""" Swin Transformer A PyTorch impl of : `Swin Transformer: Hierarchical Vision Transformer using Shifted Windows` - https://arxiv.org/pdf/2103.14030 Code/weights from https://github.com/microsoft/Swin-Transformer, original copyright/license info below """ # -------------------------------------------------------- # Swin Transformer # Copyright (c) 2021 Microsoft # Licensed under The MIT License [see LICENSE for details] # Written by Ze Liu # -------------------------------------------------------- import logging import math from copy import deepcopy from typing import Optional import torch import torch.nn as nn import torch.nn.functional as F import torch.utils.checkpoint as checkpoint import numpy as np from timm.models.layers import DropPath, to_2tuple, trunc_normal_ class Mlp(nn.Module): """Multilayer perceptron.""" def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.0): 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 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): """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=0.0, proj_drop=0.0, dim_text=None ): 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=0.02) self.softmax = nn.Softmax(dim=-1) # dim_text = 768 if dim_text is not None: self.qkv_text_i2t = nn.Linear(dim_text, dim * 2, bias=qkv_bias) self.qkv_i2t = nn.Linear(dim, dim, bias=qkv_bias) self.attn_drop_i2t = nn.Dropout(attn_drop) self.proj_i2t = nn.Linear(dim, dim) self.proj_drop_i2t = nn.Dropout(proj_drop) # self.proj_i2t = nn.Linear(dim, dim) self.alpha_i2t = nn.Parameter(torch.Tensor([0])) # self.gate_i2t = nn.Linear(2*dim, 1) # self.gate_i2t = nn.Linear(2*dim, dim) # self.sigmoid_i2t = nn.Sigmoid() """self.i2t_relative_position_bias = nn.Parameter( torch.zeros(2, num_heads, ntext)) # (2, nH, ntext) self.t2t_relative_position_bias = nn.Parameter( torch.zeros(num_heads, ntext, ntext)) # (nH, ntext, ntext) trunc_normal_(self.i2t_relative_position_bias, std=.02) trunc_normal_(self.t2t_relative_position_bias, std=.02)#""" def forward(self, x, mask: Optional[torch.Tensor] = None, y=None, y_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) attn = self.attn_drop(attn) x = (attn @ v).transpose(1, 2).reshape(B_, N, C) x = self.proj(x) x = self.proj_drop(x) if y is not None: B_text, N_text, C_text = y.shape nW = B_ // B_text # number of windows assert B_text * nW == B_, "B_ is not a multiplier of B_text in window attention" # notice that after qkv_text, the hidden dimension is C instead of C_text qkv_text = ( self.qkv_text_i2t(y) .reshape(B_text, N_text, 2, self.num_heads, C // self.num_heads) .permute(2, 0, 3, 1, 4) ) k_text, v_text = qkv_text[0], qkv_text[1] k_text = torch.repeat_interleave(k_text, nW, dim=0) v_text = torch.repeat_interleave(v_text, nW, dim=0) # TODO: remove q_text q_i2t = self.qkv_i2t(x).reshape(B_, N, 1, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4) q_i2t = q_i2t[0] # image to text attention # attn_i2t = (q_i2t @ torch.repeat_interleave(k_text, nW, dim=0).transpose(-2, -1)) # B_, nH, N, N_text # print(q_i2t.size()) # print(k_text.size()) # torch.Size([4096, 4, 49, 32]) # torch.Size([4096, 4, 50, 32]) text_scale = k_text.size(-1) ** -0.5 q_i2t = q_i2t * text_scale attn_i2t = q_i2t @ k_text.transpose(-2, -1) # B_, nH, N, N_text # add image to text bias and text_mask if y_mask is not None: mask_and_i2t_bias = y_mask.view( B_text, 1, 1, N_text ) # + self.i2t_relative_position_bias[:1].expand(B_text, -1, -1).unsqueeze(-2) # B_text, nH, 1, N_text attn_i2t = attn_i2t + torch.repeat_interleave(mask_and_i2t_bias, nW, dim=0) attn_i2t = self.softmax(attn_i2t) attn_i2t = self.attn_drop_i2t(attn_i2t) # torch.Size([4096, 4, 49, 50]) # torch.Size([64, 4, 50, 32]) # print(attn_i2t.size()) # print(v_text.size()) # 1/0 y = (attn_i2t @ v_text).transpose(1, 2).reshape(B_, N, C) y = self.proj_i2t(y) y = self.proj_drop_i2t(y) # g = torch.cat([x, y], dim=-1) # g = (self.gate_i2t(g)) # g = self.sigmoid_i2t(self.gate_i2t(g)) # x = x+g*y x = x + self.alpha_i2t * y return x class SwinTransformerBlock(nn.Module): """Swin Transformer Block. Args: dim (int): Number of input channels. 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 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, num_heads, window_size=7, shift_size=0, mlp_ratio=4.0, qkv_bias=True, qk_scale=None, drop=0.0, attn_drop=0.0, drop_path=0.0, act_layer=nn.GELU, norm_layer=nn.LayerNorm, dim_text=None, ): super().__init__() self.dim = dim self.num_heads = num_heads self.window_size = window_size self.shift_size = shift_size self.mlp_ratio = mlp_ratio 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, dim_text=dim_text, ) self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity() self.norm2 = norm_layer(dim) mlp_hidden_dim = int(dim * mlp_ratio) self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop) self.H = None self.W = None def forward(self, x, mask_matrix, x_text=None, mask_text=None): B, L, C = x.shape H, W = self.H, self.W assert L == H * W, "input feature has wrong size" shortcut = x x = self.norm1(x) x = x.view(B, H, W, C) # pad feature maps to multiples of window size pad_l = pad_t = 0 pad_r = (self.window_size - W % self.window_size) % self.window_size pad_b = (self.window_size - H % self.window_size) % self.window_size x = F.pad(x, (0, 0, pad_l, pad_r, pad_t, pad_b)) _, Hp, Wp, _ = x.shape # cyclic shift if self.shift_size > 0: shifted_x = torch.roll(x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2)) attn_mask = mask_matrix else: shifted_x = x attn_mask = None # 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 = self.attn( x_windows, mask=attn_mask, y=x_text, y_mask=mask_text ) # 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, Hp, Wp) # 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 if pad_r > 0 or pad_b > 0: x = x[:, :H, :W, :].contiguous() 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 class PatchMerging(nn.Module): """Patch Merging Layer Args: dim (int): Number of input channels. norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm """ def __init__(self, dim, norm_layer=nn.LayerNorm): super().__init__() self.dim = dim self.reduction = nn.Linear(4 * dim, 2 * dim, bias=False) self.norm = norm_layer(4 * dim) def forward(self, x, H, W): """Forward function. Args: x: Input feature, tensor size (B, H*W, C). H, W: Spatial resolution of the input feature. """ B, L, C = x.shape assert L == H * W, "input feature has wrong size" # TODO: Keep? assert H % 2 == 0 and W % 2 == 0, f"x size ({H}*{W}) are not even." x = x.view(B, H, W, C) # padding pad_input = (H % 2 == 1) or (W % 2 == 1) if pad_input: x = F.pad(x, (0, 0, 0, W % 2, 0, H % 2)) 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 # TODO: Keep? # def extra_repr(self) -> str: # return f"input_resolution={self.input_resolution}, dim={self.dim}" # # def flops(self): # H, W = self.input_resolution # flops = H * W * self.dim # flops += (H // 2) * (W // 2) * 4 * self.dim * 2 * self.dim # return flops class BasicLayer(nn.Module): """A basic Swin Transformer layer for one stage. Args: dim (int): Number of feature channels depth (int): Depths of this stage. num_heads (int): Number of attention head. window_size (int): Local window size. Default: 7. mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4. 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, depth, num_heads, window_size, mlp_ratio=4.0, qkv_bias=True, qk_scale=None, drop=0.0, attn_drop=0.0, drop_path=0.0, norm_layer=nn.LayerNorm, downsample=None, use_checkpoint=False, dim_text=None, ): super().__init__() self.window_size = window_size self.shift_size = window_size // 2 self.depth = depth self.use_checkpoint = use_checkpoint # build blocks self.blocks = nn.ModuleList( [ SwinTransformerBlock( dim=dim, 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, dim_text=(768 if i >= 14 else dim_text), ) for i in range(depth) ] ) # patch merging layer if downsample is not None: self.downsample = downsample(dim=dim, norm_layer=norm_layer) else: self.downsample = None def get_attention_mask(self, H, W, device): # calculate attention mask for SW-MSA Hp = int(np.ceil(H / self.window_size)) * self.window_size Wp = int(np.ceil(W / self.window_size)) * self.window_size img_mask = torch.zeros((1, Hp, Wp, 1), device=device) # 1 Hp Wp 1 h_slices = ( slice(0, -self.window_size), slice(-self.window_size, -self.shift_size), slice(-self.shift_size, None), ) w_slices = ( slice(0, -self.window_size), slice(-self.window_size, -self.shift_size), slice(-self.shift_size, None), ) cnt = 0 for h in h_slices: for w in w_slices: img_mask[:, h, w, :] = cnt cnt += 1 mask_windows = window_partition(img_mask, self.window_size) # nW, window_size, window_size, 1 mask_windows = mask_windows.view(-1, self.window_size * self.window_size) attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2) attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(attn_mask == 0, float(0.0)) return attn_mask def forward(self, x, H, W, x_text=None, mask_text=None): """Forward function. Args: x: Input feature, tensor size (B, H*W, C). H, W: Spatial resolution of the input feature. x_text: input text features with shape of (B_text, N_text, C_text) mask_text: (0/-inf) mask with shape of (B_text, N_text) or None; """ attn_mask = self.get_attention_mask(H, W, x.device) for blk in self.blocks: blk.H, blk.W = H, W if not torch.jit.is_scripting() and self.use_checkpoint: x = checkpoint.checkpoint(blk, x, attn_mask, x_text, mask_text) else: x = blk(x, mask_matrix=attn_mask, x_text=x_text, mask_text=mask_text) # print(x.size()) if self.downsample is not None: x_down = self.downsample(x, H, W) Wh, Ww = (H + 1) // 2, (W + 1) // 2 return x, H, W, x_down, Wh, Ww else: return x, H, W, x, H, W # TODO: Keep? # def extra_repr(self) -> str: # return f"dim={self.dim}, input_resolution={self.input_resolution}, depth={self.depth}" class PatchEmbed(nn.Module): """Image to Patch Embedding Args: patch_size (int): Patch token size. Default: 4. in_chans (int): Number of input image channels. Default: 3. embed_dim (int): Number of linear projection output channels. Default: 96. norm_layer (nn.Module, optional): Normalization layer. Default: None """ def __init__(self, patch_size=4, in_chans=3, embed_dim=96, norm_layer=None): super().__init__() patch_size = to_2tuple(patch_size) self.patch_size = patch_size self.in_chans = in_chans self.embed_dim = embed_dim self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size) if norm_layer is not None: self.norm = norm_layer(embed_dim) else: self.norm = None def forward(self, x): """Forward function.""" # padding _, _, H, W = x.size() if W % self.patch_size[1] != 0: x = F.pad(x, (0, self.patch_size[1] - W % self.patch_size[1])) if H % self.patch_size[0] != 0: x = F.pad(x, (0, 0, 0, self.patch_size[0] - H % self.patch_size[0])) x = self.proj(x) # B C Wh Ww if self.norm is not None: Wh, Ww = x.size(2), x.size(3) x = x.flatten(2).transpose(1, 2) x = self.norm(x) x = x.transpose(1, 2).view(-1, self.embed_dim, Wh, Ww) return x class SwinTransformer(nn.Module): """Swin Transformer backbone. A PyTorch impl of : `Swin Transformer: Hierarchical Vision Transformer using Shifted Windows` - https://arxiv.org/pdf/2103.14030 Args: pretrain_img_size (int): Input image size for training the pretrained model, used in absolute postion embedding. Default 224. patch_size (int | tuple(int)): Patch size. Default: 4. in_chans (int): Number of input image channels. Default: 3. embed_dim (int): Number of linear projection output channels. Default: 96. depths (tuple[int]): Depths of each Swin Transformer stage. num_heads (tuple[int]): Number of attention head of each stage. window_size (int): Window size. Default: 7. 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. drop_rate (float): Dropout rate. attn_drop_rate (float): Attention dropout rate. Default: 0. drop_path_rate (float): Stochastic depth rate. Default: 0.2. 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. out_indices (Sequence[int]): Output from which stages. frozen_stages (int): Stages to be frozen (stop grad and set eval mode). -1 means not freezing any parameters. use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False. """ def __init__( self, pretrain_img_size=224, patch_size=4, in_chans=3, embed_dim=96, depths=[2, 2, 6, 2], num_heads=[3, 6, 12, 24], window_size=7, mlp_ratio=4.0, qkv_bias=True, qk_scale=None, drop_rate=0.0, attn_drop_rate=0.0, drop_path_rate=0.2, norm_layer=nn.LayerNorm, ape=False, patch_norm=True, frozen_stages=-1, use_checkpoint=False, out_features=["stage2", "stage3", "stage4", "stage5"], backbone_arch="SWINT-FPN-RETINANET", max_query_len=None, lang_dim=None, ): super(SwinTransformer, self).__init__() print("VISION BACKBONE USE GRADIENT CHECKPOINTING: ", use_checkpoint) self.pretrain_img_size = pretrain_img_size self.num_layers = len(depths) self.embed_dim = embed_dim self.ape = ape self.patch_norm = patch_norm self.frozen_stages = frozen_stages self.out_features = out_features # split image into non-overlapping patches self.patch_embed = PatchEmbed( patch_size=patch_size, in_chans=in_chans, embed_dim=embed_dim, norm_layer=norm_layer if self.patch_norm else None, ) # absolute position embedding if self.ape: pretrain_img_size = to_2tuple(pretrain_img_size) patch_size = to_2tuple(patch_size) patches_resolution = [pretrain_img_size[0] // patch_size[0], pretrain_img_size[1] // patch_size[1]] self.absolute_pos_embed = nn.Parameter( torch.zeros(1, embed_dim, patches_resolution[0], patches_resolution[1]) ) trunc_normal_(self.absolute_pos_embed, std=0.02) self.pos_drop = nn.Dropout(p=drop_rate) # stochastic depth dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))] # stochastic depth decay rule self._out_feature_strides = {} self._out_feature_channels = {} # build layers self.layers = nn.ModuleList() for i_layer in range(self.num_layers): layer = BasicLayer( dim=int(embed_dim * 2**i_layer), depth=depths[i_layer], num_heads=num_heads[i_layer], window_size=window_size, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale, drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[sum(depths[:i_layer]) : sum(depths[: i_layer + 1])], norm_layer=norm_layer, downsample=PatchMerging if (i_layer < self.num_layers - 1) else None, use_checkpoint=use_checkpoint and i_layer > self.frozen_stages - 1, dim_text=(768 if i_layer == 3 else None), ) # TODO: Make this general : lang_dim not 768 self.layers.append(layer) stage = f"stage{i_layer + 2}" if stage in self.out_features: self._out_feature_channels[stage] = embed_dim * 2**i_layer self._out_feature_strides[stage] = 4 * 2**i_layer num_features = [int(embed_dim * 2**i) for i in range(self.num_layers)] self.num_features = num_features # TODO : need this? # assert weight_init in ('jax', 'jax_nlhb', 'nlhb', '') # head_bias = -math.log(self.num_classes) if 'nlhb' in weight_init else 0. # if weight_init.startswith('jax'): # for n, m in self.named_modules(): # _init_vit_weights(m, n, head_bias=head_bias, jax_impl=True) # else: # self.apply(_init_vit_weights) # add a norm layer for each output for i_layer in range(self.num_layers): stage = f"stage{i_layer + 2}" if stage in self.out_features: if i_layer == 0 and backbone_arch.endswith("RETINANET"): layer = nn.Identity() else: layer = norm_layer(num_features[i_layer]) layer_name = f"norm{i_layer}" self.add_module(layer_name, layer) self._freeze_stages() def _freeze_stages(self): if self.frozen_stages >= 0: self.patch_embed.eval() for param in self.patch_embed.parameters(): param.requires_grad = False if self.frozen_stages >= 1 and self.ape: self.absolute_pos_embed.requires_grad = False if self.frozen_stages >= 2: self.pos_drop.eval() for i in range(0, self.frozen_stages - 1): m = self.layers[i] m.eval() for param in m.parameters(): param.requires_grad = False def init_weights(self, pretrained=None): """Initialize the weights in backbone. Args: pretrained (str, optional): Path to pre-trained weights. Defaults to None. """ def _init_weights(m): if isinstance(m, nn.Linear): trunc_normal_(m.weight, std=0.02) if isinstance(m, nn.Linear) and m.bias is not None: nn.init.constant_(m.bias, 0) elif isinstance(m, nn.LayerNorm): nn.init.constant_(m.bias, 0) nn.init.constant_(m.weight, 1.0) self.apply(_init_weights) def forward(self, inputs): """Forward function.""" x = inputs["img"] language_dict_features = inputs["lang"] x = self.patch_embed(x) Wh, Ww = x.size(2), x.size(3) if self.ape: # interpolate the position embedding to the corresponding size absolute_pos_embed = F.interpolate(self.absolute_pos_embed, size=(Wh, Ww), mode="bicubic") x = (x + absolute_pos_embed).flatten(2).transpose(1, 2) # B Wh*Ww C else: x = x.flatten(2).transpose(1, 2) x = self.pos_drop(x) x_text = language_dict_features["hidden"] if "masks" in language_dict_features: mask_text = 1.0 - language_dict_features["masks"] # (B, N_text) 0 means not to be masked out mask_text.masked_fill_(mask_text.bool(), -float("inf")) else: mask_text = None outs = [] for layer_i, layer in enumerate(self.layers): # if layer_i > 1: # if layer_i > 2: if layer_i > -1: x_out, H, W, x, Wh, Ww = layer(x, Wh, Ww, x_text=x_text, mask_text=mask_text) else: x_out, H, W, x, Wh, Ww = layer(x, Wh, Ww, x_text=None, mask_text=None) name = f"stage{layer_i + 2}" if name in self.out_features: norm_layer = getattr(self, f"norm{layer_i}") x_out = norm_layer(x_out) out = x_out.view(-1, H, W, self.num_features[layer_i]).permute(0, 3, 1, 2).contiguous() outs.append(out) # Here the text features are just combined directly with the image features, so language_dict_features is unchanged return outs, language_dict_features @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 train(self, mode=True): """Convert the model into training mode while keep layers freezed.""" super(SwinTransformer, self).train(mode) self._freeze_stages() class FusionSwinTransformer(nn.Module): def __init__(self, vision_backbone, language_backbone, add_linear_layer=False): super().__init__() self.backbone = vision_backbone self.language_backbone = language_backbone # self.cross_modal_image_transform2 = nn.Linear(1024, 768) # self.cross_modal_image_transform3 = nn.Linear(1024, 768) self.add_linear_layer = add_linear_layer if self.add_linear_layer: self.tunable_linear = torch.nn.Linear( self.language_backbone.body.cfg.MODEL.LANGUAGE_BACKBONE.LANG_DIM, 1000, bias=False ) self.tunable_linear.weight.data.fill_(0.0) def forward( self, tokenizer_input, images, ): # Fusion in the backbone forward - interleaves the passed through the langauge and image backbone. x = images.tensors # Embed the image x = self.backbone.body.patch_embed(x) Wh, Ww = x.size(2), x.size(3) if self.backbone.body.ape: # interpolate the position embedding to the corresponding size absolute_pos_embed = F.interpolate(self.backbone.body.absolute_pos_embed, size=(Wh, Ww), mode="bicubic") x = (x + absolute_pos_embed).flatten(2).transpose(1, 2) # B Wh*Ww C else: x = x.flatten(2).transpose(1, 2) image_embeds = self.backbone.body.pos_drop(x) # Embed the text text_embeds = self.language_backbone.body.model.embeddings(input_ids=tokenizer_input["input_ids"]) input_shape = tokenizer_input["attention_mask"].size() extended_text_masks = self.language_backbone.body.model.get_extended_attention_mask( tokenizer_input["attention_mask"], input_shape, device=tokenizer_input["attention_mask"].device ) if self.add_linear_layer: text_embeds = self.tunable_linear.weight[: text_embeds.size(1), :].unsqueeze(0) + text_embeds outs = [] # Pass the text through the first 10 layers num_pre_text = 6 for layer_i, layer in enumerate(self.language_backbone.body.model.encoder.layer[:num_pre_text]): text_embeds = layer(text_embeds, extended_text_masks)[0] # Pass through first 2 image backbone layers num_pre_vision = 2 for layer_i, layer in enumerate(self.backbone.body.layers[:num_pre_vision]): x_out, H, W, image_embeds, Wh, Ww = layer(image_embeds, Wh, Ww, x_text=None, mask_text=None) name = f"stage{layer_i + 2}" if name in self.backbone.body.out_features: norm_layer = getattr(self.backbone.body, f"norm{layer_i}") x_out = norm_layer(x_out) out = x_out.view(-1, H, W, self.backbone.body.num_features[layer_i]).permute(0, 3, 1, 2).contiguous() outs.append(out) num_pre_block = 14 # Get the attention mask for the third layer: attn_mask = self.backbone.body.layers[num_pre_vision].get_attention_mask(Wh, Ww, image_embeds.device) for blk_cnt, blk in enumerate(self.backbone.body.layers[num_pre_vision].blocks): blk.H, blk.W = Wh, Ww if blk_cnt < num_pre_block: if not torch.jit.is_scripting() and self.backbone.body.layers[num_pre_vision].use_checkpoint: image_embeds = checkpoint.checkpoint(blk, image_embeds, attn_mask) else: image_embeds = blk(image_embeds, attn_mask) else: if not torch.jit.is_scripting() and self.backbone.body.layers[num_pre_vision].use_checkpoint: fused_image_embeds = checkpoint.checkpoint( blk, image_embeds, attn_mask, text_embeds, extended_text_masks ) else: fused_image_embeds = blk(image_embeds, attn_mask, text_embeds, extended_text_masks) text_embeds = self.language_backbone.body.model.encoder.layer[blk_cnt - num_pre_block + num_pre_text]( text_embeds, extended_text_masks, encoder_hidden_states=(image_embeds) )[0] image_embeds = fused_image_embeds # Apply layer norm after 3rd layer and take output name = f"stage{num_pre_vision + 2}" if name in self.backbone.body.out_features: norm_layer = getattr(self.backbone.body, f"norm{num_pre_vision}") x_out = norm_layer(image_embeds) out = ( x_out.view(-1, Wh, Ww, self.backbone.body.num_features[num_pre_vision]).permute(0, 3, 1, 2).contiguous() ) outs.append(out) # Apply downsampling if we need to at the output of third layer for input to next layer if self.backbone.body.layers[num_pre_vision].downsample is not None: image_embeds = self.backbone.body.layers[num_pre_vision].downsample(image_embeds, Wh, Ww) Wh, Ww = (Wh + 1) // 2, (Ww + 1) // 2 # Final layer # Get attention mask for 4th layer attn_mask = self.backbone.body.layers[num_pre_vision + 1].get_attention_mask(Wh, Ww, image_embeds.device) blk = self.backbone.body.layers[num_pre_vision + 1].blocks[0] blk.H, blk.W = Wh, Ww fuse_image_embeds = blk( x=image_embeds, mask_matrix=attn_mask, x_text=text_embeds, mask_text=extended_text_masks ) fuse_text_embeds = self.language_backbone.body.model.encoder.layer[-2]( text_embeds, extended_text_masks, encoder_hidden_states=(image_embeds) )[0] text_embeds, image_embeds = fuse_text_embeds, fuse_image_embeds blk = self.backbone.body.layers[num_pre_vision + 1].blocks[1] blk.H, blk.W = Wh, Ww fuse_image_embeds = self.backbone.body.layers[num_pre_vision + 1].blocks[1]( x=image_embeds, mask_matrix=attn_mask, x_text=text_embeds, mask_text=extended_text_masks ) fuse_text_embeds = self.language_backbone.body.model.encoder.layer[-1]( text_embeds, extended_text_masks, encoder_hidden_states=(image_embeds) )[0] text_embeds, image_embeds = fuse_text_embeds, fuse_image_embeds # Apply layer norm after 4th layer and take output name = f"stage{num_pre_vision + 1 + 2}" if name in self.backbone.body.out_features: norm_layer = getattr(self.backbone.body, f"norm{num_pre_vision + 1}") x_out = norm_layer(image_embeds) out = ( x_out.view(-1, Wh, Ww, self.backbone.body.num_features[num_pre_vision + 1]) .permute(0, 3, 1, 2) .contiguous() ) outs.append(out) language_dict_features = self.language_backbone.body.get_aggregated_output( text_embeds, tokenizer_input["input_ids"], tokenizer_input["attention_mask"] ) # Apply fpn visual_features = self.backbone.fpn(outs) # None for now, need to add if we want to add shallow contrastive loss? swint_feature_c4 = None return visual_features, language_dict_features, swint_feature_c4 def build_swint_backbone(cfg): """ Create a SwinT instance from config. Returns: VoVNet: a :class:`VoVNet` instance. """ return SwinTransformer( patch_size=4, in_chans=3, embed_dim=cfg.MODEL.SWINT.EMBED_DIM, depths=cfg.MODEL.SWINT.DEPTHS, num_heads=cfg.MODEL.SWINT.NUM_HEADS, window_size=cfg.MODEL.SWINT.WINDOW_SIZE, mlp_ratio=cfg.MODEL.SWINT.MLP_RATIO, qkv_bias=True, qk_scale=None, drop_rate=0.0, attn_drop_rate=0.0, drop_path_rate=cfg.MODEL.SWINT.DROP_PATH_RATE, norm_layer=nn.LayerNorm, ape=cfg.MODEL.SWINT.APE, patch_norm=True, frozen_stages=cfg.MODEL.BACKBONE.FREEZE_CONV_BODY_AT, backbone_arch=cfg.MODEL.BACKBONE.CONV_BODY, use_checkpoint=cfg.MODEL.BACKBONE.USE_CHECKPOINT, out_features=cfg.MODEL.BACKBONE.OUT_FEATURES, max_query_len=cfg.MODEL.LANGUAGE_BACKBONE.MAX_QUERY_LEN, lang_dim=cfg.MODEL.LANGUAGE_BACKBONE.LANG_DIM, ) def build_combined_backbone(vision_backbone, language_backbone, add_linear_layer=False): return FusionSwinTransformer(vision_backbone, language_backbone, add_linear_layer=add_linear_layer)