 64
+    window_size = 1024
+
+    # PROJECTION SPACE CONFIG 
+    d_proj = 1024
+    temperature = 0.003
+
+    # TRAINING AND EVALUATION CONFIG
+    num_classes = 527
+    batch_size = 1024
+    demo = False
+    
+
+clapConfig = ClapConfig()
+clap = CLAP(
+            audioenc_name=clapConfig.audioenc_name,
+            sample_rate=clapConfig.sample_rate,
+            window_size=clapConfig.window_size,
+            hop_size=clapConfig.hop_size,
+            mel_bins=clapConfig.mel_bins,
+            fmin=clapConfig.fmin,
+            fmax=clapConfig.fmax,
+            classes_num=clapConfig.num_classes,
+            out_emb=clapConfig.out_emb,
+            text_model=clapConfig.text_model,
+            transformer_embed_dim=clapConfig.transformer_embed_dim,
+            d_proj=clapConfig.d_proj
+        )
+
+model_repo = "microsoft/msclap"
+model_name = {
+    '2022': 'CLAP_weights_2022.pth',
+    '2023': 'CLAP_weights_2023.pth',
+    'clapcap': 'clapcap_weights_2023.pth'
+}
+
+version = '2023'
+model_fp = hf_hub_download(model_repo, model_name[version])
+
+model_state_dict = torch.load(model_fp, map_location=torch.device('cpu'))['model']
+clap.load_state_dict(model_state_dict, strict=False)
+clap.eval()
+
+clap_audio_encoder = clap.audio_encoder.eval().to(device)
+
+
+# ENGLISH_AUDIO_DIR = r"/home/IITB/ai-at-ieor/23m1521/datasets/Vaani/Audios/English"
+# audio_files = [os.path.join(ENGLISH_AUDIO_DIR, i) for i in os.listdir(ENGLISH_AUDIO_DIR) if i.endswith(".wav")]
+# audio_embedding = get_audio_embeddings(audio_files, clap_audio_encoder)
+# print("CLAP Audio Encoder Embeddings:", audio_embedding.shape) # [5, 1024]
+
+
+# ==================================================================
+#                 C L A P - L o R A -  M O D E L
+# ==================================================================
+LoRAconfig = {
+    "peft_type": "LORA",
+    "task_type": "FEATURE_EXTRACTION",
+    "inference_mode": False,
+    "r": 16,
+    "target_modules": ["qkv", "fc1", "fc2", "proj", "linear1", "linear2"],
+    "lora_alpha": 32,
+    "lora_dropout": 0.05,
+    "fan_in_fan_out": False,
+    "bias": "all",
+}
+peft_config = get_peft_config(LoRAconfig)
+
+peft_model = get_peft_model(clap_audio_encoder, peft_config)
+
+peft_model.print_trainable_parameters()
+
+peft_clap_audio_encoder = peft_model.base_model
+# audio_embedding = get_audio_embeddings(audio_files, peft_clap_audio_encoder)
+# print("CLAP LoRA Audio Encoder Embeddings:", audio_embedding.shape) # [5, 1024]
+
+
+
+# ==================================================================
+#                      C L I P - M O D E L
+# ==================================================================
+from transformers import CLIPImageProcessorFast, CLIPImageProcessor
+clip_vision_model = CLIPVisionModel.from_pretrained("openai/clip-vit-base-patch32").to(device)
+# clip_vision_processor = AutoProcessor.from_pretrained("openai/clip-vit-base-patch32")
+clip_vision_processor = CLIPImageProcessorFast.from_pretrained("openai/clip-vit-base-patch32")
+# clip_vision_processor = CLIPImageProcessor.from_pretrained("openai/clip-vit-base-patch32")
+
+# image = Image.open("/home/IITB/ai-at-ieor/23m1521/ashish/MTP/Vaani/Img_Audio_Alignment/000000039769.jpg")
+# inputs = clip_vision_processor(images=image, return_tensors="pt")
+# print("CLIP input image:", inputs['pixel_values'].shape)
+# # input_data = {'pixel_values': inputs}
+
+# IMAGE_SIZE = 224
+# dummy_input = torch.randn(1, 3, IMAGE_SIZE, IMAGE_SIZE) # [1, 3, 224, 224]
+# # dummy_input_data = {'pixel_values': dummy_input}
+
+# output = clip_vision_model(inputs['pixel_values'])
+# print("CLIP Image Encoder Embeddings:", output.last_hidden_state.shape) # [1, 50, 768]
+# print("CLIP Image Encoder Pooled Output:", output.pooler_output.shape) # [1, 768]
+
+
+# ==================================================================
+#                       C  S  I  P - M O D U L E
+# ==================================================================
+class CSIP(nn.Module):
+    def __init__(self, image_encoder, audio_encoder, dim_img=None, dim_audio=1024, dim_emb=768):
+        super(CSIP, self).__init__()
+        self.image_encoder = image_encoder     # CLIPVisionModel
+        self.audio_encoder = audio_encoder     # CLAP_audio_encoder
+
+        # self.image_proj = nn.Linear(dim_img, dim_emb)
+        self.audio_proj = nn.Linear(dim_audio, dim_emb)
+
+        # Learnable temperature parameter
+        self.log_temp = nn.Parameter(torch.tensor(0.07).log())
+
+    def forward(self, images, audios):
+        # Step 1: Feature extraction
+        with torch.no_grad():
+            with torch.inference_mode():
+                image_features = self.image_encoder(images).pooler_output     # shape: [n, dim_img]
+        audio_features = self.audio_encoder(audios)[0]                          # shape: [n, dim_audio]
+
+        # Step 2: Project and normalize
+        image_embeds  = F.normalize(image_features)                           # [n, dim_emb]
+        audio_embeds  = F.normalize(self.audio_proj(audio_features), dim=1)    # [n, dim_emb]
+
+        # Step 3: Cosine similarity with temperature
+        logits = torch.matmul(image_embeds, audio_embeds.T) * self.log_temp.exp()  # [n, n]
+        probs = logits.softmax(dim=1)
+
+        # Step 4: Symmetric cross-entropy loss
+        labels = torch.arange(len(images), device=images.device)
+        loss_i = F.cross_entropy(logits, labels)
+        loss_t = F.cross_entropy(logits.T, labels)
+        loss = (loss_i + loss_t) / 2
+        return loss, logits, probs
+
+
+# ==================================================================
+#                 I M A G E - A U D I O - D A T A S E T
+# ==================================================================
+class VaaniImageAudioDataset(torch.utils.data.Dataset):
+    def __init__(self, df):
+        self.image_paths = df.image_path.tolist()
+        self.audio_paths = df.audio_path.tolist()
+
+    def __len__(self):
+        return len(self.audio_paths)
+
+    def __getitem__(self, idx):
+        return {
+            'image_path': self.image_paths[idx], 
+            'audio_path': self.audio_paths[idx]
+        }
+
+
+def collate_fn(batch):
+    image_tensor = clip_vision_processor([Image.open(item['image_path']) for item in batch])['pixel_values']
+    audio_tensor = CLAPAudioProcessor([item['audio_path'] for item in batch], resample=True)
+    return {'image_tensor': torch.stack(image_tensor), 'audio_tensor': audio_tensor}
+
+
+# preprocessed_audio = CLAPAudioProcessor(audio_files, resample=True)
+# clip_vision_processor = clip_vision_processor
+
+train_df = pd.read_csv("/home/IITB/ai-at-ieor/23m1521/ashish/MTP/Vaani/Img_Audio_Alignment/available_img_audios_TRAIN.csv")
+test_df = pd.read_csv("/home/IITB/ai-at-ieor/23m1521/ashish/MTP/Vaani/Img_Audio_Alignment/available_img_audios_TEST.csv")
+train_dataset = VaaniImageAudioDataset(train_df)
+test_dataset = VaaniImageAudioDataset(test_df)
+BATCH_SIZE = 32
+
+print('Train Dataset:', len(train_dataset))
+print('Test Dataset:', len(test_dataset))
+
+
+train_dataloader = torch.utils.data.DataLoader(
+    train_dataset,
+    batch_size=BATCH_SIZE, 
+    shuffle=True, 
+    num_workers=48,
+    collate_fn=collate_fn,
+    pin_memory=True,
+    drop_last=True,
+    persistent_workers=True
+)
+
+test_dataloader = torch.utils.data.DataLoader(
+    test_dataset,
+    batch_size=BATCH_SIZE, 
+    shuffle=False, 
+    num_workers=48,
+    collate_fn=collate_fn,
+    pin_memory=True,
+    drop_last=False,
+    persistent_workers=True
+)
+
+batch = next(iter(train_dataloader))
+print("Image batch:", batch['image_tensor'].shape)
+print("Audio batch:", batch['audio_tensor'].shape)
+
+
+csip_model = CSIP(clip_vision_model.eval(), peft_clap_audio_encoder).to(device)
+
+
+# loss, logits, probs = csip_model(batch['image_tensor'].to(device), batch['audio_tensor'].to(device))
+# loss, logits, probs, logits.shape, probs.shape
+
+
+def train_batch(model, images, audio, optimizer):
+    model.train()
+    optimizer.zero_grad()
+    loss, logits, probs = model(images, audio)
+    loss.backward()
+    optimizer.step()
+    return loss.item(), logits, probs
+
+@torch.no_grad()
+def evaluate_batch(model, images, audio):
+    model.eval()
+    loss, logits, probs = model(images, audio)
+    return loss.item(), logits, probs
+
+def save_checkpoint(state, checkpoint_dir, epoch, max_checkpoints=2):
+    filename = f"csip_best_epoch_{epoch+1}.pt"
+    path = os.path.join(checkpoint_dir, filename)
+    torch.save(state, path)
+    checkpoints = sorted(
+        [f for f in os.listdir(checkpoint_dir) if f.startswith("csip_best_epoch_")],
+        key=lambda x: int(x.split("_")[-1].split(".")[0])
+    )
+    while len(checkpoints) > max_checkpoints:
+        to_delete = checkpoints.pop(0)
+        os.remove(os.path.join(checkpoint_dir, to_delete))
+
+
+def load_checkpoint(checkpoint_dir, model, optimizer):
+    checkpoints = sorted(
+        [f for f in os.listdir(checkpoint_dir) if f.startswith("csip_best_epoch_")],
+        key=lambda x: int(x.split("_")[-1].split(".")[0])
+    )
+    if not checkpoints:
+        print("No checkpoint found to resume from.")
+        return 0, float("inf")
+
+    best_ckpt = checkpoints[-1]
+    path = os.path.join(checkpoint_dir, best_ckpt)
+    checkpoint = torch.load(path)
+    model.load_state_dict(checkpoint['model_state'])
+    optimizer.load_state_dict(checkpoint['optimizer_state'])
+    start_epoch = checkpoint['epoch']
+    best_loss = checkpoint['best_loss']
+    print(f"Resumed training from epoch {start_epoch+1} with best loss {best_loss:.4f}")
+    return start_epoch, best_loss
+
+
+def fig_to_tensor(fig):
+    """Convert a Matplotlib figure to a tensor suitable for TensorBoard."""
+    buf = io.BytesIO()
+    fig.savefig(buf, format='png')
+    buf.seek(0)
+    image = Image.open(buf).convert("RGB")
+    tensor = tv.transforms.functional.to_tensor(image)
+    buf.close()
+    plt.close(fig)
+    return tensor
+
+def save_similarity_heatmaps(logits, epoch, loss, save_dir, writer):
+    os.makedirs(os.path.join(save_dir, 'logits'), exist_ok=True)
+    os.makedirs(os.path.join(save_dir, 'probs'), exist_ok=True)
+
+    # --- Raw logits heatmap ---
+    logits_np = logits.detach().cpu().numpy()
+    fig_logits = plt.figure(figsize=(8, 6))
+    sns.heatmap(logits_np, square=True, cmap="Blues", cbar=True, annot=True, fmt=".1f")
+    plt.title(f"Raw Logits Heatmap — Epoch {epoch+1}, Loss {loss:.4f}")
+    plt.xlabel("Audio Index")
+    plt.ylabel("Image Index")
+    raw_path = os.path.join(save_dir, 'logits', f"raw_logits_epoch_{epoch+1}_loss_{loss:.4f}.png")
+    fig_logits.savefig(raw_path)
+    writer.add_image("Heatmap/RawLogits", fig_to_tensor(fig_logits), global_step=epoch+1)
+
+    # --- Softmax probs heatmap ---
+    probs_np = logits.softmax(dim=1).cpu().numpy()
+    fig_probs = plt.figure(figsize=(8, 6))
+    sns.heatmap(probs_np, square=True, cmap="Blues", cbar=True, annot=True, fmt=".1f")
+    plt.title(f"Softmax Probabilities Heatmap — Epoch {epoch+1}, Loss {loss:.4f}")
+    plt.xlabel("Audio Index")
+    plt.ylabel("Image Index")
+    prob_path = os.path.join(save_dir, "probs", f"probs_epoch_{epoch+1}_loss_{loss:.4f}.png")
+    fig_probs.savefig(prob_path)
+    writer.add_image("Heatmap/SoftmaxProbs", fig_to_tensor(fig_probs), global_step=epoch+1)
+
+
+
+def train_model(model, train_loader, test_loader, 
+                optimizer, device, log_dir, 
+                checkpoint_dir, resume=False, epochs=10):
+
+    os.makedirs(log_dir, exist_ok=True)
+    os.makedirs(checkpoint_dir, exist_ok=True)
+    csv_path = os.path.join(log_dir, "training_log.csv")
+
+    writer = SummaryWriter(log_dir=log_dir)
+
+    start_epoch = 0
+    best_loss = float("inf")
+    best_epoch = -1
+
+    if resume:
+        start_epoch, best_loss = load_checkpoint(checkpoint_dir, model, optimizer)
+
+    # If resuming, don't overwrite the CSV
+    if not (resume and os.path.exists(csv_path)):
+        with open(csv_path, mode='w', newline='') as f:
+            writer_csv = csv.writer(f)
+            writer_csv.writerow(["Epoch", "Train Loss", "Test Loss", "Best Loss", "Best Epoch"])
+
+    for epoch in trange(start_epoch, epochs, colour='yellow', dynamic_ncols=True):
+        train_losses = []
+        test_losses = []
+
+        train_loop = tqdm(train_loader, desc=f"[Train Epoch {epoch+1}]", colour='blue', dynamic_ncols=True)
+        for batch in train_loop:
+            images = batch['image_tensor'].to(device)
+            audios = batch['audio_tensor'].to(device)
+            loss, logits, probs = train_batch(model, images, audios, optimizer)
+            train_losses.append(loss)
+            train_loop.set_postfix(train_loss=loss)
+
+        test_loop = tqdm(test_loader, desc=f"[Test Epoch {epoch+1}]", colour='red', dynamic_ncols=True)
+        for batch in test_loop:
+            images = batch['image_tensor'].to(device)
+            audios = batch['audio_tensor'].to(device)
+            loss, logits, probs = evaluate_batch(model, images, audios)
+            test_losses.append(loss)
+            test_loop.set_postfix(test_loss=loss)
+
+        avg_train_loss = sum(train_losses) / len(train_losses)
+        avg_test_loss = sum(test_losses) / len(test_losses)
+
+        writer.add_scalar("Loss/Train", avg_train_loss, epoch + 1)
+        writer.add_scalar("Loss/Test", avg_test_loss, epoch + 1)
+
+        print(f"Epoch {epoch+1} | Train Loss: {avg_train_loss:.4f} | Test Loss: {avg_test_loss:.4f}")
+
+        if avg_test_loss < best_loss:
+            save_similarity_heatmaps(logits, epoch, avg_test_loss, checkpoint_dir, writer)
+            best_loss = avg_test_loss
+            best_epoch = epoch + 1
+            save_checkpoint({
+                'epoch': epoch,
+                'model_state': model.state_dict(),
+                'optimizer_state': optimizer.state_dict(),
+                'best_loss': best_loss
+            }, checkpoint_dir, epoch)
+            print(f">>> Saved new best model at epoch {epoch+1}")
+
+        # Append row to CSV
+        with open(csv_path, mode='a', newline='') as f:
+            writer_csv = csv.writer(f)
+            writer_csv.writerow([epoch + 1, avg_train_loss, avg_test_loss, best_loss, best_epoch])
+
+    writer.close()
+
+
+
+
+learning_rate = 1e-20
+epochs = 400
+optimizer = torch.optim.AdamW(csip_model.parameters(), lr=learning_rate)
+train_model(
+        model=csip_model,
+        train_loader=train_dataloader,
+        test_loader=test_dataloader,
+        optimizer=optimizer,
+        device=device,
+        log_dir="runs/csip",
+        checkpoint_dir="checkpoints/csip",
+        resume=True,
+        epochs=epochs
+    )
+
+
+# tensorboard --logdir=/home/IITB/ai-at-ieor/23m1521/ashish/MTP/Vaani/Img_Audio_Alignment/runs --port=6006
+# 127.0.0.1:40697

Vaani/Img_Audio_Alignment/_2_Train.ipynb

@@ -0,0 +1,1938 @@
+{
+ "cells": [
+  {
+   "cell_type": "code",
+   "execution_count": 1,
+   "id": "5d25d5c5",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "Using device: cpu\n"
+     ]
+    }
+   ],
+   "source": [
+    "# ==================================================================\n",
+    "#             L A T E N T   D I F F U S I O N   M O D E L\n",
+    "# ==================================================================\n",
+    "# Author    : Ashish Kumar Uchadiya\n",
+    "# Created   : May 11, 2025\n",
+    "# Description: This script implements the training of a VQ-VAE model for\n",
+    "# image reconstruction, integrated with Latent Diffusion Models (LDMs) and\n",
+    "# audio conditioning. The VQ-VAE maps images to a discrete latent space, \n",
+    "# which is then modeled by the LDM for learning a diffusion process over the \n",
+    "# compressed representation. Audio features are used as conditioning inputs \n",
+    "# to guide the generation process. The training minimizes a combination of \n",
+    "# LPIPS (Learned Perceptual Image Patch Similarity) loss for perceptual \n",
+    "# fidelity and PatchGAN loss to enforce local realism. This setup enables \n",
+    "# efficient and semantically-aware generation of high-quality images driven \n",
+    "# by audio cues.\n",
+    "# ==================================================================\n",
+    "#                         I M P O R T S\n",
+    "# ==================================================================\n",
+    "from __future__ import annotations\n",
+    "import warnings\n",
+    "warnings.filterwarnings(\"ignore\")\n",
+    "\n",
+    "\n",
+    "import os\n",
+    "import io\n",
+    "import sys\n",
+    "import math\n",
+    "import random\n",
+    "import collections\n",
+    "import collections.abc\n",
+    "import re\n",
+    "from itertools import repeat\n",
+    "from pathlib import Path\n",
+    "from typing import Optional, Tuple, Union, List, Dict\n",
+    "\n",
+    "import numpy as np\n",
+    "import pandas as pd\n",
+    "from PIL import Image\n",
+    "import seaborn as sns\n",
+    "import matplotlib.pyplot as plt\n",
+    "from tqdm.auto import trange, tqdm\n",
+    "\n",
+    "import torch\n",
+    "import torch.nn.functional as F\n",
+    "from torch import nn\n",
+    "from torch.nn.init import _calculate_fan_in_and_fan_out\n",
+    "import torch.utils.checkpoint as checkpoint\n",
+    "\n",
+    "import torchvision as tv\n",
+    "from torchvision.transforms import v2\n",
+    "from torch.utils.tensorboard import SummaryWriter\n",
+    "\n",
+    "os.environ[\"CUDA_VISIBLE_DEVICES\"] = \"1\"\n",
+    "device = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\n",
+    "print(f\"Using device: {device}\")\n",
+    "\n",
+    "import torchaudio\n",
+    "import torchaudio.transforms as T\n",
+    "from torchlibrosa.stft import Spectrogram, LogmelFilterBank\n",
+    "from torchlibrosa.augmentation import SpecAugmentation\n",
+    "\n",
+    "from transformers import AutoModel, AutoTokenizer, logging\n",
+    "from huggingface_hub.file_download import hf_hub_download\n",
+    "from huggingface_hub.file_download import hf_hub_download\n",
+    "from peft import get_peft_config, get_peft_model\n",
+    "from transformers import CLIPVisionModel, AutoProcessor"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 2,
+   "id": "a41df980",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# ==================================================================\n",
+    "#                        H T S - A T\n",
+    "# ==================================================================\n",
+    "class HTSATConfig:\n",
+    "    # Ke Chen\n",
+    "    # [email protected]\n",
+    "    # HTS-AT: A HIERARCHICAL TOKEN-SEMANTIC AUDIO TRANSFORMER FOR SOUND CLASSIFICATION AND DETECTION\n",
+    "    # The configuration for training the model\n",
+    "\n",
+    "    exp_name = \"exp_htsat_pretrain\" # the saved ckpt prefix name of the model \n",
+    "    workspace = \"/home/kechen/Research/HTSAT\" # the folder of your code\n",
+    "    dataset_path = \"/home/Research/audioset\" # the dataset path\n",
+    "    desed_folder = \"/home/Research/DESED\" # the desed file\n",
+    "\n",
+    "    dataset_type = \"audioset\" # \"audioset\" \"esc-50\" \"scv2\"\n",
+    "    index_type = \"full_train\" # only works for audioset\n",
+    "    balanced_data = True # only works for audioset\n",
+    "\n",
+    "    loss_type = \"clip_bce\" # \n",
+    "    # AudioSet & SCV2: \"clip_bce\" |  ESC-50: \"clip_ce\" \n",
+    "\n",
+    "    # trained from a checkpoint, or evaluate a single model \n",
+    "    resume_checkpoint = None \n",
+    "    # \"/home/Research/model_backup/AudioSet/HTSAT_AudioSet_Saved_1.ckpt\"\n",
+    "    \n",
+    "    esc_fold = 0 # just for esc dataset, select the fold you need for evaluation and (+1) validation\n",
+    "\n",
+    "\n",
+    "    debug = False\n",
+    "\n",
+    "    random_seed = 970131 # 19970318 970131 12412 127777 1009 34047\n",
+    "    batch_size = 32 * 4 # batch size per GPU x GPU number , default is 32 x 4 = 128\n",
+    "    learning_rate = 1e-3 # 1e-4 also workable \n",
+    "    max_epoch = 100\n",
+    "    num_workers = 3\n",
+    "\n",
+    "    lr_scheduler_epoch = [10,20,30]\n",
+    "    lr_rate = [0.02, 0.05, 0.1]\n",
+    "\n",
+    "    # these data preparation optimizations do not bring many improvements, so deprecated\n",
+    "    enable_token_label = False # token label\n",
+    "    class_map_path = \"class_hier_map.npy\"\n",
+    "    class_filter = None \n",
+    "    retrieval_index = [15382, 9202, 130, 17618, 17157, 17516, 16356, 6165, 13992, 9238, 5550, 5733, 1914, 1600, 3450, 13735, 11108, 3762, \n",
+    "        9840, 11318, 8131, 4429, 16748, 4992, 16783, 12691, 4945, 8779, 2805, 9418, 2797, 14357, 5603, 212, 3852, 12666, 1338, 10269, 2388, 8260, 4293, 14454, 7677, 11253, 5060, 14938, 8840, 4542, 2627, 16336, 8992, 15496, 11140, 446, 6126, 10691, 8624, 10127, 9068, 16710, 10155, 14358, 7567, 5695, 2354, 8057, 17635, 133, 16183, 14535, 7248, 4560, 14429, 2463, 10773, 113, 2462, 9223, 4929, 14274, 4716, 17307, 4617, 2132, 11083, 1039, 1403, 9621, 13936, 2229, 2875, 17840, 9359, 13311, 9790, 13288, 4750, 17052, 8260, 14900]\n",
+    "    token_label_range = [0.2,0.6]\n",
+    "    enable_time_shift = False # shift time\n",
+    "    enable_label_enhance = False # enhance hierarchical label\n",
+    "    enable_repeat_mode = False # repeat the spectrogram / reshape the spectrogram\n",
+    "\n",
+    "\n",
+    "\n",
+    "    # for model's design\n",
+    "    enable_tscam = True # enbale the token-semantic layer\n",
+    "\n",
+    "    # for signal processing\n",
+    "    sample_rate = 32000 # 16000 for scv2, 32000 for audioset and esc-50\n",
+    "    clip_samples = sample_rate * 10 # audio_set 10-sec clip\n",
+    "    window_size = 1024\n",
+    "    hop_size = 320 # 160 for scv2, 320 for audioset and esc-50\n",
+    "    mel_bins = 64\n",
+    "    fmin = 50\n",
+    "    fmax = 14000\n",
+    "    shift_max = int(clip_samples * 0.5)\n",
+    "\n",
+    "    # for data collection\n",
+    "    classes_num = 527 # esc: 50 | audioset: 527 | scv2: 35\n",
+    "    patch_size = (25, 4) # deprecated\n",
+    "    crop_size = None # int(clip_samples * 0.5) deprecated\n",
+    "\n",
+    "    # for htsat hyperparamater\n",
+    "    htsat_window_size = 8\n",
+    "    htsat_spec_size =  256\n",
+    "    htsat_patch_size = 4 \n",
+    "    htsat_stride = (4, 4)\n",
+    "    htsat_num_head = [4,8,16,32]\n",
+    "    htsat_dim = 96 \n",
+    "    htsat_depth = [2,2,6,2]\n",
+    "\n",
+    "    swin_pretrain_path = None\n",
+    "    # \"/home/Research/model_backup/pretrain/swin_tiny_c24_patch4_window8_256.pth\"\n",
+    "\n",
+    "    # Some Deprecated Optimization in the model design, check the model code for details\n",
+    "    htsat_attn_heatmap = False\n",
+    "    htsat_hier_output = False \n",
+    "    htsat_use_max = False\n",
+    "\n",
+    "\n",
+    "    # for ensemble test \n",
+    "\n",
+    "    ensemble_checkpoints = []\n",
+    "    ensemble_strides = []\n",
+    "\n",
+    "\n",
+    "    # weight average folder\n",
+    "    wa_folder = \"/home/version_0/checkpoints/\"\n",
+    "    # weight average output filename\n",
+    "    wa_model_path = \"HTSAT_AudioSet_Saved_x.ckpt\"\n",
+    "\n",
+    "    esm_model_pathes = [\n",
+    "        \"/home/Research/model_backup/AudioSet/HTSAT_AudioSet_Saved_1.ckpt\",\n",
+    "        \"/home/Research/model_backup/AudioSet/HTSAT_AudioSet_Saved_2.ckpt\",\n",
+    "        \"/home/Research/model_backup/AudioSet/HTSAT_AudioSet_Saved_3.ckpt\",\n",
+    "        \"/home/Research/model_backup/AudioSet/HTSAT_AudioSet_Saved_4.ckpt\",\n",
+    "        \"/home/Research/model_backup/AudioSet/HTSAT_AudioSet_Saved_5.ckpt\",\n",
+    "        \"/home/Research/model_backup/AudioSet/HTSAT_AudioSet_Saved_6.ckpt\"\n",
+    "    ]\n",
+    "\n",
+    "    # for framewise localization\n",
+    "    heatmap_dir = \"/home/Research/heatmap_output\"\n",
+    "    test_file = \"htsat-test-ensemble\"\n",
+    "    fl_local = False # indicate if we need to use this dataset for the framewise detection\n",
+    "    fl_dataset = \"/home/Research/desed/desedim_embval.npy\"  \n",
+    "    fl_class_num = [\n",
+    "        \"Speech\", \"Frying\", \"Dishes\", \"Running_water\",\n",
+    "        \"Blender\", \"Electric_shaver_toothbrush\", \"Alarm_bell_ringing\",\n",
+    "        \"Cat\", \"Dog\", \"Vacuum_cleaner\"\n",
+    "    ]\n",
+    "\n",
+    "    # map 527 classes into 10 classes\n",
+    "    fl_audioset_mapping = [\n",
+    "        [0,1,2,3,4,5,6,7],\n",
+    "        [366, 367, 368],\n",
+    "        [364],\n",
+    "        [288, 289, 290, 291, 292, 293, 294, 295, 296, 297],\n",
+    "        [369],\n",
+    "        [382],\n",
+    "        [310, 388, 389, 390, 391, 392, 393, 394, 395, 396, 397, 398, 399, 400, 401, 402],\n",
+    "        [81, 82, 83, 84, 85],\n",
+    "        [74, 75, 76, 77, 78, 79],\n",
+    "        [377]\n",
+    "    ]\n",
+    "\n",
+    "\n",
+    "\n",
+    "def _ntuple(n):\n",
+    "    def parse(x):\n",
+    "        if isinstance(x, collections.abc.Iterable):\n",
+    "            return x\n",
+    "        return tuple(repeat(x, n))\n",
+    "    return parse\n",
+    "\n",
+    "to_1tuple = _ntuple(1)\n",
+    "to_2tuple = _ntuple(2)\n",
+    "to_3tuple = _ntuple(3)\n",
+    "to_4tuple = _ntuple(4)\n",
+    "to_ntuple = _ntuple\n",
+    "\n",
+    "def do_mixup(x, mixup_lambda):\n",
+    "    \"\"\"Mixup x of even indexes (0, 2, 4, ...) with x of odd indexes \n",
+    "    (1, 3, 5, ...).\n",
+    "    Args:\n",
+    "      x: (batch_size * 2, ...)\n",
+    "      mixup_lambda: (batch_size * 2,)\n",
+    "    Returns:\n",
+    "      out: (batch_size, ...)\n",
+    "    \"\"\"\n",
+    "    out = (x[0 :: 2].transpose(0, -1) * mixup_lambda[0 :: 2] + \\\n",
+    "        x[1 :: 2].transpose(0, -1) * mixup_lambda[1 :: 2]).transpose(0, -1)\n",
+    "    return out\n",
+    "\n",
+    "def interpolate(x, ratio):\n",
+    "    \"\"\"Interpolate data in time domain. This is used to compensate the \n",
+    "    resolution reduction in downsampling of a CNN.\n",
+    "    \n",
+    "    Args:\n",
+    "      x: (batch_size, time_steps, classes_num)\n",
+    "      ratio: int, ratio to interpolate\n",
+    "    Returns:\n",
+    "      upsampled: (batch_size, time_steps * ratio, classes_num)\n",
+    "    \"\"\"\n",
+    "    (batch_size, time_steps, classes_num) = x.shape\n",
+    "    upsampled = x[:, :, None, :].repeat(1, 1, ratio, 1)\n",
+    "    upsampled = upsampled.reshape(batch_size, time_steps * ratio, classes_num)\n",
+    "    return upsampled\n",
+    "\n",
+    "\n",
+    "def drop_path(x, drop_prob: float = 0., training: bool = False):\n",
+    "    \"\"\"Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).\n",
+    "    This is the same as the DropConnect impl I created for EfficientNet, etc networks, however,\n",
+    "    the original name is misleading as 'Drop Connect' is a different form of dropout in a separate paper...\n",
+    "    See discussion: https://github.com/tensorflow/tpu/issues/494#issuecomment-532968956 ... I've opted for\n",
+    "    changing the layer and argument names to 'drop path' rather than mix DropConnect as a layer name and use\n",
+    "    'survival rate' as the argument.\n",
+    "    \"\"\"\n",
+    "    if drop_prob == 0. or not training:\n",
+    "        return x\n",
+    "    keep_prob = 1 - drop_prob\n",
+    "    shape = (x.shape[0],) + (1,) * (x.ndim - 1)  # work with diff dim tensors, not just 2D ConvNets\n",
+    "    random_tensor = keep_prob + torch.rand(shape, dtype=x.dtype, device=x.device)\n",
+    "    random_tensor.floor_()  # binarize\n",
+    "    output = x.div(keep_prob) * random_tensor\n",
+    "    return output\n",
+    "\n",
+    "\n",
+    "class DropPath(nn.Module):\n",
+    "    \"\"\"Drop paths (Stochastic Depth) per sample  (when applied in main path of residual blocks).\n",
+    "    \"\"\"\n",
+    "    def __init__(self, drop_prob=None):\n",
+    "        super(DropPath, self).__init__()\n",
+    "        self.drop_prob = drop_prob\n",
+    "\n",
+    "    def forward(self, x):\n",
+    "        return drop_path(x, self.drop_prob, self.training)\n",
+    "\n",
+    "class PatchEmbed(nn.Module):\n",
+    "    \"\"\" 2D Image to Patch Embedding\n",
+    "    \"\"\"\n",
+    "    def __init__(self, img_size=224, patch_size=16, in_chans=3, embed_dim=768, norm_layer=None, flatten=True, patch_stride = 16):\n",
+    "        super().__init__()\n",
+    "        img_size = to_2tuple(img_size)\n",
+    "        patch_size = to_2tuple(patch_size)\n",
+    "        patch_stride = to_2tuple(patch_stride)\n",
+    "        self.img_size = img_size\n",
+    "        self.patch_size = patch_size\n",
+    "        self.patch_stride = patch_stride\n",
+    "        self.grid_size = (img_size[0] // patch_stride[0], img_size[1] // patch_stride[1])\n",
+    "        self.num_patches = self.grid_size[0] * self.grid_size[1]\n",
+    "        self.flatten = flatten\n",
+    "        self.in_chans = in_chans\n",
+    "        self.embed_dim = embed_dim\n",
+    "        \n",
+    "        padding = ((patch_size[0] - patch_stride[0]) // 2, (patch_size[1] - patch_stride[1]) // 2)\n",
+    "\n",
+    "        self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_stride, padding=padding)\n",
+    "        self.norm = norm_layer(embed_dim) if norm_layer else nn.Identity()\n",
+    "\n",
+    "    def forward(self, x):\n",
+    "        B, C, H, W = x.shape\n",
+    "        assert H == self.img_size[0] and W == self.img_size[1], \\\n",
+    "            f\"Input image size ({H}*{W}) doesn't match model ({self.img_size[0]}*{self.img_size[1]}).\"\n",
+    "        x = self.proj(x)\n",
+    "        if self.flatten:\n",
+    "            x = x.flatten(2).transpose(1, 2)  # BCHW -> BNC\n",
+    "        x = self.norm(x)\n",
+    "        return x\n",
+    "\n",
+    "class Mlp(nn.Module):\n",
+    "    \"\"\" MLP as used in Vision Transformer, MLP-Mixer and related networks\n",
+    "    \"\"\"\n",
+    "    def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):\n",
+    "        super().__init__()\n",
+    "        out_features = out_features or in_features\n",
+    "        hidden_features = hidden_features or in_features\n",
+    "        self.fc1 = nn.Linear(in_features, hidden_features)\n",
+    "        self.act = act_layer()\n",
+    "        self.fc2 = nn.Linear(hidden_features, out_features)\n",
+    "        self.drop = nn.Dropout(drop)\n",
+    "\n",
+    "    def forward(self, x):\n",
+    "        x = self.fc1(x)\n",
+    "        x = self.act(x)\n",
+    "        x = self.drop(x)\n",
+    "        x = self.fc2(x)\n",
+    "        x = self.drop(x)\n",
+    "        return x\n",
+    "\n",
+    "def _no_gradim_audiorunc_normal_(tensor, mean, std, a, b):\n",
+    "    # Cut & paste from PyTorch official master until it's in a few official releases - RW\n",
+    "    # Method based on https://people.sc.fsu.edu/~jburkardt/presentations/truncated_normal.pdf\n",
+    "    def norm_cdf(x):\n",
+    "        # Computes standard normal cumulative distribution function\n",
+    "        return (1. + math.erf(x / math.sqrt(2.))) / 2.\n",
+    "\n",
+    "    if (mean < a - 2 * std) or (mean > b + 2 * std):\n",
+    "        warnings.warn(\"mean is more than 2 std from [a, b] in nn.init.trunc_normal_. \"\n",
+    "                      \"The distribution of values may be incorrect.\",\n",
+    "                      stacklevel=2)\n",
+    "\n",
+    "    with torch.no_grad():\n",
+    "        # Values are generated by using a truncated uniform distribution and\n",
+    "        # then using the inverse CDF for the normal distribution.\n",
+    "        # Get upper and lower cdf values\n",
+    "        l = norm_cdf((a - mean) / std)\n",
+    "        u = norm_cdf((b - mean) / std)\n",
+    "\n",
+    "        # Uniformly fill tensor with values from [l, u], then translate to\n",
+    "        # [2l-1, 2u-1].\n",
+    "        tensor.uniform_(2 * l - 1, 2 * u - 1)\n",
+    "\n",
+    "        # Use inverse cdf transform for normal distribution to get truncated\n",
+    "        # standard normal\n",
+    "        tensor.erfinv_()\n",
+    "\n",
+    "        # Transform to proper mean, std\n",
+    "        tensor.mul_(std * math.sqrt(2.))\n",
+    "        tensor.add_(mean)\n",
+    "\n",
+    "        # Clamp to ensure it's in the proper range\n",
+    "        tensor.clamp_(min=a, max=b)\n",
+    "        return tensor\n",
+    "\n",
+    "\n",
+    "def trunc_normal_(tensor, mean=0., std=1., a=-2., b=2.):\n",
+    "    # type: (Tensor, float, float, float, float) -> Tensor\n",
+    "    r\"\"\"Fills the input Tensor with values drawn from a truncated\n",
+    "    normal distribution. The values are effectively drawn from the\n",
+    "    normal distribution :math:`\\mathcal{N}(\\text{mean}, \\text{std}^2)`\n",
+    "    with values outside :math:`[a, b]` redrawn until they are within\n",
+    "    the bounds. The method used for generating the random values works\n",
+    "    best when :math:`a \\leq \\text{mean} \\leq b`.\n",
+    "    Args:\n",
+    "        tensor: an n-dimensional `torch.Tensor`\n",
+    "        mean: the mean of the normal distribution\n",
+    "        std: the standard deviation of the normal distribution\n",
+    "        a: the minimum cutoff value\n",
+    "        b: the maximum cutoff value\n",
+    "    Examples:\n",
+    "        >>> w = torch.empty(3, 5)\n",
+    "        >>> nn.init.trunc_normal_(w)\n",
+    "    \"\"\"\n",
+    "    return _no_gradim_audiorunc_normal_(tensor, mean, std, a, b)\n",
+    "\n",
+    "\n",
+    "def variance_scaling_(tensor, scale=1.0, mode='fan_in', distribution='normal'):\n",
+    "    fan_in, fan_out = _calculate_fan_in_and_fan_out(tensor)\n",
+    "    if mode == 'fan_in':\n",
+    "        denom = fan_in\n",
+    "    elif mode == 'fan_out':\n",
+    "        denom = fan_out\n",
+    "    elif mode == 'fan_avg':\n",
+    "        denom = (fan_in + fan_out) / 2\n",
+    "\n",
+    "    variance = scale / denom\n",
+    "\n",
+    "    if distribution == \"truncated_normal\":\n",
+    "        # constant is stddev of standard normal truncated to (-2, 2)\n",
+    "        trunc_normal_(tensor, std=math.sqrt(variance) / .87962566103423978)\n",
+    "    elif distribution == \"normal\":\n",
+    "        tensor.normal_(std=math.sqrt(variance))\n",
+    "    elif distribution == \"uniform\":\n",
+    "        bound = math.sqrt(3 * variance)\n",
+    "        tensor.uniform_(-bound, bound)\n",
+    "    else:\n",
+    "        raise ValueError(f\"invalid distribution {distribution}\")\n",
+    "\n",
+    "\n",
+    "def lecun_normal_(tensor):\n",
+    "    variance_scaling_(tensor, mode='fan_in', distribution='truncated_normal')\n",
+    "\n",
+    "\n",
+    "# below codes are based and referred from https://github.com/microsoft/Swin-Transformer\n",
+    "# Swin Transformer for Computer Vision: https://arxiv.org/pdf/2103.14030.pdf\n",
+    "\n",
+    "def window_partition(x, window_size):\n",
+    "    \"\"\"\n",
+    "    Args:\n",
+    "        x: (B, H, W, C)\n",
+    "        window_size (int): window size\n",
+    "    Returns:\n",
+    "        windows: (num_windows*B, window_size, window_size, C)\n",
+    "    \"\"\"\n",
+    "    B, H, W, C = x.shape\n",
+    "    x = x.view(B, H // window_size, window_size, W // window_size, window_size, C)\n",
+    "    windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C)\n",
+    "    return windows\n",
+    "\n",
+    "\n",
+    "def window_reverse(windows, window_size, H, W):\n",
+    "    \"\"\"\n",
+    "    Args:\n",
+    "        windows: (num_windows*B, window_size, window_size, C)\n",
+    "        window_size (int): Window size\n",
+    "        H (int): Height of image\n",
+    "        W (int): Width of image\n",
+    "    Returns:\n",
+    "        x: (B, H, W, C)\n",
+    "    \"\"\"\n",
+    "    B = int(windows.shape[0] / (H * W / window_size / window_size))\n",
+    "    x = windows.view(B, H // window_size, W // window_size, window_size, window_size, -1)\n",
+    "    x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, -1)\n",
+    "    return x\n",
+    "\n",
+    "\n",
+    "class WindowAttention(nn.Module):\n",
+    "    r\"\"\" Window based multi-head self attention (W-MSA) module with relative position bias.\n",
+    "    It supports both of shifted and non-shifted window.\n",
+    "    Args:\n",
+    "        dim (int): Number of input channels.\n",
+    "        window_size (tuple[int]): The height and width of the window.\n",
+    "        num_heads (int): Number of attention heads.\n",
+    "        qkv_bias (bool, optional):  If True, add a learnable bias to query, key, value. Default: True\n",
+    "        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set\n",
+    "        attn_drop (float, optional): Dropout ratio of attention weight. Default: 0.0\n",
+    "        proj_drop (float, optional): Dropout ratio of output. Default: 0.0\n",
+    "    \"\"\"\n",
+    "\n",
+    "    def __init__(self, dim, window_size, num_heads, qkv_bias=True, qk_scale=None, attn_drop=0., proj_drop=0.):\n",
+    "\n",
+    "        super().__init__()\n",
+    "        self.dim = dim\n",
+    "        self.window_size = window_size  # Wh, Ww\n",
+    "        self.num_heads = num_heads\n",
+    "        head_dim = dim // num_heads\n",
+    "        self.scale = qk_scale or head_dim ** -0.5\n",
+    "\n",
+    "        # define a parameter table of relative position bias\n",
+    "        self.relative_position_bias_table = nn.Parameter(\n",
+    "            torch.zeros((2 * window_size[0] - 1) * (2 * window_size[1] - 1), num_heads))  # 2*Wh-1 * 2*Ww-1, nH\n",
+    "\n",
+    "        # get pair-wise relative position index for each token inside the window\n",
+    "        coords_h = torch.arange(self.window_size[0])\n",
+    "        coords_w = torch.arange(self.window_size[1])\n",
+    "        coords = torch.stack(torch.meshgrid([coords_h, coords_w]))  # 2, Wh, Ww\n",
+    "        coords_flatten = torch.flatten(coords, 1)  # 2, Wh*Ww\n",
+    "        relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :]  # 2, Wh*Ww, Wh*Ww\n",
+    "        relative_coords = relative_coords.permute(1, 2, 0).contiguous()  # Wh*Ww, Wh*Ww, 2\n",
+    "        relative_coords[:, :, 0] += self.window_size[0] - 1  # shift to start from 0\n",
+    "        relative_coords[:, :, 1] += self.window_size[1] - 1\n",
+    "        relative_coords[:, :, 0] *= 2 * self.window_size[1] - 1\n",
+    "        relative_position_index = relative_coords.sum(-1)  # Wh*Ww, Wh*Ww\n",
+    "        self.register_buffer(\"relative_position_index\", relative_position_index)\n",
+    "\n",
+    "        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)\n",
+    "        self.attn_drop = nn.Dropout(attn_drop)\n",
+    "        self.proj = nn.Linear(dim, dim)\n",
+    "        self.proj_drop = nn.Dropout(proj_drop)\n",
+    "\n",
+    "        trunc_normal_(self.relative_position_bias_table, std=.02)\n",
+    "        self.softmax = nn.Softmax(dim=-1)\n",
+    "\n",
+    "    def forward(self, x, mask=None):\n",
+    "        \"\"\"\n",
+    "        Args:\n",
+    "            x: input features with shape of (num_windows*B, N, C)\n",
+    "            mask: (0/-inf) mask with shape of (num_windows, Wh*Ww, Wh*Ww) or None\n",
+    "        \"\"\"\n",
+    "        B_, N, C = x.shape\n",
+    "        qkv = self.qkv(x).reshape(B_, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)\n",
+    "        q, k, v = qkv[0], qkv[1], qkv[2]  # make torchscript happy (cannot use tensor as tuple)\n",
+    "\n",
+    "        q = q * self.scale\n",
+    "        attn = (q @ k.transpose(-2, -1))\n",
+    "\n",
+    "        relative_position_bias = self.relative_position_bias_table[self.relative_position_index.view(-1)].view(\n",
+    "            self.window_size[0] * self.window_size[1], self.window_size[0] * self.window_size[1], -1)  # Wh*Ww,Wh*Ww,nH\n",
+    "        relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous()  # nH, Wh*Ww, Wh*Ww\n",
+    "        attn = attn + relative_position_bias.unsqueeze(0)\n",
+    "\n",
+    "        if mask is not None:\n",
+    "            nW = mask.shape[0]\n",
+    "            attn = attn.view(B_ // nW, nW, self.num_heads, N, N) + mask.unsqueeze(1).unsqueeze(0)\n",
+    "            attn = attn.view(-1, self.num_heads, N, N)\n",
+    "            attn = self.softmax(attn)\n",
+    "        else:\n",
+    "            attn = self.softmax(attn)\n",
+    "\n",
+    "        attn = self.attn_drop(attn)\n",
+    "\n",
+    "        x = (attn @ v).transpose(1, 2).reshape(B_, N, C)\n",
+    "        x = self.proj(x)\n",
+    "        x = self.proj_drop(x)\n",
+    "        return x, attn\n",
+    "\n",
+    "    def extra_repr(self):\n",
+    "        return f'dim={self.dim}, window_size={self.window_size}, num_heads={self.num_heads}'\n",
+    "\n",
+    "\n",
+    "# We use the model based on Swintransformer Block, therefore we can use the swin-transformer pretrained model\n",
+    "class SwinTransformerBlock(nn.Module):\n",
+    "    r\"\"\" Swin Transformer Block.\n",
+    "    Args:\n",
+    "        dim (int): Number of input channels.\n",
+    "        input_resolution (tuple[int]): Input resulotion.\n",
+    "        num_heads (int): Number of attention heads.\n",
+    "        window_size (int): Window size.\n",
+    "        shift_size (int): Shift size for SW-MSA.\n",
+    "        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.\n",
+    "        qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True\n",
+    "        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.\n",
+    "        drop (float, optional): Dropout rate. Default: 0.0\n",
+    "        attn_drop (float, optional): Attention dropout rate. Default: 0.0\n",
+    "        drop_path (float, optional): Stochastic depth rate. Default: 0.0\n",
+    "        act_layer (nn.Module, optional): Activation layer. Default: nn.GELU\n",
+    "        norm_layer (nn.Module, optional): Normalization layer.  Default: nn.LayerNorm\n",
+    "    \"\"\"\n",
+    "\n",
+    "    def __init__(self, dim, input_resolution, num_heads, window_size=7, shift_size=0,\n",
+    "                 mlp_ratio=4., qkv_bias=True, qk_scale=None, drop=0., attn_drop=0., drop_path=0.,\n",
+    "                 act_layer=nn.GELU, norm_layer=nn.LayerNorm, norm_before_mlp='ln'):\n",
+    "        super().__init__()\n",
+    "        self.dim = dim\n",
+    "        self.input_resolution = input_resolution\n",
+    "        self.num_heads = num_heads\n",
+    "        self.window_size = window_size\n",
+    "        self.shift_size = shift_size\n",
+    "        self.mlp_ratio = mlp_ratio\n",
+    "        self.norm_before_mlp = norm_before_mlp\n",
+    "        if min(self.input_resolution) <= self.window_size:\n",
+    "            # if window size is larger than input resolution, we don't partition windows\n",
+    "            self.shift_size = 0\n",
+    "            self.window_size = min(self.input_resolution)\n",
+    "        assert 0 <= self.shift_size < self.window_size, \"shift_size must in 0-window_size\"\n",
+    "\n",
+    "        self.norm1 = norm_layer(dim)\n",
+    "        self.attn = WindowAttention(\n",
+    "            dim, window_size=to_2tuple(self.window_size), num_heads=num_heads,\n",
+    "            qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop)\n",
+    "\n",
+    "        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()\n",
+    "        if self.norm_before_mlp == 'ln':\n",
+    "            self.norm2 = nn.LayerNorm(dim)\n",
+    "        elif self.norm_before_mlp == 'bn':\n",
+    "            self.norm2 = lambda x: nn.BatchNorm1d(dim)(x.transpose(1, 2)).transpose(1, 2)\n",
+    "        else:\n",
+    "            raise NotImplementedError\n",
+    "        mlp_hidden_dim = int(dim * mlp_ratio)\n",
+    "        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)\n",
+    "\n",
+    "        if self.shift_size > 0:\n",
+    "            # calculate attention mask for SW-MSA\n",
+    "            H, W = self.input_resolution\n",
+    "            img_mask = torch.zeros((1, H, W, 1))  # 1 H W 1\n",
+    "            h_slices = (slice(0, -self.window_size),\n",
+    "                        slice(-self.window_size, -self.shift_size),\n",
+    "                        slice(-self.shift_size, None))\n",
+    "            w_slices = (slice(0, -self.window_size),\n",
+    "                        slice(-self.window_size, -self.shift_size),\n",
+    "                        slice(-self.shift_size, None))\n",
+    "            cnt = 0\n",
+    "            for h in h_slices:\n",
+    "                for w in w_slices:\n",
+    "                    img_mask[:, h, w, :] = cnt\n",
+    "                    cnt += 1\n",
+    "\n",
+    "            mask_windows = window_partition(img_mask, self.window_size)  # nW, window_size, window_size, 1\n",
+    "            mask_windows = mask_windows.view(-1, self.window_size * self.window_size)\n",
+    "            attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2)\n",
+    "            attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(attn_mask == 0, float(0.0))\n",
+    "        else:\n",
+    "            attn_mask = None\n",
+    "\n",
+    "        self.register_buffer(\"attn_mask\", attn_mask)\n",
+    "\n",
+    "    def forward(self, x):\n",
+    "        # pdb.set_trace()\n",
+    "        H, W = self.input_resolution\n",
+    "        # print(\"H: \", H)\n",
+    "        # print(\"W: \", W)\n",
+    "        # pdb.set_trace()\n",
+    "        B, L, C = x.shape\n",
+    "        # assert L == H * W, \"input feature has wrong size\"\n",
+    "\n",
+    "        shortcut = x\n",
+    "        x = self.norm1(x)\n",
+    "        x = x.view(B, H, W, C)\n",
+    "\n",
+    "        # cyclic shift\n",
+    "        if self.shift_size > 0:\n",
+    "            shifted_x = torch.roll(x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2))\n",
+    "        else:\n",
+    "            shifted_x = x\n",
+    "\n",
+    "        # partition windows\n",
+    "        x_windows = window_partition(shifted_x, self.window_size)  # nW*B, window_size, window_size, C\n",
+    "        x_windows = x_windows.view(-1, self.window_size * self.window_size, C)  # nW*B, window_size*window_size, C\n",
+    "\n",
+    "        # W-MSA/SW-MSA\n",
+    "        attn_windows, attn = self.attn(x_windows, mask=self.attn_mask)  # nW*B, window_size*window_size, C\n",
+    "\n",
+    "        # merge windows\n",
+    "        attn_windows = attn_windows.view(-1, self.window_size, self.window_size, C)\n",
+    "        shifted_x = window_reverse(attn_windows, self.window_size, H, W)  # B H' W' C\n",
+    "\n",
+    "        # reverse cyclic shift\n",
+    "        if self.shift_size > 0:\n",
+    "            x = torch.roll(shifted_x, shifts=(self.shift_size, self.shift_size), dims=(1, 2))\n",
+    "        else:\n",
+    "            x = shifted_x\n",
+    "        x = x.view(B, H * W, C)\n",
+    "\n",
+    "        # FFN\n",
+    "        x = shortcut + self.drop_path(x)\n",
+    "        x = x + self.drop_path(self.mlp(self.norm2(x)))\n",
+    "\n",
+    "        return x, attn\n",
+    "\n",
+    "    def extra_repr(self):\n",
+    "        return f\"dim={self.dim}, input_resolution={self.input_resolution}, num_heads={self.num_heads}, \" \\\n",
+    "               f\"window_size={self.window_size}, shift_size={self.shift_size}, mlp_ratio={self.mlp_ratio}\"\n",
+    "\n",
+    "\n",
+    "\n",
+    "class PatchMerging(nn.Module):\n",
+    "    r\"\"\" Patch Merging Layer.\n",
+    "    Args:\n",
+    "        input_resolution (tuple[int]): Resolution of input feature.\n",
+    "        dim (int): Number of input channels.\n",
+    "        norm_layer (nn.Module, optional): Normalization layer.  Default: nn.LayerNorm\n",
+    "    \"\"\"\n",
+    "\n",
+    "    def __init__(self, input_resolution, dim, norm_layer=nn.LayerNorm):\n",
+    "        super().__init__()\n",
+    "        self.input_resolution = input_resolution\n",
+    "        self.dim = dim\n",
+    "        self.reduction = nn.Linear(4 * dim, 2 * dim, bias=False)\n",
+    "        self.norm = norm_layer(4 * dim)\n",
+    "\n",
+    "    def forward(self, x):\n",
+    "        \"\"\"\n",
+    "        x: B, H*W, C\n",
+    "        \"\"\"\n",
+    "        H, W = self.input_resolution\n",
+    "        B, L, C = x.shape\n",
+    "        assert L == H * W, \"input feature has wrong size\"\n",
+    "        assert H % 2 == 0 and W % 2 == 0, f\"x size ({H}*{W}) are not even.\"\n",
+    "\n",
+    "        x = x.view(B, H, W, C)\n",
+    "\n",
+    "        x0 = x[:, 0::2, 0::2, :]  # B H/2 W/2 C\n",
+    "        x1 = x[:, 1::2, 0::2, :]  # B H/2 W/2 C\n",
+    "        x2 = x[:, 0::2, 1::2, :]  # B H/2 W/2 C\n",
+    "        x3 = x[:, 1::2, 1::2, :]  # B H/2 W/2 C\n",
+    "        x = torch.cat([x0, x1, x2, x3], -1)  # B H/2 W/2 4*C\n",
+    "        x = x.view(B, -1, 4 * C)  # B H/2*W/2 4*C\n",
+    "\n",
+    "        x = self.norm(x)\n",
+    "        x = self.reduction(x)\n",
+    "\n",
+    "        return x\n",
+    "\n",
+    "    def extra_repr(self):\n",
+    "        return f\"input_resolution={self.input_resolution}, dim={self.dim}\"\n",
+    "\n",
+    "\n",
+    "class BasicLayer(nn.Module):\n",
+    "    \"\"\" A basic Swin Transformer layer for one stage.\n",
+    "    Args:\n",
+    "        dim (int): Number of input channels.\n",
+    "        input_resolution (tuple[int]): Input resolution.\n",
+    "        depth (int): Number of blocks.\n",
+    "        num_heads (int): Number of attention heads.\n",
+    "        window_size (int): Local window size.\n",
+    "        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.\n",
+    "        qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True\n",
+    "        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.\n",
+    "        drop (float, optional): Dropout rate. Default: 0.0\n",
+    "        attn_drop (float, optional): Attention dropout rate. Default: 0.0\n",
+    "        drop_path (float | tuple[float], optional): Stochastic depth rate. Default: 0.0\n",
+    "        norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm\n",
+    "        downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default: None\n",
+    "        use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False.\n",
+    "    \"\"\"\n",
+    "\n",
+    "    def __init__(self, dim, input_resolution, depth, num_heads, window_size,\n",
+    "                 mlp_ratio=4., qkv_bias=True, qk_scale=None, drop=0., attn_drop=0.,\n",
+    "                 drop_path=0., norm_layer=nn.LayerNorm, downsample=None, use_checkpoint=False,\n",
+    "                 norm_before_mlp='ln'):\n",
+    "\n",
+    "        super().__init__()\n",
+    "        self.dim = dim\n",
+    "        self.input_resolution = input_resolution\n",
+    "        self.depth = depth\n",
+    "        self.use_checkpoint = use_checkpoint\n",
+    "\n",
+    "        # build blocks\n",
+    "        self.blocks = nn.ModuleList([\n",
+    "            SwinTransformerBlock(dim=dim, input_resolution=input_resolution,\n",
+    "                                 num_heads=num_heads, window_size=window_size,\n",
+    "                                 shift_size=0 if (i % 2 == 0) else window_size // 2,\n",
+    "                                 mlp_ratio=mlp_ratio,\n",
+    "                                 qkv_bias=qkv_bias, qk_scale=qk_scale,\n",
+    "                                 drop=drop, attn_drop=attn_drop,\n",
+    "                                 drop_path=drop_path[i] if isinstance(drop_path, list) else drop_path,\n",
+    "                                 norm_layer=norm_layer, norm_before_mlp=norm_before_mlp)\n",
+    "            for i in range(depth)])\n",
+    "\n",
+    "        # patch merging layer\n",
+    "        if downsample is not None:\n",
+    "            self.downsample = downsample(input_resolution, dim=dim, norm_layer=norm_layer)\n",
+    "        else:\n",
+    "            self.downsample = None\n",
+    "\n",
+    "    def forward(self, x):\n",
+    "        attns = []\n",
+    "        for blk in self.blocks:\n",
+    "            if self.use_checkpoint:\n",
+    "                x = checkpoint.checkpoint(blk, x)\n",
+    "            else:\n",
+    "                x, attn = blk(x)\n",
+    "                if not self.training:\n",
+    "                    attns.append(attn.unsqueeze(0))\n",
+    "        if self.downsample is not None:\n",
+    "            x = self.downsample(x)\n",
+    "        if not self.training:\n",
+    "            attn = torch.cat(attns, dim = 0)\n",
+    "            attn = torch.mean(attn, dim = 0)\n",
+    "        return x, attn\n",
+    "\n",
+    "    def extra_repr(self):\n",
+    "        return f\"dim={self.dim}, input_resolution={self.input_resolution}, depth={self.depth}\"\n",
+    "\n",
+    "\n",
+    "# The Core of HTSAT\n",
+    "class HTSAT_Swin_Transformer(nn.Module):\n",
+    "    r\"\"\"HTSAT based on the Swin Transformer\n",
+    "    Args:\n",
+    "        spec_size (int | tuple(int)): Input Spectrogram size. Default 256\n",
+    "        patch_size (int | tuple(int)): Patch size. Default: 4\n",
+    "        path_stride (iot | tuple(int)): Patch Stride for Frequency and Time Axis. Default: 4\n",
+    "        in_chans (int): Number of input image channels. Default: 1 (mono)\n",
+    "        num_classes (int): Number of classes for classification head. Default: 527\n",
+    "        embed_dim (int): Patch embedding dimension. Default: 96\n",
+    "        depths (tuple(int)): Depth of each HTSAT-Swin Transformer layer.\n",
+    "        num_heads (tuple(int)): Number of attention heads in different layers.\n",
+    "        window_size (int): Window size. Default: 8\n",
+    "        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4\n",
+    "        qkv_bias (bool): If True, add a learnable bias to query, key, value. Default: True\n",
+    "        qk_scale (float): Override default qk scale of head_dim ** -0.5 if set. Default: None\n",
+    "        drop_rate (float): Dropout rate. Default: 0\n",
+    "        attn_drop_rate (float): Attention dropout rate. Default: 0\n",
+    "        drop_path_rate (float): Stochastic depth rate. Default: 0.1\n",
+    "        norm_layer (nn.Module): Normalization layer. Default: nn.LayerNorm.\n",
+    "        ape (bool): If True, add absolute position embedding to the patch embedding. Default: False\n",
+    "        patch_norm (bool): If True, add normalization after patch embedding. Default: True\n",
+    "        use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False\n",
+    "        config (module): The configuration Module from config.py (HTSATConfig Class)\n",
+    "    \"\"\"\n",
+    "\n",
+    "    def __init__(self, spec_size=256, patch_size=4, patch_stride=(4,4), \n",
+    "                in_chans=1, num_classes=527,\n",
+    "                 embed_dim=96, depths=[2, 2, 6, 2], num_heads=[4, 8, 16, 32],\n",
+    "                 window_size=8, mlp_ratio=4., qkv_bias=True, qk_scale=None,\n",
+    "                 drop_rate=0., attn_drop_rate=0., drop_path_rate=0.1,\n",
+    "                 norm_layer=nn.LayerNorm, \n",
+    "                 ape=False, patch_norm=True,\n",
+    "                 use_checkpoint=False, norm_before_mlp='ln', config = None, **kwargs):\n",
+    "        super(HTSAT_Swin_Transformer, self).__init__()\n",
+    "\n",
+    "        self.config = config\n",
+    "        self.spec_size = spec_size \n",
+    "        self.patch_stride = patch_stride\n",
+    "        self.patch_size = patch_size\n",
+    "        self.window_size = window_size\n",
+    "        self.embed_dim = embed_dim\n",
+    "        self.depths = depths\n",
+    "        self.ape = ape\n",
+    "        self.in_chans = in_chans\n",
+    "        self.num_classes = num_classes\n",
+    "        self.num_heads = num_heads\n",
+    "        self.num_layers = len(self.depths)\n",
+    "        self.num_features = int(self.embed_dim * 2 ** (self.num_layers - 1))\n",
+    "        \n",
+    "        self.drop_rate = drop_rate\n",
+    "        self.attn_drop_rate = attn_drop_rate\n",
+    "        self.drop_path_rate = drop_path_rate\n",
+    "\n",
+    "        self.qkv_bias = qkv_bias\n",
+    "        self.qk_scale = None\n",
+    "\n",
+    "        self.patch_norm = patch_norm\n",
+    "        self.norm_layer = norm_layer if self.patch_norm else None\n",
+    "        self.norm_before_mlp = norm_before_mlp\n",
+    "        self.mlp_ratio = mlp_ratio\n",
+    "\n",
+    "        self.use_checkpoint = use_checkpoint\n",
+    "\n",
+    "        #  process mel-spec ; used only once\n",
+    "        self.freq_ratio = self.spec_size // self.config.mel_bins\n",
+    "        window = 'hann'\n",
+    "        center = True\n",
+    "        pad_mode = 'reflect'\n",
+    "        ref = 1.0\n",
+    "        amin = 1e-10\n",
+    "        top_db = None\n",
+    "        self.interpolate_ratio = 32     # Downsampled ratio\n",
+    "        # Spectrogram extractor\n",
+    "        self.spectrogram_extractor = Spectrogram(n_fft=config.window_size, hop_length=config.hop_size, \n",
+    "            win_length=config.window_size, window=window, center=center, pad_mode=pad_mode, \n",
+    "            freeze_parameters=True)\n",
+    "        # Logmel feature extractor\n",
+    "        self.logmel_extractor = LogmelFilterBank(sr=config.sample_rate, n_fft=config.window_size, \n",
+    "            n_mels=config.mel_bins, fmin=config.fmin, fmax=config.fmax, ref=ref, amin=amin, top_db=top_db, \n",
+    "            freeze_parameters=True)\n",
+    "        # Spec augmenter\n",
+    "        self.spec_augmenter = SpecAugmentation(time_drop_width=64, time_stripes_num=2, \n",
+    "            freq_drop_width=8, freq_stripes_num=2) # 2 2\n",
+    "        self.bn0 = nn.BatchNorm2d(self.config.mel_bins)\n",
+    "\n",
+    "\n",
+    "        # split spctrogram into non-overlapping patches\n",
+    "        self.patch_embed = PatchEmbed(\n",
+    "            img_size=self.spec_size, patch_size=self.patch_size, in_chans=self.in_chans, \n",
+    "            embed_dim=self.embed_dim, norm_layer=self.norm_layer, patch_stride = patch_stride)\n",
+    "\n",
+    "        num_patches = self.patch_embed.num_patches\n",
+    "        patches_resolution = self.patch_embed.grid_size\n",
+    "        self.patches_resolution = patches_resolution\n",
+    "\n",
+    "        # absolute position embedding\n",
+    "        if self.ape:\n",
+    "            self.absolute_pos_embed = nn.Parameter(torch.zeros(1, num_patches, self.embed_dim))\n",
+    "            trunc_normal_(self.absolute_pos_embed, std=.02)\n",
+    "\n",
+    "        self.pos_drop = nn.Dropout(p=self.drop_rate)\n",
+    "\n",
+    "        # stochastic depth\n",
+    "        dpr = [x.item() for x in torch.linspace(0, self.drop_path_rate, sum(self.depths))]  # stochastic depth decay rule\n",
+    "\n",
+    "        # build layers\n",
+    "        self.layers = nn.ModuleList()\n",
+    "        for i_layer in range(self.num_layers):\n",
+    "            layer = BasicLayer(dim=int(self.embed_dim * 2 ** i_layer),\n",
+    "                input_resolution=(patches_resolution[0] // (2 ** i_layer),\n",
+    "                                    patches_resolution[1] // (2 ** i_layer)),\n",
+    "                depth=self.depths[i_layer],\n",
+    "                num_heads=self.num_heads[i_layer],\n",
+    "                window_size=self.window_size,\n",
+    "                mlp_ratio=self.mlp_ratio,\n",
+    "                qkv_bias=self.qkv_bias, qk_scale=self.qk_scale,\n",
+    "                drop=self.drop_rate, attn_drop=self.attn_drop_rate,\n",
+    "                drop_path=dpr[sum(self.depths[:i_layer]):sum(self.depths[:i_layer + 1])],\n",
+    "                norm_layer=self.norm_layer,\n",
+    "                downsample=PatchMerging if (i_layer < self.num_layers - 1) else None,\n",
+    "                use_checkpoint=use_checkpoint,\n",
+    "                norm_before_mlp=self.norm_before_mlp)\n",
+    "            self.layers.append(layer)\n",
+    "\n",
+    "        self.norm = self.norm_layer(self.num_features)\n",
+    "        self.avgpool = nn.AdaptiveAvgPool1d(1)\n",
+    "        self.maxpool = nn.AdaptiveMaxPool1d(1)\n",
+    "\n",
+    "        if self.config.enable_tscam:\n",
+    "            SF = self.spec_size // (2 ** (len(self.depths) - 1)) // self.patch_stride[0] // self.freq_ratio\n",
+    "            self.tscam_conv = nn.Conv2d(\n",
+    "                in_channels = self.num_features,\n",
+    "                out_channels = self.num_classes,\n",
+    "                kernel_size = (SF,3),\n",
+    "                padding = (0,1)\n",
+    "            )\n",
+    "            self.head = nn.Linear(num_classes, num_classes)\n",
+    "        else:\n",
+    "            self.head = nn.Linear(self.num_features, num_classes) if num_classes > 0 else nn.Identity()\n",
+    "\n",
+    "        self.apply(self._init_weights)\n",
+    "\n",
+    "    def _init_weights(self, m):\n",
+    "        if isinstance(m, nn.Linear):\n",
+    "            trunc_normal_(m.weight, std=.02)\n",
+    "            if isinstance(m, nn.Linear) and m.bias is not None:\n",
+    "                nn.init.constant_(m.bias, 0)\n",
+    "        elif isinstance(m, nn.LayerNorm):\n",
+    "            nn.init.constant_(m.bias, 0)\n",
+    "            nn.init.constant_(m.weight, 1.0)\n",
+    "\n",
+    "    @torch.jit.ignore\n",
+    "    def no_weight_decay(self):\n",
+    "        return {'absolute_pos_embed'}\n",
+    "\n",
+    "    @torch.jit.ignore\n",
+    "    def no_weight_decay_keywords(self):\n",
+    "        return {'relative_position_bias_table'}\n",
+    "\n",
+    "    def forward_features(self, x):\n",
+    "        frames_num = x.shape[2]        \n",
+    "        x = self.patch_embed(x)\n",
+    "        if self.ape:\n",
+    "            x = x + self.absolute_pos_embed\n",
+    "        x = self.pos_drop(x)\n",
+    "        for i, layer in enumerate(self.layers):\n",
+    "            x, attn = layer(x)\n",
+    "\n",
+    "        if self.config.enable_tscam:\n",
+    "            # for x\n",
+    "            x = self.norm(x)\n",
+    "            B, N, C = x.shape\n",
+    "            SF = frames_num // (2 ** (len(self.depths) - 1)) // self.patch_stride[0]\n",
+    "            ST = frames_num // (2 ** (len(self.depths) - 1)) // self.patch_stride[1]\n",
+    "            x = x.permute(0,2,1).contiguous().reshape(B, C, SF, ST)\n",
+    "            B, C, F, T = x.shape\n",
+    "            # group 2D CNN\n",
+    "            c_freq_bin = F // self.freq_ratio\n",
+    "            x = x.reshape(B, C, F // c_freq_bin, c_freq_bin, T)\n",
+    "            x = x.permute(0,1,3,2,4).contiguous().reshape(B, C, c_freq_bin, -1)\n",
+    "\n",
+    "            # get latent_output\n",
+    "            latent_output = self.avgpool(torch.flatten(x,2))\n",
+    "            latent_output = torch.flatten(latent_output, 1)\n",
+    "\n",
+    "            # display the attention map, if needed\n",
+    "            if self.config.htsat_attn_heatmap:\n",
+    "                # for attn\n",
+    "                attn = torch.mean(attn, dim = 1)\n",
+    "                attn = torch.mean(attn, dim = 1)\n",
+    "                attn = attn.reshape(B, SF, ST)\n",
+    "                c_freq_bin = SF // self.freq_ratio\n",
+    "                attn = attn.reshape(B, SF // c_freq_bin, c_freq_bin, ST) \n",
+    "                attn = attn.permute(0,2,1,3).contiguous().reshape(B, c_freq_bin, -1)\n",
+    "                attn = attn.mean(dim = 1)\n",
+    "                attn_max = torch.max(attn, dim = 1, keepdim = True)[0]\n",
+    "                attn_min = torch.min(attn, dim = 1, keepdim = True)[0]\n",
+    "                attn = ((attn * 0.15) + (attn_max * 0.85 - attn_min)) / (attn_max - attn_min)\n",
+    "                attn = attn.unsqueeze(dim = 2)\n",
+    "\n",
+    "            x = self.tscam_conv(x)\n",
+    "            x = torch.flatten(x, 2) # B, C, T\n",
+    "\n",
+    "            if self.config.htsat_attn_heatmap:\n",
+    "                fpx = interpolate(torch.sigmoid(x).permute(0,2,1).contiguous() * attn, 8 * self.patch_stride[1]) \n",
+    "            else: \n",
+    "                fpx = interpolate(torch.sigmoid(x).permute(0,2,1).contiguous(), 8 * self.patch_stride[1]) \n",
+    "            \n",
+    "            x = self.avgpool(x)\n",
+    "            x = torch.flatten(x, 1)\n",
+    "\n",
+    "            if self.config.loss_type == \"clip_ce\":\n",
+    "                output_dict = {\n",
+    "                    'framewise_output': fpx, # already sigmoided\n",
+    "                    'clipwise_output': x,\n",
+    "                    'latent_output': latent_output\n",
+    "                }\n",
+    "            else:\n",
+    "                output_dict = {\n",
+    "                    'framewise_output': fpx, # already sigmoided\n",
+    "                    'clipwise_output': torch.sigmoid(x),\n",
+    "                    'latent_output': latent_output\n",
+    "                }\n",
+    "           \n",
+    "        else:\n",
+    "            x = self.norm(x)  # B N C\n",
+    "            B, N, C = x.shape\n",
+    "            \n",
+    "            fpx = x.permute(0,2,1).contiguous().reshape(B, C, frames_num // (2 ** (len(self.depths) + 1)), frames_num // (2 ** (len(self.depths) + 1)) )\n",
+    "            B, C, F, T = fpx.shape\n",
+    "            c_freq_bin = F // self.freq_ratio\n",
+    "            fpx = fpx.reshape(B, C, F // c_freq_bin, c_freq_bin, T)\n",
+    "            fpx = fpx.permute(0,1,3,2,4).contiguous().reshape(B, C, c_freq_bin, -1)\n",
+    "            fpx = torch.sum(fpx, dim = 2)\n",
+    "            fpx = interpolate(fpx.permute(0,2,1).contiguous(), 8 * self.patch_stride[1]) \n",
+    "            x = self.avgpool(x.transpose(1, 2))  # B C 1\n",
+    "            x = torch.flatten(x, 1)\n",
+    "            if self.num_classes > 0:\n",
+    "                x = self.head(x)\n",
+    "                fpx = self.head(fpx)\n",
+    "            output_dict = {'framewise_output': torch.sigmoid(fpx), \n",
+    "                'clipwise_output': torch.sigmoid(x)}\n",
+    "        return output_dict\n",
+    "\n",
+    "    def crop_wav(self, x, crop_size, spe_pos = None):\n",
+    "        time_steps = x.shape[2]\n",
+    "        tx = torch.zeros(x.shape[0], x.shape[1], crop_size, x.shape[3]).to(x.device)\n",
+    "        for i in range(len(x)):\n",
+    "            if spe_pos is None:\n",
+    "                crop_pos = random.randint(0, time_steps - crop_size - 1)\n",
+    "            else:\n",
+    "                crop_pos = spe_pos\n",
+    "            tx[i][0] = x[i, 0, crop_pos:crop_pos + crop_size,:]\n",
+    "        return tx\n",
+    "\n",
+    "    # Reshape the wavform to a img size, if you want to use the pretrained swin transformer model\n",
+    "    def reshape_wav2img(self, x):\n",
+    "        B, C, T, F = x.shape\n",
+    "        target_T = int(self.spec_size * self.freq_ratio)\n",
+    "        target_F = self.spec_size // self.freq_ratio\n",
+    "        assert T <= target_T and F <= target_F, \"the wav size should less than or equal to the swin input size\"\n",
+    "        # to avoid bicubic zero error\n",
+    "        if T < target_T:\n",
+    "            x = nn.functional.interpolate(x, (target_T, x.shape[3]), mode=\"bicubic\", align_corners=True)\n",
+    "        if F < target_F:\n",
+    "            x = nn.functional.interpolate(x, (x.shape[2], target_F), mode=\"bicubic\", align_corners=True)\n",
+    "        x = x.permute(0,1,3,2).contiguous()\n",
+    "        x = x.reshape(x.shape[0], x.shape[1], x.shape[2], self.freq_ratio, x.shape[3] // self.freq_ratio)\n",
+    "        # print(x.shape)\n",
+    "        x = x.permute(0,1,3,2,4).contiguous()\n",
+    "        x = x.reshape(x.shape[0], x.shape[1], x.shape[2] * x.shape[3], x.shape[4])\n",
+    "        return x\n",
+    "    \n",
+    "    # Repeat the wavform to a img size, if you want to use the pretrained swin transformer model\n",
+    "    def repeat_wat2img(self, x, cur_pos):\n",
+    "        B, C, T, F = x.shape\n",
+    "        target_T = int(self.spec_size * self.freq_ratio)\n",
+    "        target_F = self.spec_size // self.freq_ratio\n",
+    "        assert T <= target_T and F <= target_F, \"the wav size should less than or equal to the swin input size\"\n",
+    "        # to avoid bicubic zero error\n",
+    "        if T < target_T:\n",
+    "            x = nn.functional.interpolate(x, (target_T, x.shape[3]), mode=\"bicubic\", align_corners=True)\n",
+    "        if F < target_F:\n",
+    "            x = nn.functional.interpolate(x, (x.shape[2], target_F), mode=\"bicubic\", align_corners=True)  \n",
+    "        x = x.permute(0,1,3,2).contiguous() # B C F T\n",
+    "        x = x[:,:,:,cur_pos:cur_pos + self.spec_size]\n",
+    "        x = x.repeat(repeats = (1,1,4,1))\n",
+    "        return x\n",
+    "\n",
+    "    def forward(self, x: torch.Tensor, mixup_lambda = None, infer_mode = False):# out_feat_keys: List[str] = None):\n",
+    "        x = self.spectrogram_extractor(x)   # (batch_size, 1, time_steps, freq_bins)\n",
+    "        x = self.logmel_extractor(x)    # (batch_size, 1, time_steps, mel_bins)\n",
+    "        \n",
+    "        \n",
+    "        x = x.transpose(1, 3)\n",
+    "        x = self.bn0(x)\n",
+    "        x = x.transpose(1, 3)\n",
+    "        if self.training:\n",
+    "            x = self.spec_augmenter(x)\n",
+    "        if self.training and mixup_lambda is not None:\n",
+    "            x = do_mixup(x, mixup_lambda)\n",
+    "        \n",
+    "        if infer_mode:\n",
+    "            # in infer mode. we need to handle different length audio input\n",
+    "            frame_num = x.shape[2]\n",
+    "            target_T = int(self.spec_size * self.freq_ratio)\n",
+    "            repeat_ratio = math.floor(target_T / frame_num)\n",
+    "            x = x.repeat(repeats=(1,1,repeat_ratio,1))\n",
+    "            x = self.reshape_wav2img(x)\n",
+    "            output_dict = self.forward_features(x)\n",
+    "        elif self.config.enable_repeat_mode:\n",
+    "            if self.training:\n",
+    "                cur_pos = random.randint(0, (self.freq_ratio - 1) * self.spec_size - 1)\n",
+    "                x = self.repeat_wat2img(x, cur_pos)\n",
+    "                output_dict = self.forward_features(x)\n",
+    "            else:\n",
+    "                output_dicts = []\n",
+    "                for cur_pos in range(0, (self.freq_ratio - 1) * self.spec_size + 1, self.spec_size):\n",
+    "                    tx = x.clone()\n",
+    "                    tx = self.repeat_wat2img(tx, cur_pos)\n",
+    "                    output_dicts.append(self.forward_features(tx))\n",
+    "                clipwise_output = torch.zeros_like(output_dicts[0][\"clipwise_output\"]).float().to(x.device)\n",
+    "                framewise_output = torch.zeros_like(output_dicts[0][\"framewise_output\"]).float().to(x.device)\n",
+    "                for d in output_dicts:\n",
+    "                    clipwise_output += d[\"clipwise_output\"]\n",
+    "                    framewise_output += d[\"framewise_output\"]\n",
+    "                clipwise_output  = clipwise_output / len(output_dicts)\n",
+    "                framewise_output = framewise_output / len(output_dicts)\n",
+    "\n",
+    "                output_dict = {\n",
+    "                    'framewise_output': framewise_output, \n",
+    "                    'clipwise_output': clipwise_output\n",
+    "                }\n",
+    "        else:\n",
+    "            if x.shape[2] > self.freq_ratio * self.spec_size:\n",
+    "                if self.training:\n",
+    "                    x = self.crop_wav(x, crop_size=self.freq_ratio * self.spec_size)\n",
+    "                    x = self.reshape_wav2img(x)\n",
+    "                    output_dict = self.forward_features(x)\n",
+    "                else:\n",
+    "                    # Change: Hard code here\n",
+    "                    overlap_size = 344 #(x.shape[2] - 1) // 4\n",
+    "                    output_dicts = []\n",
+    "                    crop_size = 689 #(x.shape[2] - 1) // 2\n",
+    "                    for cur_pos in range(0, x.shape[2] - crop_size - 1, overlap_size):\n",
+    "                        tx = self.crop_wav(x, crop_size = crop_size, spe_pos = cur_pos)\n",
+    "                        tx = self.reshape_wav2img(tx)\n",
+    "                        output_dicts.append(self.forward_features(tx))\n",
+    "                    clipwise_output = torch.zeros_like(output_dicts[0][\"clipwise_output\"]).float().to(x.device)\n",
+    "                    framewise_output = torch.zeros_like(output_dicts[0][\"framewise_output\"]).float().to(x.device)\n",
+    "                    latent_output = torch.zeros_like(output_dicts[0][\"latent_output\"]).float().to(x.device)\n",
+    "                    for d in output_dicts:\n",
+    "                        clipwise_output += d[\"clipwise_output\"]\n",
+    "                        framewise_output += d[\"framewise_output\"]\n",
+    "                        latent_output += d[\"latent_output\"]\n",
+    "                    clipwise_output  = clipwise_output / len(output_dicts)\n",
+    "                    framewise_output = framewise_output / len(output_dicts)\n",
+    "                    latent_output = latent_output / len(output_dicts)\n",
+    "                    output_dict = {\n",
+    "                        'framewise_output': framewise_output, \n",
+    "                        'clipwise_output': clipwise_output,\n",
+    "                        'latent_output': latent_output,\n",
+    "                    }\n",
+    "            else: # this part is typically used, and most easy one\n",
+    "                x = self.reshape_wav2img(x)\n",
+    "                output_dict = self.forward_features(x)\n",
+    "        # x = self.head(x)\n",
+    "        return output_dict\n",
+    "\n",
+    "class HTSATWrapper(nn.Module):\n",
+    "    def __init__(self, sample_rate, window_size, hop_size, mel_bins, fmin, \n",
+    "        fmax, classes_num, out_emb):\n",
+    "        super().__init__()\n",
+    "\n",
+    "        # print(\"parameters are being overidden when using HTSAT\")\n",
+    "        # print(\"HTSAT only support loading a pretrained model on AudioSet\")\n",
+    "        # @TODO later look at what parameters are same and can be merged\n",
+    "\n",
+    "        self.htsat = HTSAT_Swin_Transformer(config=HTSATConfig())\n",
+    "\n",
+    "    def forward(self, x):\n",
+    "        out_dict = self.htsat(x)\n",
+    "        out_dict['embedding'] = out_dict['latent_output']\n",
+    "        return out_dict\n",
+    "\n",
+    "\n",
+    "def get_audio_encoder(name: str):\n",
+    "    if name == \"HTSAT\":\n",
+    "        return HTSATWrapper\n",
+    "    else:\n",
+    "        raise Exception('The audio encoder name {} is incorrect or not supported'.format(name))\n",
+    "\n",
+    "class Projection(nn.Module):\n",
+    "    def __init__(self, dim_imgn: int, d_out: int, p: float=0.5) -> None:\n",
+    "        super().__init__()\n",
+    "        self.linear1 = nn.Linear(dim_imgn, d_out, bias=False)\n",
+    "        self.linear2 = nn.Linear(d_out, d_out, bias=False)\n",
+    "        self.layer_norm = nn.LayerNorm(d_out)\n",
+    "        self.drop = nn.Dropout(p)\n",
+    "\n",
+    "    def forward(self, x: torch.Tensor) -> torch.Tensor:\n",
+    "        embed1 = self.linear1(x)\n",
+    "        embed2 = self.drop(self.linear2(F.gelu(embed1)))\n",
+    "        embeds = self.layer_norm(embed1 + embed2)\n",
+    "        return embeds\n",
+    "\n",
+    "class AudioEncoder(nn.Module):\n",
+    "    def __init__(self, audioenc_name:str, dim_imgn: int, d_out: int, sample_rate: int, window_size: int,\n",
+    "            hop_size: int, mel_bins: int, fmin: int, fmax: int, classes_num: int) -> None:\n",
+    "        super().__init__()\n",
+    "\n",
+    "        audio_encoder = get_audio_encoder(audioenc_name)\n",
+    "\n",
+    "        self.base = audio_encoder(\n",
+    "            sample_rate, window_size,\n",
+    "            hop_size, mel_bins, fmin, fmax,\n",
+    "            classes_num, dim_imgn)\n",
+    "\n",
+    "        self.projection = Projection(dim_imgn, d_out)\n",
+    "\n",
+    "    def forward(self, x):\n",
+    "        out_dict = self.base(x)\n",
+    "        audio_features, audio_classification_output = out_dict['embedding'], out_dict['clipwise_output']\n",
+    "        projected_vec = self.projection(audio_features)\n",
+    "        return projected_vec, audio_classification_output\n",
+    "\n",
+    "class TextEncoder(nn.Module):\n",
+    "    def __init__(self, d_out: int, text_model: str, transformer_embed_dim: int) -> None:\n",
+    "        super().__init__()\n",
+    "        self.text_model = text_model\n",
+    "        self.base = AutoModel.from_pretrained(text_model)\n",
+    "\n",
+    "        if 'clip' in text_model:\n",
+    "            self.clip_text_projection = self.base.text_projection\n",
+    "            self.base = self.base.text_model\n",
+    "            if 'base' in text_model:\n",
+    "                transformer_embed_dim = 512\n",
+    "        \n",
+    "        self.projection = Projection(transformer_embed_dim, d_out)\n",
+    "\n",
+    "    def forward(self, x):\n",
+    "        if 'clip' in self.text_model:\n",
+    "            pooled_output = self.base(**x)[1] # get pooled output\n",
+    "            out = self.clip_text_projection(pooled_output)  # get CLS token output\n",
+    "        elif 'gpt' in self.text_model:\n",
+    "            batch_size = x['input_ids'].shape[0]\n",
+    "            hidden_states = self.base(**x)[0] # (batch_size=4, seq_len, 768)\n",
+    "\n",
+    "            sequence_lengths = torch.ne(x['input_ids'], 0).sum(-1) - 1 # tensor([13, 14, 18, 17])\n",
+    "            out = hidden_states[torch.arange(batch_size, device=hidden_states.device), sequence_lengths] # [batch_size, 768] = [4, 768]\n",
+    "        else:\n",
+    "            out = self.base(**x)[0]\n",
+    "            out = out[:, 0, :]  # get CLS token output\n",
+    "        \n",
+    "        projected_vec = self.projection(out)\n",
+    "\n",
+    "        return projected_vec\n",
+    "\n",
+    "class CLAP(nn.Module):\n",
+    "    def __init__(self,\n",
+    "                # audio\n",
+    "                audioenc_name: str,\n",
+    "                sample_rate: int, \n",
+    "                window_size: int, \n",
+    "                hop_size: int, \n",
+    "                mel_bins: int, \n",
+    "                fmin: int, \n",
+    "                fmax: int, \n",
+    "                classes_num: int, \n",
+    "                out_emb: int,\n",
+    "                # text\n",
+    "                text_model: str,\n",
+    "                transformer_embed_dim: int,\n",
+    "                # common\n",
+    "                d_proj: int,\n",
+    "                ):\n",
+    "        super().__init__()\n",
+    "\n",
+    "        \n",
+    "        self.audio_encoder = AudioEncoder(\n",
+    "            audioenc_name, out_emb, d_proj,\n",
+    "            sample_rate, window_size, hop_size, mel_bins, fmin, fmax, classes_num)\n",
+    "\n",
+    "        self.caption_encoder = TextEncoder(\n",
+    "            d_proj, text_model, transformer_embed_dim\n",
+    "        )\n",
+    "\n",
+    "        self.logit_scale = nn.Parameter(torch.ones([]) * np.log(1 / 0.07))\n",
+    "\n",
+    "    def forward(self, audio, text):\n",
+    "        audio_embed, _ = self.audio_encoder(audio)\n",
+    "        caption_embed = self.caption_encoder(text)\n",
+    "\n",
+    "        return caption_embed, audio_embed, self.logit_scale.exp()\n",
+    "    \n",
+    "    \n",
+    "    \n",
+    "# ==================================================================\n",
+    "#             A U D I O - P R E - P R O C E S S I N G\n",
+    "# ==================================================================\n",
+    "def read_audio(audio_path, resample=True):\n",
+    "    r\"\"\"Loads audio file or array and returns a torch tensor\"\"\"\n",
+    "    # Randomly sample a segment of audio_duration from the clip or pad to match duration\n",
+    "    audio_time_series, sample_rate = torchaudio.load(audio_path)\n",
+    "\n",
+    "    resample_rate = clapConfig.sample_rate\n",
+    "    if resample and resample_rate != sample_rate:\n",
+    "        resampler = T.Resample(sample_rate, resample_rate)\n",
+    "        audio_time_series = resampler(audio_time_series)\n",
+    "    return audio_time_series, resample_rate\n",
+    "\n",
+    "def load_audio_into_tensor(audio_path, audio_duration, resample=False):\n",
+    "    r\"\"\"Loads audio file and returns raw audio.\"\"\"\n",
+    "    # Randomly sample a segment of audio_duration from the clip or pad to match duration\n",
+    "    audio_time_series, sample_rate = read_audio(audio_path, resample)\n",
+    "    audio_time_series = audio_time_series.reshape(-1)\n",
+    "\n",
+    "    # audio_time_series is shorter than predefined audio duration,\n",
+    "    # so audio_time_series is extended\n",
+    "    if audio_duration*sample_rate >= audio_time_series.shape[0]:\n",
+    "        repeat_factor = int(np.ceil((audio_duration*sample_rate) /\n",
+    "                                    audio_time_series.shape[0]))\n",
+    "        # Repeat audio_time_series by repeat_factor to match audio_duration\n",
+    "        audio_time_series = audio_time_series.repeat(repeat_factor)\n",
+    "        # remove excess part of audio_time_series\n",
+    "        audio_time_series = audio_time_series[0:audio_duration*sample_rate]\n",
+    "    else:\n",
+    "        # audio_time_series is longer than predefined audio duration,\n",
+    "        # so audio_time_series is trimmed\n",
+    "        start_index = random.randrange(\n",
+    "            audio_time_series.shape[0] - audio_duration*sample_rate)\n",
+    "        audio_time_series = audio_time_series[start_index:start_index +\n",
+    "                                                audio_duration*sample_rate]\n",
+    "    return torch.FloatTensor(audio_time_series)\n",
+    "\n",
+    "np_str_obj_array_pattern = re.compile(r'[SaUO]')\n",
+    "default_collate_err_msg_format = (\n",
+    "    \"default_collate: batch must contain tensors, numpy arrays, numbers, \"\n",
+    "    \"dicts or lists; found {}\")\n",
+    "\n",
+    "def default_collate(batch):\n",
+    "    r\"\"\"Puts each data field into a tensor with outer dimension batch size\"\"\"\n",
+    "    elem = batch[0]\n",
+    "    elem_type = type(elem)\n",
+    "    if isinstance(elem, torch.Tensor):\n",
+    "        out = None\n",
+    "        if torch.utils.data.get_worker_info() is not None:\n",
+    "            # If we're in a background process, concatenate directly into a\n",
+    "            # shared memory tensor to avoid an extra copy\n",
+    "            numel = sum([x.numel() for x in batch])\n",
+    "            storage = elem.storage()._new_shared(numel)\n",
+    "            out = elem.new(storage)\n",
+    "        return torch.stack(batch, 0, out=out)\n",
+    "    elif elem_type.__module__ == 'numpy' and elem_type.__name__ != 'str_' \\\n",
+    "            and elem_type.__name__ != 'string_':\n",
+    "        if elem_type.__name__ == 'ndarray' or elem_type.__name__ == 'memmap':\n",
+    "            # array of string classes and object\n",
+    "            if np_str_obj_array_pattern.search(elem.dtype.str) is not None:\n",
+    "                raise TypeError(\n",
+    "                    default_collate_err_msg_format.format(elem.dtype))\n",
+    "\n",
+    "            return default_collate([torch.as_tensor(b) for b in batch])\n",
+    "        elif elem.shape == ():  # scalars\n",
+    "            return torch.as_tensor(batch)\n",
+    "    elif isinstance(elem, float):\n",
+    "        return torch.tensor(batch, dtype=torch.float64)\n",
+    "    elif isinstance(elem, int):\n",
+    "        return torch.tensor(batch)\n",
+    "    elif isinstance(elem, str):\n",
+    "        return batch\n",
+    "    elif isinstance(elem, collections.abc.Mapping):\n",
+    "        return {key: default_collate([d[key] for d in batch]) for key in elem}\n",
+    "    elif isinstance(elem, tuple) and hasattr(elem, '_fields'):  # namedtuple\n",
+    "        return elem_type(*(default_collate(samples) for samples in zip(*batch)))\n",
+    "    elif isinstance(elem, collections.abc.Sequence):\n",
+    "        # check to make sure that the elements in batch have consistent size\n",
+    "        it = iter(batch)\n",
+    "        elem_size = len(next(it))\n",
+    "        if not all(len(elem) == elem_size for elem in it):\n",
+    "            raise RuntimeError(\n",
+    "                'each element in list of batch should be of equal size')\n",
+    "        transposed = zip(*batch)\n",
+    "        return [default_collate(samples) for samples in transposed]\n",
+    "\n",
+    "    raise TypeError(default_collate_err_msg_format.format(elem_type))\n",
+    "\n",
+    "def preprocess_audio(audio_files, resample):\n",
+    "    r\"\"\"Load list of audio files and return raw audio\"\"\"\n",
+    "    audio_tensors = []\n",
+    "    for audio_file in audio_files:\n",
+    "        audio_tensor = load_audio_into_tensor(\n",
+    "            audio_file, clapConfig.duration, resample)\n",
+    "        audio_tensor = audio_tensor.reshape(1, -1)\n",
+    "        audio_tensors.append(audio_tensor)\n",
+    "    return default_collate(audio_tensors)\n",
+    "\n",
+    "\n",
+    "\n",
+    "# ==================================================================\n",
+    "#            A U D I O - E M B E D D I N G S - H E L P E R\n",
+    "# ==================================================================\n",
+    "def CLAPAudioProcessor(audio_files: List[str], resample=True):\n",
+    "    preprocessed_audio = preprocess_audio(audio_files, resample)\n",
+    "    preprocessed_audio = preprocessed_audio.reshape(\n",
+    "        preprocessed_audio.shape[0], preprocessed_audio.shape[2])\n",
+    "    preprocessed_audio = preprocessed_audio\n",
+    "    return preprocessed_audio\n",
+    "\n",
+    "def get_audio_embeddings(audio_files: List[str], audio_encoder, resample=True):\n",
+    "    \"\"\"Load list of audio files and return audio embeddings\"\"\"\n",
+    "    # preprocessed_audio = preprocess_audio(audio_files, resample)\n",
+    "    # with torch.no_grad():\n",
+    "    #     preprocessed_audio = preprocessed_audio.reshape(\n",
+    "    #         preprocessed_audio.shape[0], preprocessed_audio.shape[2])\n",
+    "    with torch.no_grad():\n",
+    "        preprocessed_audio = CLAPAudioProcessor(audio_files, resample)\n",
+    "        return audio_encoder(preprocessed_audio)[0]"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 3,
+   "id": "cab13923",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "trainable params: 1,289,494 || all params: 34,355,927 || trainable%: 3.7533\n"
+     ]
+    }
+   ],
+   "source": [
+    "# ==================================================================\n",
+    "#                            C L A P\n",
+    "# ==================================================================\n",
+    "class ClapConfig:\n",
+    "    # TEXT ENCODER CONFIG\n",
+    "    text_model = 'gpt2'\n",
+    "    text_len = 77\n",
+    "    transformer_embed_dim = 768\n",
+    "    freeze_text_encoder_weights = True\n",
+    "\n",
+    "    # AUDIO ENCODER CONFIG\n",
+    "    audioenc_name = 'HTSAT'\n",
+    "    out_emb = 768\n",
+    "    sample_rate = 44100\n",
+    "    duration = 7\n",
+    "    fmin = 50\n",
+    "    fmax = 8000  # 14000\n",
+    "    n_fft = 1024  # 1028\n",
+    "    hop_size = 320\n",
+    "    mel_bins = 64\n",
+    "    window_size = 1024\n",
+    "\n",
+    "    # PROJECTION SPACE CONFIG \n",
+    "    d_proj = 1024\n",
+    "    temperature = 0.003\n",
+    "\n",
+    "    # TRAINING AND EVALUATION CONFIG\n",
+    "    num_classes = 527\n",
+    "    batch_size = 1024\n",
+    "    demo = False\n",
+    "    \n",
+    "\n",
+    "clapConfig = ClapConfig()\n",
+    "clap = CLAP(\n",
+    "            audioenc_name=clapConfig.audioenc_name,\n",
+    "            sample_rate=clapConfig.sample_rate,\n",
+    "            window_size=clapConfig.window_size,\n",
+    "            hop_size=clapConfig.hop_size,\n",
+    "            mel_bins=clapConfig.mel_bins,\n",
+    "            fmin=clapConfig.fmin,\n",
+    "            fmax=clapConfig.fmax,\n",
+    "            classes_num=clapConfig.num_classes,\n",
+    "            out_emb=clapConfig.out_emb,\n",
+    "            text_model=clapConfig.text_model,\n",
+    "            transformer_embed_dim=clapConfig.transformer_embed_dim,\n",
+    "            d_proj=clapConfig.d_proj\n",
+    "        )\n",
+    "\n",
+    "model_repo = \"microsoft/msclap\"\n",
+    "model_name = {\n",
+    "    '2022': 'CLAP_weights_2022.pth',\n",
+    "    '2023': 'CLAP_weights_2023.pth',\n",
+    "    'clapcap': 'clapcap_weights_2023.pth'\n",
+    "}\n",
+    "\n",
+    "version = '2023'\n",
+    "model_fp = hf_hub_download(model_repo, model_name[version])\n",
+    "\n",
+    "model_state_dict = torch.load(model_fp, map_location=torch.device('cpu'))['model']\n",
+    "clap.load_state_dict(model_state_dict, strict=False)\n",
+    "clap.eval()\n",
+    "\n",
+    "clap_audio_encoder = clap.audio_encoder.eval().to(device)\n",
+    "\n",
+    "\n",
+    "# ENGLISH_AUDIO_DIR = r\"/home/IITB/ai-at-ieor/23m1521/datasets/Vaani/Audios/English\"\n",
+    "# audio_files = [os.path.join(ENGLISH_AUDIO_DIR, i) for i in os.listdir(ENGLISH_AUDIO_DIR) if i.endswith(\".wav\")]\n",
+    "# audio_embedding = get_audio_embeddings(audio_files, clap_audio_encoder)\n",
+    "# print(\"CLAP Audio Encoder Embeddings:\", audio_embedding.shape) # [5, 1024]\n",
+    "\n",
+    "\n",
+    "# ==================================================================\n",
+    "#                 C L A P - L o R A -  M O D E L\n",
+    "# ==================================================================\n",
+    "LoRAconfig = {\n",
+    "    \"peft_type\": \"LORA\",\n",
+    "    \"task_type\": \"FEATURE_EXTRACTION\",\n",
+    "    \"inference_mode\": False,\n",
+    "    \"r\": 16,\n",
+    "    \"target_modules\": [\"qkv\", \"fc1\", \"fc2\", \"proj\", \"linear1\", \"linear2\"],\n",
+    "    \"lora_alpha\": 32,\n",
+    "    \"lora_dropout\": 0.05,\n",
+    "    \"fan_in_fan_out\": False,\n",
+    "    \"bias\": \"all\",\n",
+    "}\n",
+    "peft_config = get_peft_config(LoRAconfig)\n",
+    "\n",
+    "peft_model = get_peft_model(clap_audio_encoder, peft_config)\n",
+    "\n",
+    "peft_model.print_trainable_parameters()\n",
+    "\n",
+    "peft_clap_audio_encoder = peft_model.base_model\n",
+    "# audio_embedding = get_audio_embeddings(audio_files, peft_clap_audio_encoder)\n",
+    "# print(\"CLAP LoRA Audio Encoder Embeddings:\", audio_embedding.shape) # [5, 1024]"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 4,
+   "id": "f9998f39",
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "'1.26.0'"
+      ]
+     },
+     "execution_count": 4,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "np.__version__"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 5,
+   "id": "16c61e94",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# ==================================================================\n",
+    "#                      C L I P - M O D E L\n",
+    "# ==================================================================\n",
+    "from transformers import CLIPImageProcessorFast, CLIPImageProcessor\n",
+    "clip_vision_model = CLIPVisionModel.from_pretrained(\"openai/clip-vit-base-patch32\").to(device)\n",
+    "# clip_vision_processor = AutoProcessor.from_pretrained(\"openai/clip-vit-base-patch32\")\n",
+    "clip_vision_processor = CLIPImageProcessorFast.from_pretrained(\"openai/clip-vit-base-patch32\")\n",
+    "# clip_vision_processor = CLIPImageProcessor.from_pretrained(\"openai/clip-vit-base-patch32\")\n",
+    "\n",
+    "# image = Image.open(\"/home/IITB/ai-at-ieor/23m1521/ashish/MTP/Vaani/Img_Audio_Alignment/000000039769.jpg\")\n",
+    "# inputs = clip_vision_processor(images=image, return_tensors=\"pt\")\n",
+    "# print(\"CLIP input image:\", inputs['pixel_values'].shape)\n",
+    "# # input_data = {'pixel_values': inputs}\n",
+    "\n",
+    "# IMAGE_SIZE = 224\n",
+    "# dummy_input = torch.randn(1, 3, IMAGE_SIZE, IMAGE_SIZE) # [1, 3, 224, 224]\n",
+    "# # dummy_input_data = {'pixel_values': dummy_input}\n",
+    "\n",
+    "# output = clip_vision_model(inputs['pixel_values'])\n",
+    "# print(\"CLIP Image Encoder Embeddings:\", output.last_hidden_state.shape) # [1, 50, 768]\n",
+    "# print(\"CLIP Image Encoder Pooled Output:\", output.pooler_output.shape) # [1, 768]"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 6,
+   "id": "61ef98b9",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# ==================================================================\n",
+    "#                       C  S  I  P - M O D U L E\n",
+    "# ==================================================================\n",
+    "class CSIP(nn.Module):\n",
+    "    def __init__(self, image_encoder, audio_encoder, dim_img=None, dim_audio=1024, dim_emb=768):\n",
+    "        super(CSIP, self).__init__()\n",
+    "        self.image_encoder = image_encoder     # CLIPVisionModel\n",
+    "        self.audio_encoder = audio_encoder     # CLAP_audio_encoder\n",
+    "\n",
+    "        # self.image_proj = nn.Linear(dim_img, dim_emb)\n",
+    "        self.audio_proj = nn.Linear(dim_audio, dim_emb)\n",
+    "\n",
+    "        # Learnable temperature parameter\n",
+    "        self.log_temp = nn.Parameter(torch.tensor(0.07).log())\n",
+    "\n",
+    "    def forward(self, images, audios):\n",
+    "        # Step 1: Feature extraction\n",
+    "        with torch.no_grad():\n",
+    "            with torch.inference_mode():\n",
+    "                image_features = self.image_encoder(images).pooler_output     # shape: [n, dim_img]\n",
+    "        audio_features = self.audio_encoder(audios)[0]                          # shape: [n, dim_audio]\n",
+    "\n",
+    "        # Step 2: Project and normalize\n",
+    "        image_embeds  = F.normalize(image_features)                           # [n, dim_emb]\n",
+    "        audio_embeds  = F.normalize(self.audio_proj(audio_features), dim=1)    # [n, dim_emb]\n",
+    "\n",
+    "        # Step 3: Cosine similarity with temperature\n",
+    "        logits = torch.matmul(image_embeds, audio_embeds.T) * self.log_temp.exp()  # [n, n]\n",
+    "        probs = logits.softmax(dim=1)\n",
+    "\n",
+    "        # Step 4: Symmetric cross-entropy loss\n",
+    "        labels = torch.arange(len(images), device=images.device)\n",
+    "        loss_i = F.cross_entropy(logits, labels)\n",
+    "        loss_t = F.cross_entropy(logits.T, labels)\n",
+    "        loss = (loss_i + loss_t) / 2\n",
+    "        return loss, logits, probs"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 7,
+   "id": "b1a15b19",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "Train Dataset: 15632\n",
+      "Test Dataset: 6700\n"
+     ]
+    }
+   ],
+   "source": [
+    "# ==================================================================\n",
+    "#                 I M A G E - A U D I O - D A T A S E T\n",
+    "# ==================================================================\n",
+    "class VaaniImageAudioDataset(torch.utils.data.Dataset):\n",
+    "    def __init__(self, df):\n",
+    "        self.image_paths = df.image_path.tolist()\n",
+    "        self.audio_paths = df.audio_path.tolist()\n",
+    "\n",
+    "    def __len__(self):\n",
+    "        return len(self.audio_paths)\n",
+    "\n",
+    "    def __getitem__(self, idx):\n",
+    "        return {\n",
+    "            'image_path': self.image_paths[idx], \n",
+    "            'audio_path': self.audio_paths[idx]\n",
+    "        }\n",
+    "\n",
+    "\n",
+    "def collate_fn(batch):\n",
+    "    image_tensor = clip_vision_processor([Image.open(item['image_path']) for item in batch])['pixel_values']\n",
+    "    audio_tensor = CLAPAudioProcessor([item['audio_path'] for item in batch], resample=True)\n",
+    "    return {'image_tensor': torch.stack(image_tensor), 'audio_tensor': audio_tensor}\n",
+    "\n",
+    "\n",
+    "# preprocessed_audio = CLAPAudioProcessor(audio_files, resample=True)\n",
+    "# clip_vision_processor = clip_vision_processor\n",
+    "\n",
+    "train_df = pd.read_csv(\"/home/IITB/ai-at-ieor/23m1521/ashish/MTP/Vaani/Img_Audio_Alignment/available_img_audios_TRAIN.csv\")\n",
+    "test_df = pd.read_csv(\"/home/IITB/ai-at-ieor/23m1521/ashish/MTP/Vaani/Img_Audio_Alignment/available_img_audios_TEST.csv\")\n",
+    "train_dataset = VaaniImageAudioDataset(train_df)\n",
+    "test_dataset = VaaniImageAudioDataset(test_df)\n",
+    "BATCH_SIZE = 32\n",
+    "\n",
+    "print('Train Dataset:', len(train_dataset))\n",
+    "print('Test Dataset:', len(test_dataset))"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 8,
+   "id": "166105cd",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "Image batch: torch.Size([32, 3, 224, 224])\n",
+      "Audio batch: torch.Size([32, 308700])\n"
+     ]
+    }
+   ],
+   "source": [
+    "train_dataloader = torch.utils.data.DataLoader(\n",
+    "    train_dataset,\n",
+    "    batch_size=BATCH_SIZE, \n",
+    "    shuffle=True, \n",
+    "    num_workers=48,\n",
+    "    collate_fn=collate_fn,\n",
+    "    pin_memory=True,\n",
+    "    drop_last=True,\n",
+    "    persistent_workers=True\n",
+    ")\n",
+    "\n",
+    "test_dataloader = torch.utils.data.DataLoader(\n",
+    "    test_dataset,\n",
+    "    batch_size=BATCH_SIZE, \n",
+    "    shuffle=False, \n",
+    "    num_workers=48,\n",
+    "    collate_fn=collate_fn,\n",
+    "    pin_memory=True,\n",
+    "    drop_last=False,\n",
+    "    persistent_workers=True\n",
+    ")\n",
+    "\n",
+    "batch = next(iter(train_dataloader))\n",
+    "print(\"Image batch:\", batch['image_tensor'].shape)\n",
+    "print(\"Audio batch:\", batch['audio_tensor'].shape)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 9,
+   "id": "d3b0a29f",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "csip_model = CSIP(clip_vision_model.eval(), peft_clap_audio_encoder).to(device)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 10,
+   "id": "2475318a",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# loss, logits, probs = csip_model(batch['image_tensor'].to(device), batch['audio_tensor'].to(device))\n",
+    "# loss, logits, probs, logits.shape, probs.shape"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 11,
+   "id": "dc748c49",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "def train_batch(model, images, audio, optimizer):\n",
+    "    model.train()\n",
+    "    optimizer.zero_grad()\n",
+    "    loss, logits, probs = model(images, audio)\n",
+    "    loss.backward()\n",
+    "    optimizer.step()\n",
+    "    return loss.item(), logits, probs\n",
+    "\n",
+    "@torch.no_grad()\n",
+    "def evaluate_batch(model, images, audio):\n",
+    "    model.eval()\n",
+    "    loss, logits, probs = model(images, audio)\n",
+    "    return loss.item(), logits, probs\n",
+    "\n",
+    "def save_checkpoint(state, checkpoint_dir, epoch, max_checkpoints=2):\n",
+    "    filename = f\"csip_best_epoch_{epoch+1}.pt\"\n",
+    "    path = os.path.join(checkpoint_dir, filename)\n",
+    "    torch.save(state, path)\n",
+    "    checkpoints = sorted(\n",
+    "        [f for f in os.listdir(checkpoint_dir) if f.startswith(\"csip_best_epoch_\")],\n",
+    "        key=lambda x: int(x.split(\"_\")[-1].split(\".\")[0])\n",
+    "    )\n",
+    "    while len(checkpoints) > max_checkpoints:\n",
+    "        to_delete = checkpoints.pop(0)\n",
+    "        os.remove(os.path.join(checkpoint_dir, to_delete))\n",
+    "\n",
+    "\n",
+    "def load_checkpoint(checkpoint_dir, model, optimizer):\n",
+    "    checkpoints = sorted(\n",
+    "        [f for f in os.listdir(checkpoint_dir) if f.startswith(\"clip_best_epoch_\")],\n",
+    "        key=lambda x: int(x.split(\"_\")[-1].split(\".\")[0])\n",
+    "    )\n",
+    "    if not checkpoints:\n",
+    "        print(\"No checkpoint found to resume from.\")\n",
+    "        return 0, float(\"inf\")\n",
+    "\n",
+    "    best_ckpt = checkpoints[-1]\n",
+    "    path = os.path.join(checkpoint_dir, best_ckpt)\n",
+    "    checkpoint = torch.load(path)\n",
+    "    model.load_state_dict(checkpoint['model_state'])\n",
+    "    optimizer.load_state_dict(checkpoint['optimizer_state'])\n",
+    "    start_epoch = checkpoint['epoch']\n",
+    "    best_loss = checkpoint['best_loss']\n",
+    "    print(f\"Resumed training from epoch {start_epoch+1} with best loss {best_loss:.4f}\")\n",
+    "    return start_epoch, best_loss\n",
+    "\n",
+    "\n",
+    "def fig_to_tensor(fig):\n",
+    "    \"\"\"Convert a Matplotlib figure to a tensor suitable for TensorBoard.\"\"\"\n",
+    "    buf = io.BytesIO()\n",
+    "    fig.savefig(buf, format='png')\n",
+    "    buf.seek(0)\n",
+    "    image = Image.open(buf).convert(\"RGB\")\n",
+    "    tensor = tv.transforms.functional.to_tensor(image)\n",
+    "    buf.close()\n",
+    "    plt.close(fig)\n",
+    "    return tensor\n",
+    "\n",
+    "def save_similarity_heatmaps(logits, epoch, loss, save_dir, writer):\n",
+    "    os.makedirs(save_dir, exist_ok=True)\n",
+    "\n",
+    "    # --- Raw logits heatmap ---\n",
+    "    logits_np = logits.detach().cpu().numpy()\n",
+    "    fig_logits = plt.figure(figsize=(8, 6))\n",
+    "    sns.heatmap(logits_np, square=True, cmap=\"viridis\", cbar=True)\n",
+    "    plt.title(f\"Raw Logits Heatmap — Epoch {epoch+1}, Loss {loss:.4f}\")\n",
+    "    plt.xlabel(\"Audio Index\")\n",
+    "    plt.ylabel(\"Image Index\")\n",
+    "    raw_path = os.path.join(save_dir, f\"raw_logits_epoch_{epoch+1}_loss_{loss:.4f}.png\")\n",
+    "    fig_logits.savefig(raw_path)\n",
+    "    writer.add_image(\"Heatmap/RawLogits\", fig_to_tensor(fig_logits), global_step=epoch+1)\n",
+    "\n",
+    "    # --- Softmax probs heatmap ---\n",
+    "    probs_np = logits.softmax(dim=1).cpu().numpy()\n",
+    "    fig_probs = plt.figure(figsize=(8, 6))\n",
+    "    sns.heatmap(probs_np, square=True, cmap=\"viridis\", cbar=True)\n",
+    "    plt.title(f\"Softmax Probabilities Heatmap — Epoch {epoch+1}, Loss {loss:.4f}\")\n",
+    "    plt.xlabel(\"Audio Index\")\n",
+    "    plt.ylabel(\"Image Index\")\n",
+    "    prob_path = os.path.join(save_dir, f\"probs_epoch_{epoch+1}_loss_{loss:.4f}.png\")\n",
+    "    fig_probs.savefig(prob_path)\n",
+    "    writer.add_image(\"Heatmap/SoftmaxProbs\", fig_to_tensor(fig_probs), global_step=epoch+1)\n",
+    "\n",
+    "\n",
+    "\n",
+    "def train_model(model, train_loader, test_loader, \n",
+    "                optimizer, device, log_dir, \n",
+    "                checkpoint_dir, resume=False, epochs=10):\n",
+    "    os.makedirs(log_dir, exist_ok=True)\n",
+    "    os.makedirs(checkpoint_dir, exist_ok=True)\n",
+    "    writer = SummaryWriter(log_dir=log_dir)\n",
+    "\n",
+    "    start_epoch = 0\n",
+    "    best_loss = float(\"inf\")\n",
+    "\n",
+    "    if resume:\n",
+    "        start_epoch, best_loss = load_checkpoint(checkpoint_dir, model, optimizer)\n",
+    "\n",
+    "    for epoch in trange(start_epoch, epochs, colour='yellow', ncols=70):\n",
+    "        train_losses = []\n",
+    "        test_losses = []\n",
+    "\n",
+    "        train_loop = tqdm(train_loader, desc=f\"[Train Epoch {epoch+1}]\", colour='blue', ncols=70)\n",
+    "        for batch in train_loop:\n",
+    "            images = batch['image_tensor'].to(device)\n",
+    "            audios = batch['audio_tensor'].to(device)\n",
+    "            loss, logits, probs = train_batch(model, images, audios, optimizer)\n",
+    "            train_losses.append(loss)\n",
+    "            train_loop.set_postfix(train_loss=loss)\n",
+    "\n",
+    "        test_loop = tqdm(test_loader, desc=f\"[Test Epoch {epoch+1}]\", colour='blue', ncols=70)\n",
+    "        for batch in test_loop:\n",
+    "            images = batch['image_tensor'].to(device)\n",
+    "            audios = batch['audio_tensor'].to(device)\n",
+    "            loss, logits, probs = evaluate_batch(model, images, audios)\n",
+    "            test_losses.append(loss)\n",
+    "            test_loop.set_postfix(test_loss=loss)\n",
+    "\n",
+    "        avg_train_loss = sum(train_losses) / len(train_losses)\n",
+    "        avg_test_loss = sum(test_losses) / len(test_losses)\n",
+    "\n",
+    "        writer.add_scalar(\"Loss/Train\", avg_train_loss, epoch + 1)\n",
+    "        writer.add_scalar(\"Loss/Test\", avg_test_loss, epoch + 1)\n",
+    "\n",
+    "        print(f\"Epoch {epoch+1} | Train Loss: {avg_train_loss:.4f} | Test Loss: {avg_test_loss:.4f}\")\n",
+    "\n",
+    "        if avg_test_loss < best_loss:\n",
+    "            save_similarity_heatmaps(logits, epoch, avg_test_loss, checkpoint_dir, writer)\n",
+    "            best_loss = avg_test_loss\n",
+    "            save_checkpoint({\n",
+    "                'epoch': epoch,\n",
+    "                'model_state': model.state_dict(),\n",
+    "                'optimizer_state': optimizer.state_dict(),\n",
+    "                'best_loss': best_loss\n",
+    "            }, checkpoint_dir, epoch)\n",
+    "            print(f\">>> Saved new best model at epoch {epoch+1}\")\n",
+    "\n",
+    "    writer.close()"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "1e29bf49",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "No checkpoint found to resume from.\n"
+     ]
+    },
+    {
+     "name": "stderr",
+     "output_type": "stream",
+     "text": [
+      "  0%|\u001b[33m                                          \u001b[0m| 0/10 [00:00<?, ?it/s]\u001b[0m"
+     ]
+    }
+   ],
+   "source": [
+    "learning_rate = 1e-4\n",
+    "epochs = 10\n",
+    "optimizer = torch.optim.AdamW(csip_model.parameters(), lr=learning_rate)\n",
+    "\n",
+    "train_model(\n",
+    "        model=csip_model,\n",
+    "        train_loader=train_dataloader,\n",
+    "        test_loader=test_dataloader,\n",
+    "        optimizer=optimizer,\n",
+    "        device=device,\n",
+    "        log_dir=\"runs/csip\",\n",
+    "        checkpoint_dir=\"checkpoints/csip\",\n",
+    "        resume=True,\n",
+    "        epochs=epochs\n",
+    "    )"
+   ]
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "Python 3",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.11.11"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 5
+}

Vaani/Img_Audio_Alignment/_2_Train.py

@@ -32,15 +32,19 @@ from pathlib import Path
 from typing import Optional, Tuple, Union, List, Dict
 
 import requests
-from PIL import Image
 import numpy as np
+import pandas as pd
+from PIL import Image
 import torch
 import torch.nn.functional as F
 from torch import nn
 from torch.nn.init import _calculate_fan_in_and_fan_out
 import torch.utils.checkpoint as checkpoint
 
-os.environ["CUDA_VISIBLE_DEVICES"] = "0"
+import torchvision as tv
+from torchvision.transforms import v2
+
+os.environ["CUDA_VISIBLE_DEVICES"] = "1"
 device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
 print(f"Using device: {device}")
 
@@ -169,7 +173,7 @@ class HTSATConfig:
     heatmap_dir = "/home/Research/heatmap_output"
     test_file = "htsat-test-ensemble"
     fl_local = False # indicate if we need to use this dataset for the framewise detection
-    fl_dataset = "/home/Research/desed/desed_eval.npy"  
+    fl_dataset = "/home/Research/desed/desedim_embval.npy"  
     fl_class_num = [
         "Speech", "Frying", "Dishes", "Running_water",
         "Blender", "Electric_shaver_toothbrush", "Alarm_bell_ringing",
@@ -314,7 +318,7 @@ class Mlp(nn.Module):
         x = self.drop(x)
         return x
 
-def _no_grad_trunc_normal_(tensor, mean, std, a, b):
+def _no_gradim_audiorunc_normal_(tensor, mean, std, a, b):
     # Cut & paste from PyTorch official master until it's in a few official releases - RW
     # Method based on https://people.sc.fsu.edu/~jburkardt/presentations/truncated_normal.pdf
     def norm_cdf(x):
@@ -368,7 +372,7 @@ def trunc_normal_(tensor, mean=0., std=1., a=-2., b=2.):
         >>> w = torch.empty(3, 5)
         >>> nn.init.trunc_normal_(w)
     """
-    return _no_grad_trunc_normal_(tensor, mean, std, a, b)
+    return _no_gradim_audiorunc_normal_(tensor, mean, std, a, b)
 
 
 def variance_scaling_(tensor, scale=1.0, mode='fan_in', distribution='normal'):
@@ -1144,9 +1148,9 @@ def get_audio_encoder(name: str):
         raise Exception('The audio encoder name {} is incorrect or not supported'.format(name))
 
 class Projection(nn.Module):
-    def __init__(self, d_in: int, d_out: int, p: float=0.5) -> None:
+    def __init__(self, dim_imgn: int, d_out: int, p: float=0.5) -> None:
         super().__init__()
-        self.linear1 = nn.Linear(d_in, d_out, bias=False)
+        self.linear1 = nn.Linear(dim_imgn, d_out, bias=False)
         self.linear2 = nn.Linear(d_out, d_out, bias=False)
         self.layer_norm = nn.LayerNorm(d_out)
         self.drop = nn.Dropout(p)
@@ -1158,7 +1162,7 @@ class Projection(nn.Module):
         return embeds
 
 class AudioEncoder(nn.Module):
-    def __init__(self, audioenc_name:str, d_in: int, d_out: int, sample_rate: int, window_size: int,
+    def __init__(self, audioenc_name:str, dim_imgn: int, d_out: int, sample_rate: int, window_size: int,
             hop_size: int, mel_bins: int, fmin: int, fmax: int, classes_num: int) -> None:
         super().__init__()
 
@@ -1167,9 +1171,9 @@ class AudioEncoder(nn.Module):
         self.base = audio_encoder(
             sample_rate, window_size,
             hop_size, mel_bins, fmin, fmax,
-            classes_num, d_in)
+            classes_num, dim_imgn)
 
-        self.projection = Projection(d_in, d_out)
+        self.projection = Projection(dim_imgn, d_out)
 
     def forward(self, x):
         out_dict = self.base(x)
@@ -1343,7 +1347,7 @@ def preprocess_audio(audio_files, resample):
     for audio_file in audio_files:
         audio_tensor = load_audio_into_tensor(
             audio_file, clapConfig.duration, resample)
-        audio_tensor = audio_tensor.reshape(1, -1).to(device)
+        audio_tensor = audio_tensor.reshape(1, -1)
         audio_tensors.append(audio_tensor)
     return default_collate(audio_tensors)
 
@@ -1352,12 +1356,21 @@ def preprocess_audio(audio_files, resample):
 # ==================================================================
 #            A U D I O - E M B E D D I N G S - H E L P E R
 # ==================================================================
+def CLAPAudioProcessor(audio_files: List[str], resample=True):
+    preprocessed_audio = preprocess_audio(audio_files, resample)
+    preprocessed_audio = preprocessed_audio.reshape(
+        preprocessed_audio.shape[0], preprocessed_audio.shape[2])
+    preprocessed_audio = preprocessed_audio
+    return preprocessed_audio
+
 def get_audio_embeddings(audio_files: List[str], audio_encoder, resample=True):
     """Load list of audio files and return audio embeddings"""
-    preprocessed_audio = preprocess_audio(audio_files, resample)
+    # preprocessed_audio = preprocess_audio(audio_files, resample)
+    # with torch.no_grad():
+    #     preprocessed_audio = preprocessed_audio.reshape(
+    #         preprocessed_audio.shape[0], preprocessed_audio.shape[2])
     with torch.no_grad():
-        preprocessed_audio = preprocessed_audio.reshape(
-            preprocessed_audio.shape[0], preprocessed_audio.shape[2])
+        preprocessed_audio = CLAPAudioProcessor(audio_files, resample)
         return audio_encoder(preprocessed_audio)[0]
 
 
@@ -1421,16 +1434,15 @@ model_fp = hf_hub_download(model_repo, model_name[version])
 
 model_state_dict = torch.load(model_fp, map_location=torch.device('cpu'))['model']
 clap.load_state_dict(model_state_dict, strict=False)
-clap.to(device)
 clap.eval()
 
-clap_audio_encoder = clap.audio_encoder.eval()
+clap_audio_encoder = clap.audio_encoder.eval().to(device)
 
 
-ENGLISH_AUDIO_DIR = r"/home/IITB/ai-at-ieor/23m1521/datasets/Vaani/Audios/English"
-audio_files = [os.path.join(ENGLISH_AUDIO_DIR, i) for i in os.listdir(ENGLISH_AUDIO_DIR) if i.endswith(".wav")]
-audio_embedding = get_audio_embeddings(audio_files, clap_audio_encoder)
-print("CLAP Audio Encoder Embeddings:", audio_embedding.shape) # [5, 1024]
+# ENGLISH_AUDIO_DIR = r"/home/IITB/ai-at-ieor/23m1521/datasets/Vaani/Audios/English"
+# audio_files = [os.path.join(ENGLISH_AUDIO_DIR, i) for i in os.listdir(ENGLISH_AUDIO_DIR) if i.endswith(".wav")]
+# audio_embedding = get_audio_embeddings(audio_files, clap_audio_encoder)
+# print("CLAP Audio Encoder Embeddings:", audio_embedding.shape) # [5, 1024]
 
 
 # ==================================================================
@@ -1449,47 +1461,302 @@ LoRAconfig = {
 }
 peft_config = get_peft_config(LoRAconfig)
 
-model = clap_audio_encoder
-peft_model = get_peft_model(model, peft_config)
+peft_model = get_peft_model(clap_audio_encoder, peft_config)
 
 peft_model.print_trainable_parameters()
 
-# peft_model.base_model
-# peft_model
-
 peft_clap_audio_encoder = peft_model.base_model
-audio_embedding = get_audio_embeddings(audio_files, peft_clap_audio_encoder)
-print("CLAP LoRA Audio Encoder Embeddings:", audio_embedding.shape) # [5, 1024]
+# audio_embedding = get_audio_embeddings(audio_files, peft_clap_audio_encoder)
+# print("CLAP LoRA Audio Encoder Embeddings:", audio_embedding.shape) # [5, 1024]
 
 
 # ==================================================================
 #                      C L I P - M O D E L
 # ==================================================================
 from transformers import CLIPImageProcessorFast, CLIPImageProcessor
-clip_vision_model = CLIPVisionModel.from_pretrained("openai/clip-vit-base-patch32")
+clip_vision_model = CLIPVisionModel.from_pretrained("openai/clip-vit-base-patch32").to(device)
 # clip_vision_processor = AutoProcessor.from_pretrained("openai/clip-vit-base-patch32")
-# clip_vision_processor = CLIPImageProcessorFast.from_pretrained("openai/clip-vit-base-patch32")
-clip_vision_processor = CLIPImageProcessor.from_pretrained("openai/clip-vit-base-patch32")
+clip_vision_processor = CLIPImageProcessorFast.from_pretrained("openai/clip-vit-base-patch32")
+# clip_vision_processor = CLIPImageProcessor.from_pretrained("openai/clip-vit-base-patch32")
 
-image = Image.open("/home/IITB/ai-at-ieor/23m1521/ashish/MTP/Vaani/Img_Audio_Alignment/000000039769.jpg")
-inputs = clip_vision_processor(images=image, return_tensors="pt")
-print("CLIP input image:", inputs['pixel_values'].shape)
-# input_data = {'pixel_values': inputs}
+# image = Image.open("/home/IITB/ai-at-ieor/23m1521/ashish/MTP/Vaani/Img_Audio_Alignment/000000039769.jpg")
+# inputs = clip_vision_processor(images=image, return_tensors="pt")
+# print("CLIP input image:", inputs['pixel_values'].shape)
+# # input_data = {'pixel_values': inputs}
 
-IMAGE_SIZE = 224
-dummy_input = torch.randn(1, 3, IMAGE_SIZE, IMAGE_SIZE) # [1, 3, 224, 224]
-# dummy_input_data = {'pixel_values': dummy_input}
+# IMAGE_SIZE = 224
+# dummy_input = torch.randn(1, 3, IMAGE_SIZE, IMAGE_SIZE) # [1, 3, 224, 224]
+# # dummy_input_data = {'pixel_values': dummy_input}
 
-# class VisionModelWrapper(nn.Module):
-#     def __init__(self, peft_model):
-#         super().__init__()
-#         self.model = peft_model.base_model
+# output = clip_vision_model(inputs['pixel_values'])
+# print("CLIP Image Encoder Embeddings:", output.last_hidden_state.shape) # [1, 50, 768]
+# print("CLIP Image Encoder Pooled Output:", output.pooler_output.shape) # [1, 768]
 
-#     def forward(self, x):
-#         return self.model(pixel_values=x).last_hidden_state
+exit()
+
+
+# ==================================================================
+#                       C  S  I  P - M O D U L E
+# ==================================================================
+class CSIP(nn.Module):
+    def __init__(self, image_encoder, audio_encoder, dim_img=None, dim_audio=1024, dim_emb=768):
+        super(CSIP, self).__init__()
+        self.image_encoder = image_encoder     # CLIPVisionModel
+        self.audio_encoder = audio_encoder     # CLAP_audio_encoder
 
-# wrapped_model = VisionModelWrapper(peft_model)
-# output = wrapped_model(input_data)
-output = clip_vision_model(inputs['pixel_values'])
-print("CLIP Image Encoder Embeddings:", output.last_hidden_state.shape) # [1, 50, 768]
-print("CLIP Image Encoder Pooled Output:", output.pooler_output.shape) # [1, 768]
+        # self.image_proj = nn.Linear(dim_img, dim_emb)
+        self.audio_proj = nn.Linear(dim_audio, dim_emb)
+
+        # Learnable temperature parameter
+        self.log_temp = nn.Parameter(torch.tensor(0.07).log())
+
+    def forward(self, images, audios):
+        # Step 1: Feature extraction
+        with torch.no_grad():
+            with torch.inference_mode():
+                image_features = self.image_encoder(images)     # shape: [n, dim_img]
+        audio_features = self.audio_encoder(audios)             # shape: [n, dim_audio]
+
+        # Step 2: Project and normalize
+        image_embeds  = F.normalize(image_features)                           # [n, dim_emb]
+        audio_embeds  = F.normalize(self.text_proj(audio_features), dim=1)    # [n, dim_emb]
+
+        # Step 3: Cosine similarity with temperature
+        logits = torch.matmul(image_embeds, audio_embeds.T) * self.log_temp.exp()  # [n, n]
+
+        # Step 4: Symmetric cross-entropy loss
+        labels = torch.arange(len(images), device=images.device)
+        loss_i = F.cross_entropy(logits, labels)
+        loss_t = F.cross_entropy(logits.T, labels)
+        loss = (loss_i + loss_t) / 2
+        return loss, logits
+
+
+
+
+# ==================================================================
+#                 I M A G E - A U D I O - D A T A S E T
+# ==================================================================
+class VaaniImageAudioDataset(torch.utils.data.Dataset):
+    def __init__(self, df):
+        self.image_paths = df.image_path.tolist()
+        self.audio_paths = df.audio_path.tolist()
+
+    def __len__(self):
+        return len(self.audio_paths)
+
+    def __getitem__(self, idx):
+        return {
+            'image_path': self.image_paths[idx], 
+            'audio_path': self.audio_paths[idx]
+        }
+
+
+def collate_fn(batch):
+    image_tensor = clip_vision_processor([Image.open(item['image_path']) for item in batch])['pixel_values']
+    audio_tensor = CLAPAudioProcessor([item['audio_path'] for item in batch], resample=True)
+    return {'image_tensor': image_tensor, 'audio_tensor': audio_tensor}
+
+
+# preprocessed_audio = CLAPAudioProcessor(audio_files, resample=True)
+# clip_vision_processor = clip_vision_processor
+
+train_df = pd.read_csv("/home/IITB/ai-at-ieor/23m1521/ashish/MTP/Vaani/Img_Audio_Alignment/available_img_audios_TRAIN.csv")
+test_df = pd.read_csv("/home/IITB/ai-at-ieor/23m1521/ashish/MTP/Vaani/Img_Audio_Alignment/available_img_audios_TEST.csv")
+train_dataset = VaaniImageAudioDataset(train_df)
+test_dataset = VaaniImageAudioDataset(test_df)
+BATCH_SIZE = 32
+
+print('Train Dataset:', len(train_dataset))
+print('Test Dataset:', len(test_dataset))
+
+train_dataloader = torch.utils.data.DataLoader(
+    train_dataset,
+    batch_size=BATCH_SIZE, 
+    shuffle=True, 
+    num_workers=48,
+    collate_fn=collate_fn,
+    pin_memory=True,
+    drop_last=True,
+    persistent_workers=True
+)
+
+test_dataloader = torch.utils.data.DataLoader(
+    test_dataset,
+    batch_size=BATCH_SIZE, 
+    shuffle=False, 
+    num_workers=48,
+    collate_fn=collate_fn,
+    pin_memory=True,
+    drop_last=False,
+    persistent_workers=True
+)
+
+batch = next(iter(train_dataloader))
+print('BATCH SHAPE:', batch['image_tensor'].shape, batch['audio_tensor'].shape)
+exit()
+
+# csip_model = CSIP(clip_vision_model.eval(), peft_clap_audio_encoder).to(device)
+# loss, logits = csip_model(images, texts)
+# print("Loss:", loss.item())
+
+
+
+
+# import torch
+# import torch.nn as nn
+# import torch.nn.functional as F
+# from torch.utils.tensorboard import SummaryWriter
+# from tqdm import tqdm
+# import os
+
+# # CLIPCore (from your previous definition)
+# class CLIPCore(nn.Module):
+#     def __init__(self, image_encoder, text_encoder, d_i, d_t, d_e):
+#         super(CLIPCore, self).__init__()
+#         self.image_encoder = image_encoder
+#         self.text_encoder = text_encoder
+#         self.image_proj = nn.Linear(d_i, d_e)
+#         self.text_proj = nn.Linear(d_t, d_e)
+#         self.log_temp = nn.Parameter(torch.tensor(0.07).log())
+
+#     def forward(self, images, texts):
+#         image_features = self.image_encoder(images)
+#         text_features = self.text_encoder(texts)
+#         image_embeds = F.normalize(self.image_proj(image_features), dim=1)
+#         text_embeds = F.normalize(self.text_proj(text_features), dim=1)
+#         logits = torch.matmul(image_embeds, text_embeds.T) * self.log_temp.exp()
+#         labels = torch.arange(len(images), device=images.device)
+#         loss_i = F.cross_entropy(logits, labels)
+#         loss_t = F.cross_entropy(logits.T, labels)
+#         loss = (loss_i + loss_t) / 2
+#         return loss, logits
+
+
+# # Dummy encoders
+# class DummyEncoder(nn.Module):
+#     def __init__(self, output_dim):
+#         super().__init__()
+#         self.fc = nn.Linear(1024, output_dim)
+#     def forward(self, x):
+#         return self.fc(x)
+
+
+# # === Helper Functions ===
+
+# def train_batch(model, images, texts, optimizer):
+#     model.train()
+#     optimizer.zero_grad()
+#     loss, _ = model(images, texts)
+#     loss.backward()
+#     optimizer.step()
+#     return loss.item()
+
+# @torch.no_grad()
+# def evaluate_batch(model, images, texts):
+#     model.eval()
+#     loss, _ = model(images, texts)
+#     return loss.item()
+
+# def save_checkpoint(state, checkpoint_dir, name="clip_best.pt"):
+#     torch.save(state, os.path.join(checkpoint_dir, name))
+
+# def load_checkpoint(checkpoint_path, model, optimizer):
+#     checkpoint = torch.load(checkpoint_path)
+#     model.load_state_dict(checkpoint['model_state'])
+#     optimizer.load_state_dict(checkpoint['optimizer_state'])
+#     start_epoch = checkpoint['epoch']
+#     best_loss = checkpoint['best_loss']
+#     print(f"Resumed training from epoch {start_epoch+1} with best loss {best_loss:.4f}")
+#     return start_epoch, best_loss
+
+
+# # === Training Loop ===
+
+# def train_model(model, train_loader, test_loader, optimizer, device, log_dir, checkpoint_dir, resume=False, epochs=10):
+#     os.makedirs(log_dir, exist_ok=True)
+#     os.makedirs(checkpoint_dir, exist_ok=True)
+#     writer = SummaryWriter(log_dir=log_dir)
+
+#     start_epoch = 0
+#     best_loss = float("inf")
+
+#     # Resume logic
+#     checkpoint_path = os.path.join(checkpoint_dir, "clip_best.pt")
+#     if resume and os.path.exists(checkpoint_path):
+#         start_epoch, best_loss = load_checkpoint(checkpoint_path, model, optimizer)
+
+#     for epoch in range(start_epoch, epochs):
+#         train_losses = []
+#         test_losses = []
+
+#         train_loop = tqdm(train_loader, desc=f"[Train Epoch {epoch+1}]")
+#         for images, texts in train_loop:
+#             images, texts = images.to(device), texts.to(device)
+#             loss = train_batch(model, images, texts, optimizer)
+#             train_losses.append(loss)
+#             train_loop.set_postfix(train_loss=loss)
+
+#         test_loop = tqdm(test_loader, desc=f"[Test Epoch {epoch+1}]")
+#         for images, texts in test_loop:
+#             images, texts = images.to(device), texts.to(device)
+#             loss = evaluate_batch(model, images, texts)
+#             test_losses.append(loss)
+#             test_loop.set_postfix(test_loss=loss)
+
+#         avg_train_loss = sum(train_losses) / len(train_losses)
+#         avg_test_loss = sum(test_losses) / len(test_losses)
+
+#         writer.add_scalar("Loss/Train", avg_train_loss, epoch + 1)
+#         writer.add_scalar("Loss/Test", avg_test_loss, epoch + 1)
+
+#         print(f"Epoch {epoch+1} | Train Loss: {avg_train_loss:.4f} | Test Loss: {avg_test_loss:.4f}")
+
+#         # Save checkpoint if test loss improves
+#         if avg_test_loss < best_loss:
+#             best_loss = avg_test_loss
+#             save_checkpoint({
+#                 'epoch': epoch,
+#                 'model_state': model.state_dict(),
+#                 'optimizer_state': optimizer.state_dict(),
+#                 'best_loss': best_loss
+#             }, checkpoint_dir)
+#             print(f">>> Saved new best model at epoch {epoch+1}")
+
+#     writer.close()
+
+
+# # Dummy DataLoader
+# def get_dummy_loader(batch_size=32, num_batches=100):
+#     for _ in range(num_batches):
+#         yield torch.randn(batch_size, 1024), torch.randn(batch_size, 1024)
+
+
+# # === Main ===
+# if __name__ == "__main__":
+#     device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+#     d_i, d_t, d_e = 512, 512, 256
+#     batch_size = 32
+#     learning_rate = 1e-4
+#     epochs = 10
+
+#     image_encoder = DummyEncoder(d_i)
+#     text_encoder = DummyEncoder(d_t)
+#     model = CLIPCore(image_encoder, text_encoder, d_i, d_t, d_e).to(device)
+
+#     optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
+
+#     train_loader = get_dummy_loader(batch_size)
+#     test_loader = get_dummy_loader(batch_size)
+
+#     train_model(
+#         model=model,
+#         train_loader=train_loader,
+#         test_loader=test_loader,
+#         optimizer=optimizer,
+#         device=device,
+#         log_dir="runs/clip_audio_image",
+#         checkpoint_dir="checkpoints/clip",
+#         resume=True,
+#         epochs=epochs
+#     )

Vaani/Img_Audio_Alignment/checkpoints/csip/csip_best_epoch_203.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:135b0d13ec7cddb29f4809cde8c3b3283c12acd952031167dd96e3a374dcfd6b
+size 509342259

Vaani/Img_Audio_Alignment/checkpoints/csip/csip_best_epoch_27.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:96492f3f12877ebf99517fd51fc483a3c102c6080cbbf474b01a31f98454dc2e
+size 509341156

Vaani/Img_Audio_Alignment/runs/csip/events.out.tfevents.1747523457.rmgpu025.3375295.0

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0427f062675aaf495fb5d3fd32554103c41a838c0d384e8416b4fc8eaef8be32
+size 792717

Vaani/Img_Audio_Alignment/runs/csip/training_log.csv

@@ -0,0 +1,367 @@
+Epoch, Train Loss, Test Loss, Best Loss, Best Epoch
+1,3.46573521709833,3.4610748007183982,3.4610748007183982,1
+2,3.4657436843778266,3.461075038001651,3.4610748007183982,1
+3,3.4657385955091384,3.4610777514321462,3.4610748007183982,1
+4,3.4657369627327217,3.4610763163793656,3.4610748007183982,1
+5,3.465738452848841,3.4610726004555112,3.4610726004555112,5
+6,3.465746810201739,3.4610738901864915,3.4610726004555112,5
+7,3.4657312768404602,3.461075624965486,3.4610726004555112,5
+8,3.4657392096324044,3.461072703770229,3.4610726004555112,5
+9,3.4657418517792813,3.4610761290504817,3.4610726004555112,5
+10,3.465733758250221,3.4610775572913033,3.4610726004555112,5
+11,3.4657396654613684,3.46107660588764,3.4610726004555112,5
+12,3.4657437903959245,3.4610769033432005,3.4610726004555112,5
+13,3.4657299968062856,3.4610789242244904,3.4610726004555112,5
+14,3.4657370570253154,3.4610737800598144,3.4610726004555112,5
+15,3.4657460182416635,3.461070890653701,3.461070890653701,15
+16,3.4657411438519836,3.461075038001651,3.461070890653701,15
+17,3.465738015096696,3.4610758043470837,3.461070890653701,15
+18,3.465739974232971,3.4610746485846384,3.461070890653701,15
+19,3.465737346254411,3.461074812071664,3.461070890653701,15
+20,3.4657459053836885,3.4610753933588665,3.461070890653701,15
+21,3.4657364262909187,3.4610756033942813,3.461070890653701,15
+22,3.4657383458536177,3.4610799448830742,3.461070890653701,15
+23,3.4657375920014304,3.4610742341904412,3.461070890653701,15
+24,3.4657370931789524,3.461076168786912,3.461070890653701,15
+25,3.465736433130796,3.4610772473471507,3.461070890653701,15
+26,3.465753781990927,3.4610763265973046,3.461070890653701,15
+27,3.4657371376381545,3.461068638165792,3.461068638165792,27
+28,3.4657417775177564,3.461077817281087,3.461068638165792,27
+29,3.4657461814215926,3.461076136997768,3.461068638165792,27
+30,3.46573303322323,3.4610767409915018,3.461068638165792,27
+31,3.465738852004536,3.4610775118782406,3.461068638165792,27
+32,3.4657424312145984,3.4610733213878815,3.461068638165792,27
+33,3.465739804701727,3.4610734939575196,3.461068638165792,27
+34,3.465737399019179,3.461074911980402,3.461068638165792,27
+35,3.4657459112464406,3.4610757736932665,3.461068638165792,27
+36,3.465745515999247,3.461076812517075,3.461068638165792,27
+37,3.4657391080113706,3.461073993501209,3.461068638165792,27
+38,3.4657376931339012,3.4610795236769176,3.461068638165792,27
+39,3.4657382471639604,3.461073828878857,3.461068638165792,27
+40,3.465732887142994,3.461078895841326,3.461068638165792,27
+41,3.4657490404902913,3.4610782282693044,3.461068638165792,27
+42,3.4657386443654046,3.461073933328901,3.461068638165792,27
+43,3.46573893896869,3.4610799698602586,3.461068638165792,27
+44,3.4657407231995316,3.461077432405381,3.461068638165792,27
+45,3.4657383854271937,3.46107694989159,3.461068638165792,27
+46,3.4657291291189973,3.4610767580214,3.461068638165792,27
+47,3.465737374102483,3.461077233723232,3.461068638165792,27
+48,3.4657390997058055,3.4610781737736294,3.461068638165792,27
+49,3.4657394270427893,3.46107615289234,3.461068638165792,27
+50,3.4657357354633143,3.4610809689476376,3.461068638165792,27
+51,3.4657413041005367,3.4610799119586035,3.461068638165792,27
+52,3.465741222999135,3.4610729069936843,3.461068638165792,27
+53,3.465745713867125,3.4610725323359173,3.461068638165792,27
+54,3.465748988702649,3.4610764003935315,3.461068638165792,27
+55,3.4657449101815456,3.461075716926938,3.461068638165792,27
+56,3.465735683675672,3.461073174930754,3.461068638165792,27
+57,3.4657397333715783,3.461076039359683,3.461068638165792,27
+58,3.4657420652811646,3.461073859532674,3.461068638165792,27
+59,3.4657352219839566,3.461078399703616,3.461068638165792,27
+60,3.465740470124073,3.461074948310852,3.461068638165792,27
+61,3.4657351355083654,3.4610742954980758,3.461068638165792,27
+62,3.4657426378766045,3.4610716240746635,3.461068638165792,27
+63,3.465735925025627,3.461075274149577,3.461068638165792,27
+64,3.465727699584648,3.4610778536115374,3.461068638165792,27
+65,3.4657338398401856,3.461080871309553,3.461068638165792,27
+66,3.4657386731906015,3.461076987357367,3.461068638165792,27
+67,3.465741856664908,3.461076105208624,3.461068638165792,27
+68,3.4657425152473764,3.461078572273254,3.461068638165792,27
+69,3.465733808572175,3.461077258700416,3.461068638165792,27
+70,3.4657452492440335,3.4610759156090873,3.461068638165792,27
+71,3.465743854397633,3.4610787323543004,3.461068638165792,27
+72,3.4657392931766196,3.46107439994812,3.461068638165792,27
+73,3.4657338339774335,3.46107584862482,3.461068638165792,27
+74,3.465745241427031,3.461076845441546,3.461068638165792,27
+75,3.4657387733459473,3.4610753434044974,3.461068638165792,27
+76,3.4657278603217643,3.461074594088963,3.461068638165792,27
+77,3.4657401466955906,3.4610771372204736,3.461068638165792,27
+78,3.4657415811155663,3.461074877920605,3.461068638165792,27
+79,3.465745365521947,3.4610784859884354,3.461068638165792,27
+80,3.4657414550663996,3.46107325894492,3.461068638165792,27
+81,3.465744232056571,3.461077724184309,3.461068638165792,27
+82,3.465740345052031,3.461073366800944,3.461068638165792,27
+83,3.4657413295057955,3.4610741683415003,3.461068638165792,27
+84,3.465736867951565,3.4610788685934883,3.461068638165792,27
+85,3.4657383502506818,3.461074101357233,3.461068638165792,27
+86,3.4657447699640618,3.4610755636578516,3.461068638165792,27
+87,3.465739328841694,3.4610791433425176,3.461068638165792,27
+88,3.4657491352714476,3.4610740661621096,3.461068638165792,27
+89,3.465734797423003,3.461073492822193,3.461068638165792,27
+90,3.465737556336356,3.461072019168309,3.461068638165792,27
+91,3.4657416778509735,3.4610769669214885,3.461068638165792,27
+92,3.465753313459334,3.4610748552140738,3.461068638165792,27
+93,3.4657432901077585,3.461076624052865,3.461068638165792,27
+94,3.4657475782222433,3.4610774267287483,3.461068638165792,27
+95,3.465736511300822,3.461075268472944,3.461068638165792,27
+96,3.46573667985494,3.4610716479165213,3.461068638165792,27
+97,3.4657364189624786,3.461074159258888,3.461068638165792,27
+98,3.465737197731362,3.461076383363633,3.461068638165792,27
+99,3.4657412313047002,3.461075404712132,3.461068638165792,27
+100,3.4657446996110384,3.4610728910991124,3.461068638165792,27
+101,3.4657335491454013,3.4610757543927146,3.461068638165792,27
+102,3.4657341920938647,3.4610740400495983,3.461068638165792,27
+103,3.4657355043731752,3.461075307074047,3.461068638165792,27
+104,3.4657418293053985,3.461076294808161,3.461068638165792,27
+105,3.4657442037199364,3.4610763561157953,3.461068638165792,27
+106,3.465734151054601,3.461075664701916,3.461068638165792,27
+107,3.4657363960000334,3.461078837939671,3.461068638165792,27
+108,3.465736574325405,3.4610768329529535,3.461068638165792,27
+109,3.465738363441874,3.461077742349534,3.461068638165792,27
+110,3.4657457519750126,3.4610780761355446,3.461068638165792,27
+111,3.4657398457409907,3.461078751654852,3.461068638165792,27
+112,3.4657469171969617,3.4610763640630813,3.461068638165792,27
+113,3.4657363588692713,3.4610767455328078,3.461068638165792,27
+114,3.4657273487966567,3.461072870663234,3.461068638165792,27
+115,3.4657473823086162,3.4610776117869784,3.461068638165792,27
+116,3.4657388109652723,3.461082070214408,3.461068638165792,27
+117,3.465734670396711,3.4610760495776223,3.461068638165792,27
+118,3.4657491802192126,3.4610756056649343,3.461068638165792,27
+119,3.465745851153233,3.4610792194093976,3.461068638165792,27
+120,3.4657440600825136,3.4610720282509213,3.461068638165792,27
+121,3.465742236766659,3.4610794657752626,3.461068638165792,27
+122,3.465734956694431,3.4610778093338013,3.461068638165792,27
+123,3.4657477277224182,3.461074746222723,3.461068638165792,27
+124,3.4657419480261256,3.46107584862482,3.461068638165792,27
+125,3.465743043383614,3.461076643353417,3.461068638165792,27
+126,3.4657389433657535,3.4610770032519387,3.461068638165792,27
+127,3.465749927231523,3.4610780579703193,3.461068638165792,27
+128,3.465737064353755,3.461076171057565,3.461068638165792,27
+129,3.465738639479778,3.461073470115662,3.461068638165792,27
+130,3.465747668606336,3.461080064092364,3.461068638165792,27
+131,3.4657384968194807,3.461079478263855,3.461068638165792,27
+132,3.465744329769103,3.4610762982141403,3.461068638165792,27
+133,3.465744828591581,3.461078797067915,3.461068638165792,27
+134,3.4657438485348813,3.461077248482477,3.461068638165792,27
+135,3.465738213941699,3.4610794907524474,3.461068638165792,27
+136,3.465734507216782,3.4610772927602134,3.461068638165792,27
+137,3.4657387093442384,3.4610828002293905,3.461068638165792,27
+138,3.4657351741048155,3.461076327732631,3.461068638165792,27
+139,3.4657402404996214,3.461069657689049,3.461068638165792,27
+140,3.46574071000834,3.461073954900106,3.461068638165792,27
+141,3.4657448583939034,3.4610777014777776,3.461068638165792,27
+142,3.4657410143828784,3.46107508455004,3.461068638165792,27
+143,3.4657372089683034,3.461074709892273,3.461068638165792,27
+144,3.465740814072187,3.4610765264147805,3.461068638165792,27
+145,3.4657404256648703,3.461073138600304,3.461068638165792,27
+146,3.4657417223101756,3.4610730829693024,3.461068638165792,27
+147,3.465751084147907,3.461075372922988,3.461068638165792,27
+148,3.4657406293955004,3.461076331138611,3.461068638165792,27
+149,3.4657458384506037,3.4610776231402443,3.461068638165792,27
+150,3.4657487248788112,3.461075837271554,3.461068638165792,27
+151,3.465731780548565,3.4610822654905773,3.461068638165792,27
+152,3.4657375988413075,3.4610815161750432,3.461068638165792,27
+153,3.4657279018495903,3.461074351129078,3.461068638165792,27
+154,3.4657473139098434,3.4610762164706275,3.461068638165792,27
+155,3.4657374415241304,3.46107755842663,3.461068638165792,27
+156,3.465745550687196,3.4610732634862265,3.461068638165792,27
+157,3.4657520876556145,3.461078104518709,3.461068638165792,27
+158,3.4657488592335435,3.461076010976519,3.461068638165792,27
+159,3.4657454524861007,3.4610791932968867,3.461068638165792,27
+160,3.4657454901054257,3.461073358853658,3.461068638165792,27
+161,3.465733320498076,3.46107577255794,3.461068638165792,27
+162,3.4657320453495277,3.4610758020764307,3.461068638165792,27
+163,3.465735073460907,3.4610742568969726,3.461068638165792,27
+164,3.4657488929443674,3.4610763243266516,3.461068638165792,27
+165,3.4657452614580997,3.4610728388740903,3.461068638165792,27
+166,3.465739756822586,3.4610755897703624,3.461068638165792,27
+167,3.465740684603081,3.461081443514143,3.461068638165792,27
+168,3.46574545786029,3.4610778956186206,3.461068638165792,27
+169,3.4657452096704575,3.461077322278704,3.461068638165792,27
+170,3.4657420359674047,3.4610769271850588,3.461068638165792,27
+171,3.465747346643542,3.4610752253305344,3.461068638165792,27
+172,3.4657437635249777,3.46107512088049,3.461068638165792,27
+173,3.4657375822301772,3.4610755273274014,3.461068638165792,27
+174,3.4657370711936326,3.461075911067781,3.461068638165792,27
+175,3.4657403939082974,3.461072755995251,3.461068638165792,27
+176,3.4657455756038917,3.4610804580506827,3.461068638165792,27
+177,3.465736765353406,3.4610756011236283,3.461068638165792,27
+178,3.4657422377437843,3.4610758338655745,3.461068638165792,27
+179,3.465742223575467,3.4610772212346395,3.461068638165792,27
+180,3.4657448574167784,3.4610780318578085,3.461068638165792,27
+181,3.465744264790269,3.461074071838742,3.461068638165792,27
+182,3.4657460778463083,3.4610787902559554,3.461068638165792,27
+183,3.4657434918841377,3.4610763617924283,3.461068638165792,27
+184,3.4657449131129217,3.4610758418128604,3.461068638165792,27
+185,3.4657399444306485,3.4610760507129488,3.461068638165792,27
+186,3.465741965125819,3.4610746156601677,3.461068638165792,27
+187,3.465730566958912,3.4610766183762323,3.461068638165792,27
+188,3.465735874703673,3.461078835669018,3.461068638165792,27
+189,3.465743572496977,3.461074141093663,3.461068638165792,27
+190,3.4657333439490836,3.461076091584705,3.461068638165792,27
+191,3.4657420017680183,3.461072128159659,3.461068638165792,27
+192,3.465742250934976,3.461073995771862,3.461068638165792,27
+193,3.465743506541018,3.461074773470561,3.461068638165792,27
+194,3.4657396053681606,3.461078999156044,3.461068638165792,27
+195,3.4657456166431553,3.461080551147461,3.461068638165792,27
+196,3.4657305117513313,3.4610739231109617,3.461068638165792,27
+197,3.465746803850424,3.461074657667251,3.461068638165792,27
+198,3.4657420735867297,3.461078974178859,3.461068638165792,27
+199,3.465739453425173,3.461075480779012,3.461068638165792,27
+200,3.4657408443630717,3.4610792137327646,3.461068638165792,27
+201,3.4657429539766467,3.461070229893639,3.461068638165792,27
+202,3.465730927029594,3.4610772870835804,3.461068638165792,27
+203,3.465742439031601,3.461066366377331,3.461066366377331,203
+204,3.4657376076354356,3.4610721542721703,3.461066366377331,203
+205,3.4657432231746736,3.461076171057565,3.461066366377331,203
+206,3.4657434997011403,3.4610708770297824,3.461066366377331,203
+207,3.465746605982546,3.461074680373782,3.461066366377331,203
+208,3.4657472215715,3.461073782330468,3.461066366377331,203
+209,3.465739699660755,3.461071228981018,3.461066366377331,203
+210,3.465738641922591,3.461079346565973,3.461066366377331,203
+211,3.4657519738205145,3.4610757759639195,3.461066366377331,203
+212,3.465741905521174,3.4610800198146277,3.461066366377331,203
+213,3.465743096636944,3.461075006212507,3.461066366377331,203
+214,3.465747090636707,3.4610739265169417,3.461066366377331,203
+215,3.465735371484131,3.46107569308508,3.461066366377331,203
+216,3.465741084247339,3.4610781680969964,3.461066366377331,203
+217,3.4657371532721597,3.4610765502566383,3.461066366377331,203
+218,3.4657441196871583,3.4610755477632793,3.461066366377331,203
+219,3.4657394568451116,3.461075703303019,3.461066366377331,203
+220,3.4657466108681727,3.461076226688567,3.461066366377331,203
+221,3.4657258948341747,3.461073694910322,3.461066366377331,203
+222,3.465742596837341,3.461073380424863,3.461066366377331,203
+223,3.4657480218371406,3.461077020281837,3.461066366377331,203
+224,3.46574360425355,3.46107980410258,3.461066366377331,203
+225,3.4657430380094247,3.461077296166193,3.461066366377331,203
+226,3.4657447709411873,3.461075229871841,3.461066366377331,203
+227,3.46574397311836,3.461075416065398,3.461066366377331,203
+228,3.465747856702961,3.4610744623910814,3.461066366377331,203
+229,3.465742313470997,3.461079010509309,3.461066366377331,203
+230,3.4657399991496662,3.461076861336118,3.461066366377331,203
+231,3.4657419118724886,3.4610717796144033,3.461066366377331,203
+232,3.465750910219599,3.4610756840024677,3.461066366377331,203
+233,3.465734547767483,3.4610780398050944,3.461066366377331,203
+234,3.4657490028709663,3.461073198772612,3.461066366377331,203
+235,3.4657469103570846,3.461076382228306,3.461066366377331,203
+236,3.465751744196063,3.4610761154265632,3.461066366377331,203
+237,3.465744873539346,3.461074137687683,3.461066366377331,203
+238,3.465736294379,3.461072364307585,3.461066366377331,203
+239,3.465735922582814,3.461075043678284,3.461066366377331,203
+240,3.465739214518031,3.4610767046610516,3.461066366377331,203
+241,3.4657417237758636,3.4610774982543218,3.461066366377331,203
+242,3.4657466211279884,3.4610769146964664,3.461066366377331,203
+243,3.4657360202953464,3.461077460788545,3.461066366377331,203
+244,3.4657325925397093,3.461074318204607,3.461066366377331,203
+245,3.46573880510252,3.46107482001895,3.461066366377331,203
+246,3.465741548381868,3.4610731794720606,3.461066366377331,203
+247,3.4657338745281345,3.4610764628364925,3.461066366377331,203
+248,3.465727367362038,3.461078473499843,3.461066366377331,203
+249,3.465733908727521,3.4610756692432223,3.461066366377331,203
+250,3.4657462312549843,3.4610773415792555,3.461066366377331,203
+251,3.465736162955644,3.4610737403233847,3.461066366377331,203
+252,3.4657466963666383,3.461074525969369,3.461066366377331,203
+253,3.465742281714424,3.461075586364383,3.461066366377331,203
+254,3.465745878512742,3.461071934018816,3.461066366377331,203
+255,3.465741225441948,3.4610752151125954,3.461066366377331,203
+256,3.4657438988568354,3.461076286860875,3.461066366377331,203
+257,3.465739455867986,3.461078098842076,3.461066366377331,203
+258,3.465738760643318,3.4610744328725906,3.461066366377331,203
+259,3.4657423857782708,3.461076878366016,3.461066366377331,203
+260,3.465746166764713,3.46107535248711,3.461066366377331,203
+261,3.465745163257005,3.4610774653298515,3.461066366377331,203
+262,3.4657451749825086,3.4610761506216865,3.461066366377331,203
+263,3.4657509776412465,3.461078325907389,3.461066366377331,203
+264,3.465736592402224,3.4610776435761226,3.461066366377331,203
+265,3.4657426901528092,3.4610785109656197,3.461066366377331,203
+266,3.4657360403264157,3.4610751708348593,3.461066366377331,203
+267,3.4657488162400294,3.4610736460912794,3.461066366377331,203
+268,3.465739801770351,3.4610775595619563,3.461066366377331,203
+269,3.4657413847133762,3.4610764934903098,3.461066366377331,203
+270,3.465745068964411,3.4610776265462238,3.461066366377331,203
+271,3.4657335256943935,3.4610766274588447,3.461066366377331,203
+272,3.46574189379567,3.46107405935015,3.461066366377331,203
+273,3.465747667140648,3.4610798540569485,3.461066366377331,203
+274,3.4657369246248337,3.461074560029166,3.461066366377331,203
+275,3.4657399385678964,3.461074930145627,3.461066366377331,203
+276,3.4657441157786573,3.461076661518642,3.461066366377331,203
+277,3.465741717424549,3.4610764764604114,3.461066366377331,203
+278,3.465741726218677,3.4610762505304247,3.461066366377331,203
+279,3.465729427142221,3.4610741024925593,3.461066366377331,203
+280,3.465743330658459,3.461078711918422,3.461066366377331,203
+281,3.4657441602378594,3.46107717809223,3.461066366377331,203
+282,3.4657420970377375,3.4610763969875515,3.461066366377331,203
+283,3.465742819133352,3.461077488036383,3.461066366377331,203
+284,3.465747262122201,3.4610763947168985,3.461066366377331,203
+285,3.4657372260679966,3.4610755057561966,3.461066366377331,203
+286,3.4657412772295904,3.4610748529434203,3.461066366377331,203
+287,3.4657379652633042,3.4610749108450753,3.461066366377331,203
+288,3.4657514979604813,3.4610778104691278,3.461066366377331,203
+289,3.4657424991248083,3.461076621782212,3.461066366377331,203
+290,3.465733950255347,3.46107615289234,3.461066366377331,203
+291,3.4657346088378156,3.461077342714582,3.461066366377331,203
+292,3.4657455931921475,3.4610693659101215,3.461066366377331,203
+293,3.4657400304176766,3.461077047529675,3.461066366377331,203
+294,3.4657420017680183,3.4610726969582695,3.461066366377331,203
+295,3.465732125962367,3.461076987357367,3.461066366377331,203
+296,3.465746620639426,3.4610745225633894,3.461066366377331,203
+297,3.4657395017928763,3.461076865877424,3.461066366377331,203
+298,3.465732188009825,3.4610753297805785,3.461066366377331,203
+299,3.4657439985236183,3.4610741831007457,3.461066366377331,203
+300,3.465742188887518,3.461076719420297,3.461066366377331,203
+301,3.465748793277584,3.4610767648333596,3.461066366377331,203
+302,3.465756238483992,3.461079292070298,3.461066366377331,203
+303,3.465737801594812,3.46107439994812,3.461066366377331,203
+304,3.4657489647630784,3.4610772541591097,3.461066366377331,203
+305,3.465739526709572,3.4610728490920293,3.461066366377331,203
+306,3.4657510636282747,3.4610797178177606,3.461066366377331,203
+307,3.465744333189042,3.461076752344767,3.461066366377331,203
+308,3.4657363842745297,3.4610820055007934,3.461066366377331,203
+309,3.4657407437191634,3.4610778819947017,3.461066366377331,203
+310,3.465737354071414,3.4610712244397117,3.461066366377331,203
+311,3.465733944392595,3.4610777854919434,3.461066366377331,203
+312,3.465738877409794,3.461076634270804,3.461066366377331,203
+313,3.4657346576940817,3.4610760473069693,3.461066366377331,203
+314,3.4657396097652247,3.461073366800944,3.461066366377331,203
+315,3.4657462034069124,3.4610737698418754,3.461066366377331,203
+316,3.4657335984902304,3.461074838184175,3.461066366377331,203
+317,3.4657493453533923,3.4610770679655527,3.461066366377331,203
+318,3.4657437620592897,3.4610746485846384,3.461066366377331,203
+319,3.4657356978439893,3.4610737278347923,3.461066366377331,203
+320,3.465740291798701,3.461077796845209,3.461066366377331,203
+321,3.465741071056147,3.4610751640229,3.461066366377331,203
+322,3.4657440131804984,3.461071945372082,3.461066366377331,203
+323,3.4657415781841903,3.461076443535941,3.461066366377331,203
+324,3.4657341906281767,3.461078177179609,3.461066366377331,203
+325,3.465742891440626,3.4610766274588447,3.461066366377331,203
+326,3.4657413475826138,3.4610791967028662,3.461066366377331,203
+327,3.4657359333311923,3.461078169232323,3.461066366377331,203
+328,3.4657344500549505,3.4610752207892284,3.461066366377331,203
+329,3.46574491653286,3.4610752616609846,3.461066366377331,203
+330,3.465749452348615,3.461077682177226,3.461066366377331,203
+331,3.4657414619062767,3.46107474395207,3.461066366377331,203
+332,3.465741016337129,3.4610730420975457,3.461066366377331,203
+333,3.4657289151285515,3.4610743670236497,3.461066366377331,203
+334,3.4657349904052546,3.4610790036973498,3.461066366377331,203
+335,3.4657439594386052,3.461073590460278,3.461066366377331,203
+336,3.465748588081266,3.4610757952644713,3.461066366377331,203
+337,3.4657394216685997,3.4610817318870906,3.461066366377331,203
+338,3.4657396126966007,3.4610796360742477,3.461066366377331,203
+339,3.465739287313868,3.4610763016201203,3.461066366377331,203
+340,3.4657487292758753,3.4610756692432223,3.461066366377331,203
+341,3.4657395208468205,3.4610739537647794,3.461066366377331,203
+342,3.4657366896261936,3.4610780239105225,3.461066366377331,203
+343,3.465738265729341,3.4610756794611612,3.461066366377331,203
+344,3.46573600808128,3.4610785302661715,3.461066366377331,203
+345,3.4657506024251217,3.4610763947168985,3.461066366377331,203
+346,3.465738793865579,3.461075832730248,3.461066366377331,203
+347,3.4657355283127456,3.461077941031683,3.461066366377331,203
+348,3.465745441737722,3.4610753638403757,3.461066366377331,203
+349,3.465739132439504,3.4610723086765836,3.461066366377331,203
+350,3.465746519018392,3.461076836358933,3.461066366377331,203
+351,3.465745436363533,3.4610771814982098,3.461066366377331,203
+352,3.465747639781139,3.461080509140378,3.461066366377331,203
+353,3.4657432231746736,3.4610794578279767,3.461066366377331,203
+354,3.4657415449619293,3.461075420606704,3.461066366377331,203
+355,3.465747449730263,3.4610786449341546,3.461066366377331,203
+356,3.465735651919099,3.4610766569773355,3.461066366377331,203
+357,3.465745558992761,3.4610742523556666,3.461066366377331,203
+358,3.4657398398782386,3.4610731238410586,3.461066366377331,203
+359,3.465734190139614,3.461073308899289,3.461066366377331,203
+360,3.465740403190988,3.4610808065959384,3.461066366377331,203
+361,3.465748734161502,3.4610764741897584,3.461066366377331,203
+362,3.4657354364629653,3.4610757623400006,3.461066366377331,203
+363,3.465733616078486,3.4610783985682896,3.461066366377331,203
+364,3.4657378011062496,3.461077472141811,3.461066366377331,203
+365,3.4657405311944056,3.461075202624003,3.461066366377331,203
+366,3.465736522537763,3.461072782107762,3.461066366377331,203

Vaani/Vaani-Images-Audio-JSON.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a10f00f0bb589b863a9f8e0eb8d8d814e91d32a057b48f2163067d44b4c83711
+size 1753651607

Vaani/VaaniLDM/ddpm_ckpt_epoch59.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ec39cb118ad8b8800ed7fef6dca84df35d367a9445a90618c1eb63e31a86b642
+size 593245482

Vaani/VaaniLDM/ddpm_ckpt_epoch60.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:28c7cfa3e05c931bbdedb69bcb7ac30326bf72c1a6364180d4902d57f69565a2
+size 593245546