Upload 4 files

LMdecoder.py

@@ -0,0 +1,169 @@
+import copy
+from doctest import ELLIPSIS_MARKER
+from functools import partial
+import json
+from turtle import forward, shape
+import einops
+import torch
+from torch import nn
+
+from mmcls.models.backbones.vision_transformer import TransformerEncoderLayer
+from transformers import GPT2Model, GPT2Config,GPT2LMHeadModel,GPTNeoForCausalLM,GPTNeoModel, \
+ BartModel, BartConfig, BartForCausalLM, BertForMaskedLM, AutoConfig, AutoModel, AutoModelForCausalLM, AutoTokenizer 
+from transformers import BitsAndBytesConfig
+
+from peft import prepare_model_for_kbit_training
+from peft import LoraConfig
+from peft import get_peft_model
+
+        
+from mmcv.cnn import build_norm_layer
+from mmcv.runner import BaseModule
+import math
+from ipdb import set_trace
+
+class mixEmbed(nn.Module):
+    def __init__(self, lm_embed: nn.Embedding , audio_embeddings, *args, **kwargs) -> None:
+        super().__init__(*args, **kwargs)
+        self.lm_embed = lm_embed
+        self.audio_embeddings = audio_embeddings # ugly but works without modifying raw model codes
+        
+    def forward(self, input_ids):
+        text_ids = torch.clamp(input_ids.clone(), 0).long()
+        
+        au_ids = torch.clamp(-(input_ids.clone() + 1), 0).long()
+        text_embeds = self.lm_embed(text_ids)
+        au_embeds = self.audio_embeddings[au_ids]
+        with torch.no_grad():
+            embed_mask = (input_ids > 0)
+        mix_embeds = au_embeds.clone()
+        mix_embeds[embed_mask] = text_embeds[embed_mask]
+        return mix_embeds
+ 
+
+class LMDecoder(nn.Module):
+    def __init__(self,
+                # num_patches=196,
+                img_size=(80,512),
+                patch_size:int=16,
+                in_chans:int=3,
+                embed_dim=1024, # encoder embed dim
+                decoder_embed_dim=512,
+                norm_cfg=dict(type='LN', eps=1e-6),
+                # patch_resolution=14,
+                decoder_type='gpt2',
+                freeze_decoder=True,
+                additional_layer:int=0,
+                ):
+        super().__init__()
+        self.decoder_type = decoder_type
+        self.load_lm()
+        
+        self.lm_embed = self.lm.get_input_embeddings()
+        try:
+            self.lm_pos_embed = self.lm.get_position_embeddings()
+        except NotImplementedError:
+            self.lm_pos_embed = None # rotrary embeds
+            
+        
+        if hasattr(self.lm,'embed_dim'):
+            self.embed_dim = self.lm.embed_dim
+        else:
+            self.embed_dim = decoder_embed_dim
+        
+        # self.asLM = asLM # if generates tokens rather than hidden states
+        # if self.asLM: # TODO: 当年写这个是为啥？
+        #     self.lm.set_output_embeddings(nn.Linear(self.embed_dim, self.self.LMconfig.vocab_size, bias=False))
+        self.freeze_decoder = False
+        if True:
+            for para in self.lm.parameters():
+                para.requires_grad = False
+        
+    def load_lm(self):
+        ## ---------------------LM setting----------------------
+        self.tokenizer = AutoTokenizer.from_pretrained(self.decoder_type)
+        if self.tokenizer.pad_token is None:
+            self.tokenizer.pad_token = self.tokenizer.eos_token
+        self.LMconfig = AutoConfig.from_pretrained(self.decoder_type, trust_remote_code=True )
+        self.lm = AutoModelForCausalLM.from_pretrained(self.decoder_type, trust_remote_code=True)
+       
+        
+    def forward(self, input_ids, flatten_embs, attention_mask, labels, **kwargs):
+        mix_embed = mixEmbed(self.lm_embed, flatten_embs)
+        self.lm.set_input_embeddings(mix_embed) # modification of the lm embed 
+        output = self.lm(input_ids=input_ids, attention_mask=attention_mask, labels=labels, output_hidden_states=True, **kwargs) 
+        self.lm.set_input_embeddings(self.lm_embed) # modification of the lm embed 
+        return output
+
+    def generate(self, input_ids, flatten_embs):
+        mix_embed = mixEmbed(self.lm_embed, flatten_embs)
+        self.lm.set_input_embeddings(mix_embed) # modification of the lm embed 
+        outputs = self.lm.generate(input_ids=input_ids, max_new_tokens=256, use_cache=False)
+        # outputs = self.lm.generate(input_ids=input_ids, 
+        #                            max_new_tokens=1024, 
+        #                            do_sample=True,
+        #                            temperature=1.5,
+        #                            num_beams=1,
+        #                            top_p=0.9,
+        #                            top_k=3,
+        #                            use_cache=False)
+        self.lm.set_input_embeddings(self.lm_embed) # modification of the lm embed 
+        return outputs
+'''
+## infer params
+max_input_tokens: 40
+batch_size_test: 16
+max_new_tokens: 64
+min_length: 2
+num_beams: 5
+length_penalty: -2.0
+top_p: 0.9
+top_k: 3
+no_repeat_ngram_size: 2
+apply_lemmatizer: False
+use_nucleus_sampling: True
+'''
+
+class LMDecoder_qlora(LMDecoder):
+    def __init__(self,
+                # num_patches=196,
+                img_size=(80,512),
+                patch_size:int=16,
+                in_chans:int=3,
+                embed_dim=1024, # encoder embed dim
+                decoder_embed_dim=512,
+                norm_cfg=dict(type='LN', eps=1e-6),
+                # patch_resolution=14,
+                decoder_type='gpt2',
+                freeze_decoder=True,
+                additional_layer:int=0,
+                ):
+        super().__init__( img_size, patch_size, in_chans, embed_dim, decoder_embed_dim, norm_cfg, decoder_type, freeze_decoder, additional_layer)
+        
+    def load_lm(self):
+        self.tokenizer = AutoTokenizer.from_pretrained(self.decoder_type)
+        self.LMconfig = AutoConfig.from_pretrained(self.decoder_type, trust_remote_code=True )
+        double_quant_config = BitsAndBytesConfig(
+            load_in_4bit=True,
+            bnb_4bit_use_double_quant=True,
+            )
+        model = AutoModelForCausalLM.from_pretrained(self.decoder_type, 
+                                                    #  device_map='auto', # if remove, can not add lora 
+                                                    # load_in_4bit=True,# if remove, can not add lora 
+                                                    # # torch_dtype=torch.bfloat16,
+                                                    #  quantization_config=double_quant_config, # if remove, can not add lora 
+                                                     trust_remote_code=True )
+
+        model.gradient_checkpointing_enable()
+        model = prepare_model_for_kbit_training(model)
+        lora_config = LoraConfig(
+            r=8, 
+            lora_alpha=32, 
+            target_modules=["query_key_value"], 
+            lora_dropout=0.05, 
+            bias="none", 
+            task_type="CAUSAL_LM"
+        )
+
+        self.lm = get_peft_model(model, lora_config)
+        self.lm.print_trainable_parameters()

htsat.py

@@ -0,0 +1,1249 @@
+# Ke Chen
+# [email protected]
+# HTS-AT: A HIERARCHICAL TOKEN-SEMANTIC AUDIO TRANSFORMER FOR SOUND CLASSIFICATION AND DETECTION
+# Some layers designed on the model
+# 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
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from itertools import repeat
+import collections.abc
+import math
+import warnings
+
+from torch.nn.init import _calculate_fan_in_and_fan_out
+import torch.utils.checkpoint as checkpoint
+
+import random
+
+from torchlibrosa.stft import Spectrogram, LogmelFilterBank
+from torchlibrosa.augmentation import SpecAugmentation
+from einops import rearrange
+from itertools import repeat
+# from .utils import interpolate
+
+# from .feature_fusion import iAFF, AFF, DAF
+
+
+'''
+Feature Fusion for Varible-Length Data Processing
+AFF/iAFF is referred and modified from https://github.com/YimianDai/open-aff/blob/master/aff_pytorch/aff_net/fusion.py
+According to the paper: Yimian Dai et al, Attentional Feature Fusion, IEEE Winter Conference on Applications of Computer Vision, WACV 2021
+'''
+
+class DAF(nn.Module):
+    '''
+    直接相加 DirectAddFuse
+    '''
+
+    def __init__(self):
+        super(DAF, self).__init__()
+
+    def forward(self, x, residual):
+        return x + residual
+
+
+class iAFF(nn.Module):
+    '''
+    多特征融合 iAFF
+    '''
+
+    def __init__(self, channels=64, r=4, type='2D'):
+        super(iAFF, self).__init__()
+        inter_channels = int(channels // r)
+
+        if type == '1D':
+            # 本地注意力
+            self.local_att = nn.Sequential(
+                nn.Conv1d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm1d(inter_channels),
+                nn.ReLU(inplace=True),
+                nn.Conv1d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm1d(channels),
+            )
+
+            # 全局注意力
+            self.global_att = nn.Sequential(
+                nn.AdaptiveAvgPool1d(1),
+                nn.Conv1d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm1d(inter_channels),
+                nn.ReLU(inplace=True),
+                nn.Conv1d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm1d(channels),
+            )
+
+            # 第二次本地注意力
+            self.local_att2 = nn.Sequential(
+                nn.Conv1d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm1d(inter_channels),
+                nn.ReLU(inplace=True),
+                nn.Conv1d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm1d(channels),
+            )
+            # 第二次全局注意力
+            self.global_att2 = nn.Sequential(
+                nn.AdaptiveAvgPool1d(1),
+                nn.Conv1d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm1d(inter_channels),
+                nn.ReLU(inplace=True),
+                nn.Conv1d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm1d(channels),
+            )
+        elif type == '2D':
+            # 本地注意力
+            self.local_att = nn.Sequential(
+                nn.Conv2d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm2d(inter_channels),
+                nn.ReLU(inplace=True),
+                nn.Conv2d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm2d(channels),
+            )
+
+            # 全局注意力
+            self.global_att = nn.Sequential(
+                nn.AdaptiveAvgPool2d(1),
+                nn.Conv2d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm2d(inter_channels),
+                nn.ReLU(inplace=True),
+                nn.Conv2d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm2d(channels),
+            )
+
+            # 第二次本地注意力
+            self.local_att2 = nn.Sequential(
+                nn.Conv2d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm2d(inter_channels),
+                nn.ReLU(inplace=True),
+                nn.Conv2d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm2d(channels),
+            )
+            # 第二次全局注意力
+            self.global_att2 = nn.Sequential(
+                nn.AdaptiveAvgPool2d(1),
+                nn.Conv2d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm2d(inter_channels),
+                nn.ReLU(inplace=True),
+                nn.Conv2d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm2d(channels),
+            )
+        else:
+            raise f'the type is not supported'
+
+        self.sigmoid = nn.Sigmoid()
+
+    def forward(self, x, residual):
+        flag = False
+        xa = x + residual
+        if xa.size(0) == 1:
+            xa = torch.cat([xa,xa],dim=0)
+            flag = True
+        xl = self.local_att(xa)
+        xg = self.global_att(xa)
+        xlg = xl + xg
+        wei = self.sigmoid(xlg)
+        xi = x * wei + residual * (1 - wei)
+
+        xl2 = self.local_att2(xi)
+        xg2 = self.global_att(xi)
+        xlg2 = xl2 + xg2
+        wei2 = self.sigmoid(xlg2)
+        xo = x * wei2 + residual * (1 - wei2)
+        if flag:
+            xo = xo[0].unsqueeze(0)
+        return xo
+
+
+class AFF(nn.Module):
+    '''
+    多特征融合 AFF
+    '''
+
+    def __init__(self, channels=64, r=4, type='2D'):
+        super(AFF, self).__init__()
+        inter_channels = int(channels // r)
+
+        if type == '1D':
+            self.local_att = nn.Sequential(
+                nn.Conv1d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm1d(inter_channels),
+                nn.ReLU(inplace=True),
+                nn.Conv1d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm1d(channels),
+            )
+            self.global_att = nn.Sequential(
+                nn.AdaptiveAvgPool1d(1),
+                nn.Conv1d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm1d(inter_channels),
+                nn.ReLU(inplace=True),
+                nn.Conv1d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm1d(channels),
+            )
+        elif type == '2D':
+            self.local_att = nn.Sequential(
+                nn.Conv2d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm2d(inter_channels),
+                nn.ReLU(inplace=True),
+                nn.Conv2d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm2d(channels),
+            )
+            self.global_att = nn.Sequential(
+                nn.AdaptiveAvgPool2d(1),
+                nn.Conv2d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm2d(inter_channels),
+                nn.ReLU(inplace=True),
+                nn.Conv2d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm2d(channels),
+            )
+        else:
+            raise f'the type is not supported.'
+        
+        self.sigmoid = nn.Sigmoid()
+
+    def forward(self, x, residual):
+        flag = False
+        xa = x + residual
+        if xa.size(0) == 1:
+            xa = torch.cat([xa,xa],dim=0)
+            flag = True
+        xl = self.local_att(xa)
+        xg = self.global_att(xa)
+        xlg = xl + xg
+        wei = self.sigmoid(xlg)
+        xo = 2 * x * wei + 2 * residual * (1 - wei)
+        if flag:
+            xo = xo[0].unsqueeze(0)
+        return xo
+
+
+# .utils
+
+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 do_mixup(x, mixup_lambda):
+    """
+    Args:
+      x: (batch_size , ...)
+      mixup_lambda: (batch_size,)
+    Returns:
+      out: (batch_size, ...)
+    """
+    out = (
+            x.transpose(0, -1) * mixup_lambda
+            + torch.flip(x, dims=[0]).transpose(0, -1) * (1 - mixup_lambda)
+    ).transpose(0, -1)
+    return out
+
+# from PyTorch internals
+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 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,
+        enable_fusion=False, fusion_type='None'):
+        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
+
+        self.enable_fusion = enable_fusion
+        self.fusion_type = fusion_type
+        
+        padding = ((patch_size[0] - patch_stride[0]) // 2, (patch_size[1] - patch_stride[1]) // 2)
+
+        if (self.enable_fusion) and (self.fusion_type == 'channel_map'):
+            self.proj = nn.Conv2d(in_chans*4, embed_dim, kernel_size=patch_size, stride=patch_stride, padding=padding)
+        else:
+            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()
+
+        if (self.enable_fusion) and (self.fusion_type in ['daf_2d','aff_2d','iaff_2d']):
+            self.mel_conv2d = nn.Conv2d(in_chans, embed_dim, kernel_size=(patch_size[0], patch_size[1]*3), stride=(patch_stride[0], patch_stride[1] * 3), padding=padding)
+            if self.fusion_type == 'daf_2d':
+                self.fusion_model = DAF()
+            elif self.fusion_type == 'aff_2d':
+                self.fusion_model = AFF(channels=embed_dim, type='2D')
+            elif self.fusion_type == 'iaff_2d':
+                self.fusion_model = iAFF(channels=embed_dim, type='2D')    
+    def forward(self, x, longer_idx = None):
+        if (self.enable_fusion) and (self.fusion_type in ['daf_2d','aff_2d','iaff_2d']):
+            global_x = x[:,0:1,:,:]
+            
+
+            # global processing
+            B, C, H, W = global_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]})."
+            global_x = self.proj(global_x)
+            TW = global_x.size(-1)
+            if len(longer_idx) > 0:
+                # local processing
+                local_x = x[longer_idx,1:,:,:].contiguous()
+                B, C, H, W = local_x.shape
+                local_x = local_x.view(B*C,1,H,W)
+                local_x = self.mel_conv2d(local_x)
+                local_x = local_x.view(B,C,local_x.size(1),local_x.size(2),local_x.size(3))
+                local_x = local_x.permute((0,2,3,1,4)).contiguous().flatten(3)
+                TB,TC,TH,_ = local_x.size()
+                if local_x.size(-1) < TW:
+                    local_x = torch.cat([local_x, torch.zeros((TB,TC,TH,TW-local_x.size(-1)), device=global_x.device)], dim=-1)
+                else:
+                    local_x = local_x[:,:,:,:TW]
+                
+                global_x[longer_idx] = self.fusion_model(global_x[longer_idx],local_x)
+            x = global_x
+        else:
+            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_grad_trunc_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_grad_trunc_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')
+
+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
+    
+        # 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
+    """
+
+    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, 
+                 enable_fusion = False, fusion_type = '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
+
+        self.enable_fusion = enable_fusion
+        self.fusion_type = fusion_type
+
+        #  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,
+            enable_fusion=self.enable_fusion, fusion_type=self.fusion_type
+            )
+
+        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)
+        
+        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)
+
+        if (self.enable_fusion) and (self.fusion_type in ['daf_1d','aff_1d','iaff_1d']):
+            self.mel_conv1d = nn.Sequential(
+                nn.Conv1d(64, 64, kernel_size=5, stride=3, padding=2),
+                nn.BatchNorm1d(64)
+            )
+            if self.fusion_type == 'daf_1d':
+                self.fusion_model = DAF()
+            elif self.fusion_type == 'aff_1d':
+                self.fusion_model = AFF(channels=64, type='1D')
+            elif self.fusion_type == 'iaff_1d':
+                self.fusion_model = iAFF(channels=64, type='1D')
+                
+        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, longer_idx = None):
+        # A deprecated optimization for using a hierarchical output from different blocks
+
+        frames_num = x.shape[2]        
+        x = self.patch_embed(x, longer_idx = longer_idx)
+        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)
+        # 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
+        fine_grained_latent_output = torch.mean(x, dim = 2)
+        fine_grained_latent_output = interpolate(fine_grained_latent_output.permute(0,2,1).contiguous(), 8 * self.patch_stride[1]) 
+        
+        latent_output = self.avgpool(torch.flatten(x,2))
+        latent_output = torch.flatten(latent_output, 1)
+
+        # display the attention map, if needed
+
+        x = self.tscam_conv(x)
+        x = torch.flatten(x, 2) # B, C, T
+ 
+        fpx = interpolate(torch.sigmoid(x).permute(0,2,1).contiguous(), 8 * self.patch_stride[1]) 
+            
+        x = self.avgpool(x)
+        x = torch.flatten(x, 1)
+
+        output_dict = {
+            'framewise_output': fpx, # already sigmoided
+            'clipwise_output': torch.sigmoid(x),
+            'fine_grained_embedding': fine_grained_latent_output,
+            'embedding': latent_output
+        }
+
+        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_generator(self, x: torch.Tensor, mixup_lambda = None, infer_mode = False, device=None):# out_feat_keys: List[str] = None):
+
+        n = int(x.shape[1]/480000)
+        assert n * 480000 == x.shape[1]
+        x = rearrange(x, 'b (n t) -> (b n) t', n=n)
+        if not self.enable_fusion:
+            # x = x["waveform"].to(device=device, non_blocking=True)
+            x = x.to(device=device, non_blocking=True)
+            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)
+                
+            x = self.reshape_wav2img(x)
+            # output_dict = self.forward_features(x)
+        
+        # A deprecated optimization for using a hierarchical output from different blocks
+        longer_idx = None
+        frames_num = x.shape[2]        
+        x = self.patch_embed(x, longer_idx = longer_idx)
+        if self.ape:
+            x = x + self.absolute_pos_embed
+        x = self.pos_drop(x)
+        for i, layer in enumerate(self.layers[:3]): # depth: [2,2,12,2]
+            if i == 2:
+                for blk in layer.blocks:
+                    x, attn = blk(x)
+                    # 512
+                    x = rearrange(x, '(b n) t c -> b (n t) c', n=n)
+                    x = x if (new_x:=(yield x)) is None else new_x
+                    x = rearrange(x, 'b (n t) c -> (b n) t c', n=n)
+            else:
+                x, attn = layer(x)
+
+
+        
+    def forward(self, x: torch.Tensor, mixup_lambda = None, infer_mode = False, device=None):# out_feat_keys: List[str] = None):
+
+        n = int(x.shape[1] / 480000)
+        assert n * 480000 == x.shape[1]
+        x = rearrange(x, 'b (n t) -> (b n) t', n = n)
+        if not self.enable_fusion:
+            # x = x["waveform"].to(device=device, non_blocking=True)
+            x = x.to(device=device, non_blocking=True)
+            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)
+                
+            x = self.reshape_wav2img(x)
+            # x = self.forward_features(x)
+            
+        longer_idx = None
+        frames_num = x.shape[2]        
+        x = self.patch_embed(x, longer_idx = longer_idx)
+        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)
+        # for x
+        x = self.norm(x)
+        x = rearrange(x, '(b n) t c -> b (n t) c', n = n)
+        return 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
+        # fine_grained_latent_output = torch.mean(x, dim = 2)
+        # fine_grained_latent_output = interpolate(fine_grained_latent_output.permute(0,2,1).contiguous(), 8 * self.patch_stride[1]) 
+        
+        # latent_output = self.avgpool(torch.flatten(x,2))
+        # latent_output = torch.flatten(latent_output, 1)
+
+        # # display the attention map, if needed
+
+        # x = self.tscam_conv(x)
+        # x = torch.flatten(x, 2) # B, C, T
+ 
+        # fpx = interpolate(torch.sigmoid(x).permute(0,2,1).contiguous(), 8 * self.patch_stride[1]) 
+            
+        # x = self.avgpool(x)
+        # x = torch.flatten(x, 1)
+        # return x
+
+def create_htsat_model(audio_cfg, enable_fusion=False, fusion_type='None'):
+    try:
+
+        assert audio_cfg.model_name in ["tiny", "base", "large"], "model name for HTS-AT is wrong!"
+        if audio_cfg.model_name == "tiny":
+            model = HTSAT_Swin_Transformer(
+                spec_size=256,
+                patch_size=4,
+                patch_stride=(4,4),
+                num_classes=audio_cfg.class_num,
+                embed_dim=96,
+                depths=[2,2,6,2],
+                num_heads=[4,8,16,32],
+                window_size=8,
+                config = audio_cfg,
+                enable_fusion = enable_fusion,
+                fusion_type = fusion_type
+            )
+        elif audio_cfg.model_name == "base":
+            model = HTSAT_Swin_Transformer(
+                spec_size=256,
+                patch_size=4,
+                patch_stride=(4,4),
+                num_classes=audio_cfg.class_num,
+                embed_dim=128,
+                depths=[2,2,12,2],
+                num_heads=[4,8,16,32],
+                window_size=8,
+                config = audio_cfg,
+                enable_fusion = enable_fusion,
+                fusion_type = fusion_type
+            )
+        elif audio_cfg.model_name == "large":
+            model = HTSAT_Swin_Transformer(
+                spec_size=256,
+                patch_size=4,
+                patch_stride=(4,4),
+                num_classes=audio_cfg.class_num,
+                embed_dim=256,
+                depths=[2,2,12,2],
+                num_heads=[4,8,16,32],
+                window_size=8,
+                config = audio_cfg,
+                enable_fusion = enable_fusion,
+                fusion_type = fusion_type
+            )
+        
+        return model
+    except:
+        raise RuntimeError(f'Import Model for {audio_cfg.model_name} not found, or the audio cfg parameters are not enough.')
+        

mae_vit.py

@@ -0,0 +1,303 @@
+import torch
+from mmcls.models import VisionTransformer
+from torch import nn
+from torch.utils.checkpoint import checkpoint
+import copy
+
+def build_2d_sincos_position_embedding(patches_resolution,
+                                       embed_dims,
+                                       temperature=10000.,
+                                       cls_token=False):
+    """The function is to build position embedding for model to obtain the
+    position information of the image patches."""
+
+    if isinstance(patches_resolution, int):
+        patches_resolution = (patches_resolution, patches_resolution)
+
+    h, w = patches_resolution
+    grid_w = torch.arange(w, dtype=torch.float32)
+    grid_h = torch.arange(h, dtype=torch.float32)
+    grid_w, grid_h = torch.meshgrid(grid_w, grid_h)
+    assert embed_dims % 4 == 0, \
+        'Embed dimension must be divisible by 4.'
+    pos_dim = embed_dims // 4
+
+    omega = torch.arange(pos_dim, dtype=torch.float32) / pos_dim
+    omega = 1. / (temperature**omega)
+    out_w = torch.einsum('m,d->md', [grid_w.flatten(), omega])
+    out_h = torch.einsum('m,d->md', [grid_h.flatten(), omega])
+
+    pos_emb = torch.cat(
+        [
+            torch.sin(out_w),
+            torch.cos(out_w),
+            torch.sin(out_h),
+            torch.cos(out_h)
+        ],
+        dim=1,
+    )[None, :, :]
+
+    if cls_token:
+        cls_token_pe = torch.zeros([1, 1, embed_dims], dtype=torch.float32)
+        pos_emb = torch.cat([cls_token_pe, pos_emb], dim=1)
+
+    return pos_emb
+
+
+
+class MAEViT(VisionTransformer):
+    """Vision Transformer for MAE pre-training.
+
+    A PyTorch implement of: `An Image is Worth 16x16 Words: Transformers
+    for Image Recognition at Scale <https://arxiv.org/abs/2010.11929>`_
+
+    Args:
+        arch (str | dict): Vision Transformer architecture
+            Default: 'b'
+        img_size (int | tuple): Input image size
+        patch_size (int | tuple): The patch size
+        out_indices (Sequence | int): Output from which stages.
+            Defaults to -1, means the last stage.
+        drop_rate (float): Probability of an element to be zeroed.
+            Defaults to 0.
+        drop_path_rate (float): stochastic depth rate. Defaults to 0.
+        norm_cfg (dict): Config dict for normalization layer.
+            Defaults to ``dict(type='LN')``.
+        final_norm (bool): Whether to add a additional layer to normalize
+            final feature map. Defaults to True.
+        output_cls_token (bool): Whether output the cls_token. If set True,
+            `with_cls_token` must be True. Defaults to True.
+        interpolate_mode (str): Select the interpolate mode for position
+            embeding vector resize. Defaults to "bicubic".
+        patch_cfg (dict): Configs of patch embeding. Defaults to an empty dict.
+        layer_cfgs (Sequence | dict): Configs of each transformer layer in
+            encoder. Defaults to an empty dict.
+        mask_ratio (bool): The ratio of total number of patches to be masked.
+            Defaults to 0.75.
+        init_cfg (dict, optional): Initialization config dict.
+            Defaults to None.
+    """
+
+    arch_zoo = {
+        **dict.fromkeys(
+            ['mocov3-s', 'mocov3-small'], {
+                'embed_dims': 384,
+                'num_layers': 12,
+                'num_heads': 12,
+                'feedforward_channels': 1536,
+            }),
+        **dict.fromkeys(
+            ['b', 'base'], {
+                'embed_dims': 768,
+                'num_layers': 12,
+                'num_heads': 12,
+                'feedforward_channels': 3072
+            }),
+    }
+
+
+
+    def __init__(self,
+                 arch='b',
+                 img_size=224,
+                 patch_size=16,
+                 out_indices=-1,
+                 drop_rate=0,
+                 drop_path_rate=0,
+                 norm_cfg=dict(type='LN', eps=1e-6),
+                 final_norm=True,
+                 output_cls_token=False,
+                 interpolate_mode='bicubic',
+                 patch_cfg=dict(),
+                 layer_cfgs=dict(),
+                 gradientCKPT=False,
+                 mask_ratio=0.75,
+                 init_cfg=None):
+        super().__init__(
+            arch=arch,
+            img_size=img_size,
+            patch_size=patch_size,
+            out_indices=out_indices,
+            drop_rate=drop_rate,
+            drop_path_rate=drop_path_rate,
+            norm_cfg=norm_cfg,
+            final_norm=final_norm,
+            output_cls_token=output_cls_token,
+            interpolate_mode=interpolate_mode,
+            patch_cfg=patch_cfg,
+            layer_cfgs=layer_cfgs,
+            init_cfg=init_cfg)
+        self.gradientCKPT = gradientCKPT
+        self.pos_embed.requires_grad = False
+        self.mask_ratio = mask_ratio
+        self.num_patches = self.patch_resolution[0] * self.patch_resolution[1]
+        # self.mask_embedding = copy.deepcopy(self.patch_embed)
+        # self.mask_embedding.norm = None
+
+    def init_weights(self):
+        super(MAEViT, self).init_weights()
+        if not (isinstance(self.init_cfg, dict)
+                and self.init_cfg['type'] == 'Pretrained'):
+            # initialize position  embedding in backbone
+            pos_embed = build_2d_sincos_position_embedding(
+                self.patch_resolution,
+                self.pos_embed.shape[-1],
+                cls_token=True)
+            self.pos_embed.data.copy_(pos_embed.float())
+
+            w = self.patch_embed.projection.weight.data
+            torch.nn.init.xavier_uniform_(w.view([w.shape[0], -1]))
+
+            torch.nn.init.normal_(self.cls_token, std=.02)
+
+            self.apply(self._init_weights)
+
+        # mask_embedding transfers pixel level mask to token level
+        # self.mask_embedding.apply(self._init_mask_embedding)
+        # for para in self.mask_embedding.parameters():
+        #     para.requires_grad = False
+
+    def _init_mask_embedding(self,m):
+        if hasattr(m,'weight'):
+            nn.init.constant_(m.weight,1.0)
+        if hasattr(m, 'bias'):
+            nn.init.constant_(m.bias,0)
+
+    def _init_weights(self, m):
+
+        if isinstance(m, nn.Linear):
+            torch.nn.init.xavier_uniform_(m.weight)
+            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)
+
+    def random_masking(self, x, mask_ratio=0.75, attn_mask=None):
+        """Generate the mask for MAE Pre-training.
+
+        Args:
+            x (torch.tensor): Image with data augmentation applied.
+            mask_ratio (float): The mask ratio of total patches.
+                Defaults to 0.75.
+
+        Returns:
+            tuple[Tensor, Tensor, Tensor]: masked image, mask and the ids
+                to restore original image.
+
+            - x_masked (Tensor): masked image.
+            - mask (Tensor): mask used to mask image.
+            - ids_restore (Tensor): ids to restore original image.
+        """
+        N, L, D = x.shape  # batch, length, dim
+        len_keep = int(L * (1 - mask_ratio))
+
+        noise = torch.rand(N, L, device=x.device)  # noise in [0, 1]
+
+        # sort noise for each sample
+        ids_shuffle = torch.argsort(
+            noise, dim=1)  # ascend: small is keep, large is remove
+        ids_restore = torch.argsort(ids_shuffle, dim=1)
+
+        # keep the first subset
+        ids_keep = ids_shuffle[:, :len_keep]
+        x_masked = torch.gather(
+            x, dim=1, index=ids_keep.unsqueeze(-1).repeat(1, 1, D))
+        # modified_attn_mask = None if attn_mask is None else torch.gather(attn_mask,dim=1, index=ids_keep)
+
+        # generate the binary mask: 0 is keep, 1 is remove
+        mask = torch.ones([N, L], device=x.device)
+        mask[:, :len_keep] = 0
+        # unshuffle to get the binary mask
+        mask = torch.gather(mask, dim=1, index=ids_restore)
+
+        return x_masked, mask, ids_restore #, modified_attn_mask
+
+    def generate_mask(self, pixel_level_attn_mask):
+        '''
+        pixel_level_attn_mask: (0,1) attn mask with the same shape as img
+        '''
+        if pixel_level_attn_mask is None: return None
+        # H, W = patch_resolution
+        # B, C = pixel_level_attn_mask.shape[:2]
+        # attn_mask = torch.ones((B,C,H,W),device=pixel_level_attn_mask) 
+        # H_splited = torch.chunk(pixel_level_attn_mask, H, -2)
+        # HW_splited_mask = (torch.chunk(Hs, W, -1) for Hs in H_splited)
+
+        #         if HW_splited_mask[:,:,hi,wi].sum().item() == 0:
+        #             attn_mask[:,:,hi,wi] = 0
+
+        # mask_patches = self.mask_embedding(pixel_level_attn_mask)[0]
+        # attn_mask = mask_patches.sum(-1) != 0
+
+        # return attn_mask
+
+    def extract_feat(self, img ,attn_mask=None):
+        x, *_ = self.forward(img,attn_mask)
+        if self.output_cls_token:
+            return x[:,0,:]
+        else:
+            return torch.mean(x,dim=1)
+
+    def forward(self, x, attn_mask=None):
+        if attn_mask is not None: assert self.output_cls_token
+        
+        B = x.shape[0]
+        x = self.patch_embed(x)[0]
+        # add pos embed w/o cls token
+        x = x + self.pos_embed[:, 1:1+x.shape[1], :]
+        # masking: length -> length * mask_ratio
+        if True:
+            assert self.mask_ratio == 0.
+        else:
+            x, mask, ids_restore = self.random_masking(x, self.mask_ratio)
+
+        # append cls token
+        cls_token = self.cls_token + self.pos_embed[:, :1, :]
+        cls_tokens = cls_token.expand(B, -1, -1)
+        x = torch.cat((cls_tokens, x), dim=1)
+        x = self.drop_after_pos(x)
+        # if attn_mask is not None: 
+        #     attn_mask = torch.concat((torch.ones((B,1),device=attn_mask.device) , attn_mask),dim=1)
+
+        for i, layer in enumerate(self.layers):
+            if self.gradientCKPT:
+                x = checkpoint(layer,x) # ,attn_mask
+            else:
+                x = layer(x) # ,attn_mask
+            if i == len(self.layers) - 1 and self.final_norm:
+                x = self.norm1(x)
+        if True: 
+            return x
+        else:
+            return (x, mask, ids_restore)
+
+    def forward_generator(self, x, attn_mask=None):
+        if attn_mask is not None: assert self.output_cls_token
+        
+        B = x.shape[0]
+        x = self.patch_embed(x)[0]
+        # add pos embed w/o cls token
+        x = x + self.pos_embed[:, 1:1+x.shape[1], :]
+
+        # append cls token
+        cls_token = self.cls_token + self.pos_embed[:, :1, :]
+        cls_tokens = cls_token.expand(B, -1, -1)
+        x = torch.cat((cls_tokens, x), dim=1)
+        x = self.drop_after_pos(x)
+
+        for i, layer in enumerate(self.layers):
+            if self.gradientCKPT:
+                x = checkpoint(layer,x) # ,attn_mask
+            else:
+                x = layer(x) # ,attn_mask
+
+            if i == len(self.layers) - 1 and self.final_norm:
+                x = self.norm1(x)
+         
+            x = x if (new_x:=(yield x)) is None else new_x 
+
+            debug = False
+            if debug:
+                print(f'layer {i}-th forwarded')
+ 

vision_transformer.py

@@ -0,0 +1,176 @@
+import math
+from functools import reduce
+from operator import mul
+from ipdb import set_trace
+
+import torch
+import torch.nn.functional as F
+import torch.nn as nn
+from mmcls.models.backbones import VisionTransformer as _VisionTransformer
+from mmcls.models.utils import to_2tuple
+from mmcv.cnn.bricks.transformer import PatchEmbed
+from torch.nn.modules.batchnorm import _BatchNorm
+
+
+def build_2d_sincos_position_embedding(patches_resolution,
+                                       embed_dims,
+                                       temperature=10000.,
+                                       cls_token=False):
+    """The function is to build position embedding for model to obtain the
+    position information of the image patches."""
+
+    if isinstance(patches_resolution, int):
+        patches_resolution = (patches_resolution, patches_resolution)
+
+    h, w = patches_resolution
+    grid_w = torch.arange(w, dtype=torch.float32)
+    grid_h = torch.arange(h, dtype=torch.float32)
+    grid_w, grid_h = torch.meshgrid(grid_w, grid_h)
+    assert embed_dims % 4 == 0, \
+        'Embed dimension must be divisible by 4.'
+    pos_dim = embed_dims // 4
+
+    omega = torch.arange(pos_dim, dtype=torch.float32) / pos_dim
+    omega = 1. / (temperature**omega)
+    out_w = torch.einsum('m,d->md', [grid_w.flatten(), omega])
+    out_h = torch.einsum('m,d->md', [grid_h.flatten(), omega])
+
+    pos_emb = torch.cat(
+        [
+            torch.sin(out_w),
+            torch.cos(out_w),
+            torch.sin(out_h),
+            torch.cos(out_h)
+        ],
+        dim=1,
+    )[None, :, :]
+
+    if cls_token:
+        cls_token_pe = torch.zeros([1, 1, embed_dims], dtype=torch.float32)
+        pos_emb = torch.cat([cls_token_pe, pos_emb], dim=1)
+
+    return pos_emb
+
+
+class VisionTransformer(_VisionTransformer):
+    """Vision Transformer.
+
+    A pytorch implement of: `An Images is Worth 16x16 Words: Transformers for
+    Image Recognition at Scale <https://arxiv.org/abs/2010.11929>`_.
+
+    Part of the code is modified from:
+    `<https://github.com/facebookresearch/moco-v3/blob/main/vits.py>`_.
+
+    Args:
+        stop_grad_conv1 (bool, optional): whether to stop the gradient of
+            convolution layer in `PatchEmbed`. Defaults to False.
+        frozen_stages (int): Stages to be frozen (stop grad and set eval mode).
+            -1 means not freezing any parameters. Defaults to -1.
+        norm_eval (bool): Whether to set norm layers to eval mode, namely,
+            freeze running stats (mean and var). Note: Effect on Batch Norm
+            and its variants only. Defaults to False.
+        init_cfg (dict or list[dict], optional): Initialization config dict.
+            Defaults to None.
+    """
+
+    arch_zoo = {
+        **dict.fromkeys(
+            ['mocov3-s', 'mocov3-small'], {
+                'embed_dims': 384,
+                'num_layers': 12,
+                'num_heads': 12,
+                'feedforward_channels': 1536,
+            }),
+        **dict.fromkeys(
+            ['b', 'base'], {
+                'embed_dims': 768,
+                'num_layers': 12,
+                'num_heads': 12,
+                'feedforward_channels': 3072
+            }),
+    }
+
+    def __init__(self,
+                 stop_grad_conv1=False,
+                 frozen_stages=-1,
+                 norm_eval=False,
+                 init_cfg=None,
+                 **kwargs):
+        super(VisionTransformer, self).__init__(init_cfg=init_cfg,)
+        self.patch_size = kwargs['patch_size']
+        self.frozen_stages = frozen_stages
+        self.norm_eval = norm_eval
+        self.init_cfg = init_cfg
+        
+        
+        if isinstance(self.patch_embed, PatchEmbed):
+            if stop_grad_conv1:
+                self.patch_embed.projection.weight.requires_grad = False
+                self.patch_embed.projection.bias.requires_grad = False
+
+        self._freeze_stages()
+
+    def init_weights(self):
+        super(VisionTransformer, self).init_weights()
+
+        if not (isinstance(self.init_cfg, dict)
+                and self.init_cfg['type'] == 'Pretrained'):
+
+            # Use fixed 2D sin-cos position embedding
+            pos_emb = build_2d_sincos_position_embedding(
+                patches_resolution=self.patch_resolution,
+                embed_dims=self.embed_dims,
+                cls_token=True)
+            self.pos_embed.data.copy_(pos_emb)
+            self.pos_embed.requires_grad = False
+
+            # xavier_uniform initialization for PatchEmbed
+            if isinstance(self.patch_embed, PatchEmbed):
+                val = math.sqrt(
+                    6. / float(3 * reduce(mul, to_2tuple(self.patch_size), 1) +
+                               self.embed_dims))
+                nn.init.uniform_(self.patch_embed.projection.weight, -val, val)
+                nn.init.zeros_(self.patch_embed.projection.bias)
+
+            # initialization for linear layers
+            for name, m in self.named_modules():
+                if isinstance(m, nn.Linear):
+                    if 'qkv' in name:
+                        # treat the weights of Q, K, V separately
+                        val = math.sqrt(
+                            6. /
+                            float(m.weight.shape[0] // 3 + m.weight.shape[1]))
+                        nn.init.uniform_(m.weight, -val, val)
+                    else:
+                        nn.init.xavier_uniform_(m.weight)
+                    nn.init.zeros_(m.bias)
+            nn.init.normal_(self.cls_token, std=1e-6)
+
+    def _freeze_stages(self):
+        """Freeze patch_embed layer, some parameters and stages."""
+        if self.frozen_stages >= 0:
+            self.patch_embed.eval()
+            for param in self.patch_embed.parameters():
+                param.requires_grad = False
+
+            self.cls_token.requires_grad = False
+            self.pos_embed.requires_grad = False
+
+        for i in range(1, self.frozen_stages + 1):
+            m = self.layers[i - 1]
+            m.eval()
+            for param in m.parameters():
+                param.requires_grad = False
+
+            if i == (self.num_layers) and self.final_norm:
+                for param in getattr(self, 'norm1').parameters():
+                    param.requires_grad = False
+
+    def train(self, mode=True):
+        super(VisionTransformer, self).train(mode)
+        self._freeze_stages()
+        if mode and self.norm_eval:
+            for m in self.modules():
+                # trick: eval have effect on BatchNorm only
+                if isinstance(m, _BatchNorm):
+                    m.eval()