BEN2.py
| 1 | |
| 2 | import math |
| 3 | import torch |
| 4 | import torch.nn as nn |
| 5 | import torch.nn.functional as F |
| 6 | from einops import rearrange |
| 7 | import torch.utils.checkpoint as checkpoint |
| 8 | import numpy as np |
| 9 | from timm.models.layers import DropPath, to_2tuple, trunc_normal_ |
| 10 | from PIL import Image, ImageOps |
| 11 | from torchvision import transforms |
| 12 | import numpy as np |
| 13 | import random |
| 14 | import cv2 |
| 15 | import os |
| 16 | import subprocess |
| 17 | import time |
| 18 | import tempfile |
| 19 | |
| 20 | |
| 21 | |
| 22 | |
| 23 | def set_random_seed(seed): |
| 24 | random.seed(seed) |
| 25 | np.random.seed(seed) |
| 26 | torch.manual_seed(seed) |
| 27 | torch.cuda.manual_seed(seed) |
| 28 | torch.cuda.manual_seed_all(seed) |
| 29 | torch.backends.cudnn.deterministic = True |
| 30 | torch.backends.cudnn.benchmark = False |
| 31 | set_random_seed(9) |
| 32 | |
| 33 | |
| 34 | torch.set_float32_matmul_precision('highest') |
| 35 | |
| 36 | |
| 37 | |
| 38 | class Mlp(nn.Module): |
| 39 | """ Multilayer perceptron.""" |
| 40 | |
| 41 | def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.): |
| 42 | super().__init__() |
| 43 | out_features = out_features or in_features |
| 44 | hidden_features = hidden_features or in_features |
| 45 | self.fc1 = nn.Linear(in_features, hidden_features) |
| 46 | self.act = act_layer() |
| 47 | self.fc2 = nn.Linear(hidden_features, out_features) |
| 48 | self.drop = nn.Dropout(drop) |
| 49 | |
| 50 | def forward(self, x): |
| 51 | x = self.fc1(x) |
| 52 | x = self.act(x) |
| 53 | x = self.drop(x) |
| 54 | x = self.fc2(x) |
| 55 | x = self.drop(x) |
| 56 | return x |
| 57 | |
| 58 | |
| 59 | def window_partition(x, window_size): |
| 60 | """ |
| 61 | Args: |
| 62 | x: (B, H, W, C) |
| 63 | window_size (int): window size |
| 64 | Returns: |
| 65 | windows: (num_windows*B, window_size, window_size, C) |
| 66 | """ |
| 67 | B, H, W, C = x.shape |
| 68 | x = x.view(B, H // window_size, window_size, W // window_size, window_size, C) |
| 69 | windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C) |
| 70 | return windows |
| 71 | |
| 72 | |
| 73 | def window_reverse(windows, window_size, H, W): |
| 74 | """ |
| 75 | Args: |
| 76 | windows: (num_windows*B, window_size, window_size, C) |
| 77 | window_size (int): Window size |
| 78 | H (int): Height of image |
| 79 | W (int): Width of image |
| 80 | Returns: |
| 81 | x: (B, H, W, C) |
| 82 | """ |
| 83 | B = int(windows.shape[0] / (H * W / window_size / window_size)) |
| 84 | x = windows.view(B, H // window_size, W // window_size, window_size, window_size, -1) |
| 85 | x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, -1) |
| 86 | return x |
| 87 | |
| 88 | |
| 89 | class WindowAttention(nn.Module): |
| 90 | """ Window based multi-head self attention (W-MSA) module with relative position bias. |
| 91 | It supports both of shifted and non-shifted window. |
| 92 | Args: |
| 93 | dim (int): Number of input channels. |
| 94 | window_size (tuple[int]): The height and width of the window. |
| 95 | num_heads (int): Number of attention heads. |
| 96 | qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True |
| 97 | qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set |
| 98 | attn_drop (float, optional): Dropout ratio of attention weight. Default: 0.0 |
| 99 | proj_drop (float, optional): Dropout ratio of output. Default: 0.0 |
| 100 | """ |
| 101 | |
| 102 | def __init__(self, dim, window_size, num_heads, qkv_bias=True, qk_scale=None, attn_drop=0., proj_drop=0.): |
| 103 | |
| 104 | super().__init__() |
| 105 | self.dim = dim |
| 106 | self.window_size = window_size # Wh, Ww |
| 107 | self.num_heads = num_heads |
| 108 | head_dim = dim // num_heads |
| 109 | self.scale = qk_scale or head_dim ** -0.5 |
| 110 | |
| 111 | # define a parameter table of relative position bias |
| 112 | self.relative_position_bias_table = nn.Parameter( |
| 113 | torch.zeros((2 * window_size[0] - 1) * (2 * window_size[1] - 1), num_heads)) # 2*Wh-1 * 2*Ww-1, nH |
| 114 | |
| 115 | # get pair-wise relative position index for each token inside the window |
| 116 | coords_h = torch.arange(self.window_size[0]) |
| 117 | coords_w = torch.arange(self.window_size[1]) |
| 118 | coords = torch.stack(torch.meshgrid([coords_h, coords_w])) # 2, Wh, Ww |
| 119 | coords_flatten = torch.flatten(coords, 1) # 2, Wh*Ww |
| 120 | relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :] # 2, Wh*Ww, Wh*Ww |
| 121 | relative_coords = relative_coords.permute(1, 2, 0).contiguous() # Wh*Ww, Wh*Ww, 2 |
| 122 | relative_coords[:, :, 0] += self.window_size[0] - 1 # shift to start from 0 |
| 123 | relative_coords[:, :, 1] += self.window_size[1] - 1 |
| 124 | relative_coords[:, :, 0] *= 2 * self.window_size[1] - 1 |
| 125 | relative_position_index = relative_coords.sum(-1) # Wh*Ww, Wh*Ww |
| 126 | self.register_buffer("relative_position_index", relative_position_index) |
| 127 | |
| 128 | self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias) |
| 129 | self.attn_drop = nn.Dropout(attn_drop) |
| 130 | self.proj = nn.Linear(dim, dim) |
| 131 | self.proj_drop = nn.Dropout(proj_drop) |
| 132 | |
| 133 | trunc_normal_(self.relative_position_bias_table, std=.02) |
| 134 | self.softmax = nn.Softmax(dim=-1) |
| 135 | |
| 136 | def forward(self, x, mask=None): |
| 137 | """ Forward function. |
| 138 | Args: |
| 139 | x: input features with shape of (num_windows*B, N, C) |
| 140 | mask: (0/-inf) mask with shape of (num_windows, Wh*Ww, Wh*Ww) or None |
| 141 | """ |
| 142 | B_, N, C = x.shape |
| 143 | qkv = self.qkv(x).reshape(B_, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4) |
| 144 | q, k, v = qkv[0], qkv[1], qkv[2] # make torchscript happy (cannot use tensor as tuple) |
| 145 | |
| 146 | q = q * self.scale |
| 147 | attn = (q @ k.transpose(-2, -1)) |
| 148 | |
| 149 | relative_position_bias = self.relative_position_bias_table[self.relative_position_index.view(-1)].view( |
| 150 | self.window_size[0] * self.window_size[1], self.window_size[0] * self.window_size[1], -1) # Wh*Ww,Wh*Ww,nH |
| 151 | relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous() # nH, Wh*Ww, Wh*Ww |
| 152 | attn = attn + relative_position_bias.unsqueeze(0) |
| 153 | |
| 154 | if mask is not None: |
| 155 | nW = mask.shape[0] |
| 156 | attn = attn.view(B_ // nW, nW, self.num_heads, N, N) + mask.unsqueeze(1).unsqueeze(0) |
| 157 | attn = attn.view(-1, self.num_heads, N, N) |
| 158 | attn = self.softmax(attn) |
| 159 | else: |
| 160 | attn = self.softmax(attn) |
| 161 | |
| 162 | attn = self.attn_drop(attn) |
| 163 | |
| 164 | x = (attn @ v).transpose(1, 2).reshape(B_, N, C) |
| 165 | x = self.proj(x) |
| 166 | x = self.proj_drop(x) |
| 167 | return x |
| 168 | |
| 169 | |
| 170 | class SwinTransformerBlock(nn.Module): |
| 171 | """ Swin Transformer Block. |
| 172 | Args: |
| 173 | dim (int): Number of input channels. |
| 174 | num_heads (int): Number of attention heads. |
| 175 | window_size (int): Window size. |
| 176 | shift_size (int): Shift size for SW-MSA. |
| 177 | mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. |
| 178 | qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True |
| 179 | qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set. |
| 180 | drop (float, optional): Dropout rate. Default: 0.0 |
| 181 | attn_drop (float, optional): Attention dropout rate. Default: 0.0 |
| 182 | drop_path (float, optional): Stochastic depth rate. Default: 0.0 |
| 183 | act_layer (nn.Module, optional): Activation layer. Default: nn.GELU |
| 184 | norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm |
| 185 | """ |
| 186 | |
| 187 | def __init__(self, dim, num_heads, window_size=7, shift_size=0, |
| 188 | mlp_ratio=4., qkv_bias=True, qk_scale=None, drop=0., attn_drop=0., drop_path=0., |
| 189 | act_layer=nn.GELU, norm_layer=nn.LayerNorm): |
| 190 | super().__init__() |
| 191 | self.dim = dim |
| 192 | self.num_heads = num_heads |
| 193 | self.window_size = window_size |
| 194 | self.shift_size = shift_size |
| 195 | self.mlp_ratio = mlp_ratio |
| 196 | assert 0 <= self.shift_size < self.window_size, "shift_size must in 0-window_size" |
| 197 | |
| 198 | self.norm1 = norm_layer(dim) |
| 199 | self.attn = WindowAttention( |
| 200 | dim, window_size=to_2tuple(self.window_size), num_heads=num_heads, |
| 201 | qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop) |
| 202 | |
| 203 | self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity() |
| 204 | self.norm2 = norm_layer(dim) |
| 205 | mlp_hidden_dim = int(dim * mlp_ratio) |
| 206 | self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop) |
| 207 | |
| 208 | self.H = None |
| 209 | self.W = None |
| 210 | |
| 211 | def forward(self, x, mask_matrix): |
| 212 | """ Forward function. |
| 213 | Args: |
| 214 | x: Input feature, tensor size (B, H*W, C). |
| 215 | H, W: Spatial resolution of the input feature. |
| 216 | mask_matrix: Attention mask for cyclic shift. |
| 217 | """ |
| 218 | B, L, C = x.shape |
| 219 | H, W = self.H, self.W |
| 220 | assert L == H * W, "input feature has wrong size" |
| 221 | |
| 222 | shortcut = x |
| 223 | x = self.norm1(x) |
| 224 | x = x.view(B, H, W, C) |
| 225 | |
| 226 | # pad feature maps to multiples of window size |
| 227 | pad_l = pad_t = 0 |
| 228 | pad_r = (self.window_size - W % self.window_size) % self.window_size |
| 229 | pad_b = (self.window_size - H % self.window_size) % self.window_size |
| 230 | x = F.pad(x, (0, 0, pad_l, pad_r, pad_t, pad_b)) |
| 231 | _, Hp, Wp, _ = x.shape |
| 232 | |
| 233 | # cyclic shift |
| 234 | if self.shift_size > 0: |
| 235 | shifted_x = torch.roll(x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2)) |
| 236 | attn_mask = mask_matrix |
| 237 | else: |
| 238 | shifted_x = x |
| 239 | attn_mask = None |
| 240 | |
| 241 | # partition windows |
| 242 | x_windows = window_partition(shifted_x, self.window_size) # nW*B, window_size, window_size, C |
| 243 | x_windows = x_windows.view(-1, self.window_size * self.window_size, C) # nW*B, window_size*window_size, C |
| 244 | |
| 245 | # W-MSA/SW-MSA |
| 246 | attn_windows = self.attn(x_windows, mask=attn_mask) # nW*B, window_size*window_size, C |
| 247 | |
| 248 | # merge windows |
| 249 | attn_windows = attn_windows.view(-1, self.window_size, self.window_size, C) |
| 250 | shifted_x = window_reverse(attn_windows, self.window_size, Hp, Wp) # B H' W' C |
| 251 | |
| 252 | # reverse cyclic shift |
| 253 | if self.shift_size > 0: |
| 254 | x = torch.roll(shifted_x, shifts=(self.shift_size, self.shift_size), dims=(1, 2)) |
| 255 | else: |
| 256 | x = shifted_x |
| 257 | |
| 258 | if pad_r > 0 or pad_b > 0: |
| 259 | x = x[:, :H, :W, :].contiguous() |
| 260 | |
| 261 | x = x.view(B, H * W, C) |
| 262 | |
| 263 | # FFN |
| 264 | x = shortcut + self.drop_path(x) |
| 265 | x = x + self.drop_path(self.mlp(self.norm2(x))) |
| 266 | |
| 267 | return x |
| 268 | |
| 269 | |
| 270 | class PatchMerging(nn.Module): |
| 271 | """ Patch Merging Layer |
| 272 | Args: |
| 273 | dim (int): Number of input channels. |
| 274 | norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm |
| 275 | """ |
| 276 | def __init__(self, dim, norm_layer=nn.LayerNorm): |
| 277 | super().__init__() |
| 278 | self.dim = dim |
| 279 | self.reduction = nn.Linear(4 * dim, 2 * dim, bias=False) |
| 280 | self.norm = norm_layer(4 * dim) |
| 281 | |
| 282 | def forward(self, x, H, W): |
| 283 | """ Forward function. |
| 284 | Args: |
| 285 | x: Input feature, tensor size (B, H*W, C). |
| 286 | H, W: Spatial resolution of the input feature. |
| 287 | """ |
| 288 | B, L, C = x.shape |
| 289 | assert L == H * W, "input feature has wrong size" |
| 290 | |
| 291 | x = x.view(B, H, W, C) |
| 292 | |
| 293 | # padding |
| 294 | pad_input = (H % 2 == 1) or (W % 2 == 1) |
| 295 | if pad_input: |
| 296 | x = F.pad(x, (0, 0, 0, W % 2, 0, H % 2)) |
| 297 | |
| 298 | x0 = x[:, 0::2, 0::2, :] # B H/2 W/2 C |
| 299 | x1 = x[:, 1::2, 0::2, :] # B H/2 W/2 C |
| 300 | x2 = x[:, 0::2, 1::2, :] # B H/2 W/2 C |
| 301 | x3 = x[:, 1::2, 1::2, :] # B H/2 W/2 C |
| 302 | x = torch.cat([x0, x1, x2, x3], -1) # B H/2 W/2 4*C |
| 303 | x = x.view(B, -1, 4 * C) # B H/2*W/2 4*C |
| 304 | |
| 305 | x = self.norm(x) |
| 306 | x = self.reduction(x) |
| 307 | |
| 308 | return x |
| 309 | |
| 310 | |
| 311 | class BasicLayer(nn.Module): |
| 312 | """ A basic Swin Transformer layer for one stage. |
| 313 | Args: |
| 314 | dim (int): Number of feature channels |
| 315 | depth (int): Depths of this stage. |
| 316 | num_heads (int): Number of attention head. |
| 317 | window_size (int): Local window size. Default: 7. |
| 318 | mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4. |
| 319 | qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True |
| 320 | qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set. |
| 321 | drop (float, optional): Dropout rate. Default: 0.0 |
| 322 | attn_drop (float, optional): Attention dropout rate. Default: 0.0 |
| 323 | drop_path (float | tuple[float], optional): Stochastic depth rate. Default: 0.0 |
| 324 | norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm |
| 325 | downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default: None |
| 326 | use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False. |
| 327 | """ |
| 328 | |
| 329 | def __init__(self, |
| 330 | dim, |
| 331 | depth, |
| 332 | num_heads, |
| 333 | window_size=7, |
| 334 | mlp_ratio=4., |
| 335 | qkv_bias=True, |
| 336 | qk_scale=None, |
| 337 | drop=0., |
| 338 | attn_drop=0., |
| 339 | drop_path=0., |
| 340 | norm_layer=nn.LayerNorm, |
| 341 | downsample=None, |
| 342 | use_checkpoint=False): |
| 343 | super().__init__() |
| 344 | self.window_size = window_size |
| 345 | self.shift_size = window_size // 2 |
| 346 | self.depth = depth |
| 347 | self.use_checkpoint = use_checkpoint |
| 348 | |
| 349 | # build blocks |
| 350 | self.blocks = nn.ModuleList([ |
| 351 | SwinTransformerBlock( |
| 352 | dim=dim, |
| 353 | num_heads=num_heads, |
| 354 | window_size=window_size, |
| 355 | shift_size=0 if (i % 2 == 0) else window_size // 2, |
| 356 | mlp_ratio=mlp_ratio, |
| 357 | qkv_bias=qkv_bias, |
| 358 | qk_scale=qk_scale, |
| 359 | drop=drop, |
| 360 | attn_drop=attn_drop, |
| 361 | drop_path=drop_path[i] if isinstance(drop_path, list) else drop_path, |
| 362 | norm_layer=norm_layer) |
| 363 | for i in range(depth)]) |
| 364 | |
| 365 | # patch merging layer |
| 366 | if downsample is not None: |
| 367 | self.downsample = downsample(dim=dim, norm_layer=norm_layer) |
| 368 | else: |
| 369 | self.downsample = None |
| 370 | |
| 371 | def forward(self, x, H, W): |
| 372 | """ Forward function. |
| 373 | Args: |
| 374 | x: Input feature, tensor size (B, H*W, C). |
| 375 | H, W: Spatial resolution of the input feature. |
| 376 | """ |
| 377 | |
| 378 | # calculate attention mask for SW-MSA |
| 379 | Hp = int(np.ceil(H / self.window_size)) * self.window_size |
| 380 | Wp = int(np.ceil(W / self.window_size)) * self.window_size |
| 381 | img_mask = torch.zeros((1, Hp, Wp, 1), device=x.device) # 1 Hp Wp 1 |
| 382 | h_slices = (slice(0, -self.window_size), |
| 383 | slice(-self.window_size, -self.shift_size), |
| 384 | slice(-self.shift_size, None)) |
| 385 | w_slices = (slice(0, -self.window_size), |
| 386 | slice(-self.window_size, -self.shift_size), |
| 387 | slice(-self.shift_size, None)) |
| 388 | cnt = 0 |
| 389 | for h in h_slices: |
| 390 | for w in w_slices: |
| 391 | img_mask[:, h, w, :] = cnt |
| 392 | cnt += 1 |
| 393 | |
| 394 | mask_windows = window_partition(img_mask, self.window_size) # nW, window_size, window_size, 1 |
| 395 | mask_windows = mask_windows.view(-1, self.window_size * self.window_size) |
| 396 | attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2) |
| 397 | attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(attn_mask == 0, float(0.0)) |
| 398 | |
| 399 | for blk in self.blocks: |
| 400 | blk.H, blk.W = H, W |
| 401 | if self.use_checkpoint: |
| 402 | x = checkpoint.checkpoint(blk, x, attn_mask) |
| 403 | else: |
| 404 | x = blk(x, attn_mask) |
| 405 | if self.downsample is not None: |
| 406 | x_down = self.downsample(x, H, W) |
| 407 | Wh, Ww = (H + 1) // 2, (W + 1) // 2 |
| 408 | return x, H, W, x_down, Wh, Ww |
| 409 | else: |
| 410 | return x, H, W, x, H, W |
| 411 | |
| 412 | |
| 413 | class PatchEmbed(nn.Module): |
| 414 | """ Image to Patch Embedding |
| 415 | Args: |
| 416 | patch_size (int): Patch token size. Default: 4. |
| 417 | in_chans (int): Number of input image channels. Default: 3. |
| 418 | embed_dim (int): Number of linear projection output channels. Default: 96. |
| 419 | norm_layer (nn.Module, optional): Normalization layer. Default: None |
| 420 | """ |
| 421 | |
| 422 | def __init__(self, patch_size=4, in_chans=3, embed_dim=96, norm_layer=None): |
| 423 | super().__init__() |
| 424 | patch_size = to_2tuple(patch_size) |
| 425 | self.patch_size = patch_size |
| 426 | |
| 427 | self.in_chans = in_chans |
| 428 | self.embed_dim = embed_dim |
| 429 | |
| 430 | self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size) |
| 431 | if norm_layer is not None: |
| 432 | self.norm = norm_layer(embed_dim) |
| 433 | else: |
| 434 | self.norm = None |
| 435 | |
| 436 | def forward(self, x): |
| 437 | """Forward function.""" |
| 438 | # padding |
| 439 | _, _, H, W = x.size() |
| 440 | if W % self.patch_size[1] != 0: |
| 441 | x = F.pad(x, (0, self.patch_size[1] - W % self.patch_size[1])) |
| 442 | if H % self.patch_size[0] != 0: |
| 443 | x = F.pad(x, (0, 0, 0, self.patch_size[0] - H % self.patch_size[0])) |
| 444 | |
| 445 | x = self.proj(x) # B C Wh Ww |
| 446 | if self.norm is not None: |
| 447 | Wh, Ww = x.size(2), x.size(3) |
| 448 | x = x.flatten(2).transpose(1, 2) |
| 449 | x = self.norm(x) |
| 450 | x = x.transpose(1, 2).view(-1, self.embed_dim, Wh, Ww) |
| 451 | |
| 452 | return x |
| 453 | |
| 454 | |
| 455 | class SwinTransformer(nn.Module): |
| 456 | """ Swin Transformer backbone. |
| 457 | A PyTorch impl of : `Swin Transformer: Hierarchical Vision Transformer using Shifted Windows` - |
| 458 | https://arxiv.org/pdf/2103.14030 |
| 459 | Args: |
| 460 | pretrain_img_size (int): Input image size for training the pretrained model, |
| 461 | used in absolute postion embedding. Default 224. |
| 462 | patch_size (int | tuple(int)): Patch size. Default: 4. |
| 463 | in_chans (int): Number of input image channels. Default: 3. |
| 464 | embed_dim (int): Number of linear projection output channels. Default: 96. |
| 465 | depths (tuple[int]): Depths of each Swin Transformer stage. |
| 466 | num_heads (tuple[int]): Number of attention head of each stage. |
| 467 | window_size (int): Window size. Default: 7. |
| 468 | mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4. |
| 469 | qkv_bias (bool): If True, add a learnable bias to query, key, value. Default: True |
| 470 | qk_scale (float): Override default qk scale of head_dim ** -0.5 if set. |
| 471 | drop_rate (float): Dropout rate. |
| 472 | attn_drop_rate (float): Attention dropout rate. Default: 0. |
| 473 | drop_path_rate (float): Stochastic depth rate. Default: 0.2. |
| 474 | norm_layer (nn.Module): Normalization layer. Default: nn.LayerNorm. |
| 475 | ape (bool): If True, add absolute position embedding to the patch embedding. Default: False. |
| 476 | patch_norm (bool): If True, add normalization after patch embedding. Default: True. |
| 477 | out_indices (Sequence[int]): Output from which stages. |
| 478 | frozen_stages (int): Stages to be frozen (stop grad and set eval mode). |
| 479 | -1 means not freezing any parameters. |
| 480 | use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False. |
| 481 | """ |
| 482 | |
| 483 | def __init__(self, |
| 484 | pretrain_img_size=224, |
| 485 | patch_size=4, |
| 486 | in_chans=3, |
| 487 | embed_dim=96, |
| 488 | depths=[2, 2, 6, 2], |
| 489 | num_heads=[3, 6, 12, 24], |
| 490 | window_size=7, |
| 491 | mlp_ratio=4., |
| 492 | qkv_bias=True, |
| 493 | qk_scale=None, |
| 494 | drop_rate=0., |
| 495 | attn_drop_rate=0., |
| 496 | drop_path_rate=0.2, |
| 497 | norm_layer=nn.LayerNorm, |
| 498 | ape=False, |
| 499 | patch_norm=True, |
| 500 | out_indices=(0, 1, 2, 3), |
| 501 | frozen_stages=-1, |
| 502 | use_checkpoint=False): |
| 503 | super().__init__() |
| 504 | |
| 505 | self.pretrain_img_size = pretrain_img_size |
| 506 | self.num_layers = len(depths) |
| 507 | self.embed_dim = embed_dim |
| 508 | self.ape = ape |
| 509 | self.patch_norm = patch_norm |
| 510 | self.out_indices = out_indices |
| 511 | self.frozen_stages = frozen_stages |
| 512 | |
| 513 | # split image into non-overlapping patches |
| 514 | self.patch_embed = PatchEmbed( |
| 515 | patch_size=patch_size, in_chans=in_chans, embed_dim=embed_dim, |
| 516 | norm_layer=norm_layer if self.patch_norm else None) |
| 517 | |
| 518 | # absolute position embedding |
| 519 | if self.ape: |
| 520 | pretrain_img_size = to_2tuple(pretrain_img_size) |
| 521 | patch_size = to_2tuple(patch_size) |
| 522 | patches_resolution = [pretrain_img_size[0] // patch_size[0], pretrain_img_size[1] // patch_size[1]] |
| 523 | |
| 524 | self.absolute_pos_embed = nn.Parameter(torch.zeros(1, embed_dim, patches_resolution[0], patches_resolution[1])) |
| 525 | trunc_normal_(self.absolute_pos_embed, std=.02) |
| 526 | |
| 527 | self.pos_drop = nn.Dropout(p=drop_rate) |
| 528 | |
| 529 | # stochastic depth |
| 530 | dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))] # stochastic depth decay rule |
| 531 | |
| 532 | # build layers |
| 533 | self.layers = nn.ModuleList() |
| 534 | for i_layer in range(self.num_layers): |
| 535 | layer = BasicLayer( |
| 536 | dim=int(embed_dim * 2 ** i_layer), |
| 537 | depth=depths[i_layer], |
| 538 | num_heads=num_heads[i_layer], |
| 539 | window_size=window_size, |
| 540 | mlp_ratio=mlp_ratio, |
| 541 | qkv_bias=qkv_bias, |
| 542 | qk_scale=qk_scale, |
| 543 | drop=drop_rate, |
| 544 | attn_drop=attn_drop_rate, |
| 545 | drop_path=dpr[sum(depths[:i_layer]):sum(depths[:i_layer + 1])], |
| 546 | norm_layer=norm_layer, |
| 547 | downsample=PatchMerging if (i_layer < self.num_layers - 1) else None, |
| 548 | use_checkpoint=use_checkpoint) |
| 549 | self.layers.append(layer) |
| 550 | |
| 551 | num_features = [int(embed_dim * 2 ** i) for i in range(self.num_layers)] |
| 552 | self.num_features = num_features |
| 553 | |
| 554 | # add a norm layer for each output |
| 555 | for i_layer in out_indices: |
| 556 | layer = norm_layer(num_features[i_layer]) |
| 557 | layer_name = f'norm{i_layer}' |
| 558 | self.add_module(layer_name, layer) |
| 559 | |
| 560 | self._freeze_stages() |
| 561 | |
| 562 | def _freeze_stages(self): |
| 563 | if self.frozen_stages >= 0: |
| 564 | self.patch_embed.eval() |
| 565 | for param in self.patch_embed.parameters(): |
| 566 | param.requires_grad = False |
| 567 | |
| 568 | if self.frozen_stages >= 1 and self.ape: |
| 569 | self.absolute_pos_embed.requires_grad = False |
| 570 | |
| 571 | if self.frozen_stages >= 2: |
| 572 | self.pos_drop.eval() |
| 573 | for i in range(0, self.frozen_stages - 1): |
| 574 | m = self.layers[i] |
| 575 | m.eval() |
| 576 | for param in m.parameters(): |
| 577 | param.requires_grad = False |
| 578 | |
| 579 | |
| 580 | def forward(self, x): |
| 581 | |
| 582 | x = self.patch_embed(x) |
| 583 | |
| 584 | Wh, Ww = x.size(2), x.size(3) |
| 585 | if self.ape: |
| 586 | # interpolate the position embedding to the corresponding size |
| 587 | absolute_pos_embed = F.interpolate(self.absolute_pos_embed, size=(Wh, Ww), mode='bicubic') |
| 588 | x = (x + absolute_pos_embed) # B Wh*Ww C |
| 589 | |
| 590 | outs = [x.contiguous()] |
| 591 | x = x.flatten(2).transpose(1, 2) |
| 592 | x = self.pos_drop(x) |
| 593 | |
| 594 | |
| 595 | for i in range(self.num_layers): |
| 596 | layer = self.layers[i] |
| 597 | x_out, H, W, x, Wh, Ww = layer(x, Wh, Ww) |
| 598 | |
| 599 | |
| 600 | if i in self.out_indices: |
| 601 | norm_layer = getattr(self, f'norm{i}') |
| 602 | x_out = norm_layer(x_out) |
| 603 | |
| 604 | out = x_out.view(-1, H, W, self.num_features[i]).permute(0, 3, 1, 2).contiguous() |
| 605 | outs.append(out) |
| 606 | |
| 607 | |
| 608 | |
| 609 | return tuple(outs) |
| 610 | |
| 611 | |
| 612 | |
| 613 | |
| 614 | |
| 615 | |
| 616 | |
| 617 | |
| 618 | def get_activation_fn(activation): |
| 619 | """Return an activation function given a string""" |
| 620 | if activation == "gelu": |
| 621 | return F.gelu |
| 622 | |
| 623 | raise RuntimeError(F"activation should be gelu, not {activation}.") |
| 624 | |
| 625 | |
| 626 | def make_cbr(in_dim, out_dim): |
| 627 | return nn.Sequential(nn.Conv2d(in_dim, out_dim, kernel_size=3, padding=1), nn.InstanceNorm2d(out_dim), nn.GELU()) |
| 628 | |
| 629 | |
| 630 | def make_cbg(in_dim, out_dim): |
| 631 | return nn.Sequential(nn.Conv2d(in_dim, out_dim, kernel_size=3, padding=1), nn.InstanceNorm2d(out_dim), nn.GELU()) |
| 632 | |
| 633 | |
| 634 | def rescale_to(x, scale_factor: float = 2, interpolation='nearest'): |
| 635 | return F.interpolate(x, scale_factor=scale_factor, mode=interpolation) |
| 636 | |
| 637 | |
| 638 | def resize_as(x, y, interpolation='bilinear'): |
| 639 | return F.interpolate(x, size=y.shape[-2:], mode=interpolation) |
| 640 | |
| 641 | |
| 642 | def image2patches(x): |
| 643 | """b c (hg h) (wg w) -> (hg wg b) c h w""" |
| 644 | x = rearrange(x, 'b c (hg h) (wg w) -> (hg wg b) c h w', hg=2, wg=2 ) |
| 645 | return x |
| 646 | |
| 647 | |
| 648 | def patches2image(x): |
| 649 | """(hg wg b) c h w -> b c (hg h) (wg w)""" |
| 650 | x = rearrange(x, '(hg wg b) c h w -> b c (hg h) (wg w)', hg=2, wg=2) |
| 651 | return x |
| 652 | |
| 653 | |
| 654 | |
| 655 | class PositionEmbeddingSine: |
| 656 | def __init__(self, num_pos_feats=64, temperature=10000, normalize=False, scale=None): |
| 657 | super().__init__() |
| 658 | self.num_pos_feats = num_pos_feats |
| 659 | self.temperature = temperature |
| 660 | self.normalize = normalize |
| 661 | if scale is not None and normalize is False: |
| 662 | raise ValueError("normalize should be True if scale is passed") |
| 663 | if scale is None: |
| 664 | scale = 2 * math.pi |
| 665 | self.scale = scale |
| 666 | self.dim_t = torch.arange(0, self.num_pos_feats, dtype=torch.float32) |
| 667 | |
| 668 | def __call__(self, b, h, w): |
| 669 | device = self.dim_t.device |
| 670 | mask = torch.zeros([b, h, w], dtype=torch.bool, device=device) |
| 671 | assert mask is not None |
| 672 | not_mask = ~mask |
| 673 | y_embed = not_mask.cumsum(dim=1, dtype=torch.float32) |
| 674 | x_embed = not_mask.cumsum(dim=2, dtype=torch.float32) |
| 675 | if self.normalize: |
| 676 | eps = 1e-6 |
| 677 | y_embed = (y_embed - 0.5) / (y_embed[:, -1:, :] + eps) * self.scale |
| 678 | x_embed = (x_embed - 0.5) / (x_embed[:, :, -1:] + eps) * self.scale |
| 679 | |
| 680 | dim_t = self.temperature ** (2 * (self.dim_t.to(device) // 2) / self.num_pos_feats) |
| 681 | pos_x = x_embed[:, :, :, None] / dim_t |
| 682 | pos_y = y_embed[:, :, :, None] / dim_t |
| 683 | |
| 684 | pos_x = torch.stack((pos_x[:, :, :, 0::2].sin(), pos_x[:, :, :, 1::2].cos()), dim=4).flatten(3) |
| 685 | pos_y = torch.stack((pos_y[:, :, :, 0::2].sin(), pos_y[:, :, :, 1::2].cos()), dim=4).flatten(3) |
| 686 | |
| 687 | return torch.cat((pos_y, pos_x), dim=3).permute(0, 3, 1, 2) |
| 688 | |
| 689 | |
| 690 | |
| 691 | class PositionEmbeddingSine: |
| 692 | def __init__(self, num_pos_feats=64, temperature=10000, normalize=False, scale=None): |
| 693 | super().__init__() |
| 694 | self.num_pos_feats = num_pos_feats |
| 695 | self.temperature = temperature |
| 696 | self.normalize = normalize |
| 697 | if scale is not None and normalize is False: |
| 698 | raise ValueError("normalize should be True if scale is passed") |
| 699 | if scale is None: |
| 700 | scale = 2 * math.pi |
| 701 | self.scale = scale |
| 702 | self.dim_t = torch.arange(0, self.num_pos_feats, dtype=torch.float32) |
| 703 | |
| 704 | def __call__(self, b, h, w): |
| 705 | device = self.dim_t.device |
| 706 | mask = torch.zeros([b, h, w], dtype=torch.bool, device=device) |
| 707 | assert mask is not None |
| 708 | not_mask = ~mask |
| 709 | y_embed = not_mask.cumsum(dim=1, dtype=torch.float32) |
| 710 | x_embed = not_mask.cumsum(dim=2, dtype=torch.float32) |
| 711 | if self.normalize: |
| 712 | eps = 1e-6 |
| 713 | y_embed = (y_embed - 0.5) / (y_embed[:, -1:, :] + eps) * self.scale |
| 714 | x_embed = (x_embed - 0.5) / (x_embed[:, :, -1:] + eps) * self.scale |
| 715 | |
| 716 | dim_t = self.temperature ** (2 * (self.dim_t.to(device) // 2) / self.num_pos_feats) |
| 717 | pos_x = x_embed[:, :, :, None] / dim_t |
| 718 | pos_y = y_embed[:, :, :, None] / dim_t |
| 719 | |
| 720 | pos_x = torch.stack((pos_x[:, :, :, 0::2].sin(), pos_x[:, :, :, 1::2].cos()), dim=4).flatten(3) |
| 721 | pos_y = torch.stack((pos_y[:, :, :, 0::2].sin(), pos_y[:, :, :, 1::2].cos()), dim=4).flatten(3) |
| 722 | |
| 723 | return torch.cat((pos_y, pos_x), dim=3).permute(0, 3, 1, 2) |
| 724 | |
| 725 | |
| 726 | class MCLM(nn.Module): |
| 727 | def __init__(self, d_model, num_heads, pool_ratios=[1, 4, 8]): |
| 728 | super(MCLM, self).__init__() |
| 729 | self.attention = nn.ModuleList([ |
| 730 | nn.MultiheadAttention(d_model, num_heads, dropout=0.1), |
| 731 | nn.MultiheadAttention(d_model, num_heads, dropout=0.1), |
| 732 | nn.MultiheadAttention(d_model, num_heads, dropout=0.1), |
| 733 | nn.MultiheadAttention(d_model, num_heads, dropout=0.1), |
| 734 | nn.MultiheadAttention(d_model, num_heads, dropout=0.1) |
| 735 | ]) |
| 736 | |
| 737 | self.linear1 = nn.Linear(d_model, d_model * 2) |
| 738 | self.linear2 = nn.Linear(d_model * 2, d_model) |
| 739 | self.linear3 = nn.Linear(d_model, d_model * 2) |
| 740 | self.linear4 = nn.Linear(d_model * 2, d_model) |
| 741 | self.norm1 = nn.LayerNorm(d_model) |
| 742 | self.norm2 = nn.LayerNorm(d_model) |
| 743 | self.dropout = nn.Dropout(0.1) |
| 744 | self.dropout1 = nn.Dropout(0.1) |
| 745 | self.dropout2 = nn.Dropout(0.1) |
| 746 | self.activation = get_activation_fn('gelu') |
| 747 | self.pool_ratios = pool_ratios |
| 748 | self.p_poses = [] |
| 749 | self.g_pos = None |
| 750 | self.positional_encoding = PositionEmbeddingSine(num_pos_feats=d_model // 2, normalize=True) |
| 751 | |
| 752 | def forward(self, l, g): |
| 753 | """ |
| 754 | l: 4,c,h,w |
| 755 | g: 1,c,h,w |
| 756 | """ |
| 757 | self.p_poses = [] |
| 758 | self.g_pos = None |
| 759 | b, c, h, w = l.size() |
| 760 | # 4,c,h,w -> 1,c,2h,2w |
| 761 | concated_locs = rearrange(l, '(hg wg b) c h w -> b c (hg h) (wg w)', hg=2, wg=2) |
| 762 | |
| 763 | pools = [] |
| 764 | for pool_ratio in self.pool_ratios: |
| 765 | # b,c,h,w |
| 766 | tgt_hw = (round(h / pool_ratio), round(w / pool_ratio)) |
| 767 | pool = F.adaptive_avg_pool2d(concated_locs, tgt_hw) |
| 768 | pools.append(rearrange(pool, 'b c h w -> (h w) b c')) |
| 769 | if self.g_pos is None: |
| 770 | pos_emb = self.positional_encoding(pool.shape[0], pool.shape[2], pool.shape[3]) |
| 771 | pos_emb = rearrange(pos_emb, 'b c h w -> (h w) b c') |
| 772 | self.p_poses.append(pos_emb) |
| 773 | pools = torch.cat(pools, 0) |
| 774 | if self.g_pos is None: |
| 775 | self.p_poses = torch.cat(self.p_poses, dim=0) |
| 776 | pos_emb = self.positional_encoding(g.shape[0], g.shape[2], g.shape[3]) |
| 777 | self.g_pos = rearrange(pos_emb, 'b c h w -> (h w) b c') |
| 778 | |
| 779 | device = pools.device |
| 780 | self.p_poses = self.p_poses.to(device) |
| 781 | self.g_pos = self.g_pos.to(device) |
| 782 | |
| 783 | |
| 784 | # attention between glb (q) & multisensory concated-locs (k,v) |
| 785 | g_hw_b_c = rearrange(g, 'b c h w -> (h w) b c') |
| 786 | |
| 787 | |
| 788 | g_hw_b_c = g_hw_b_c + self.dropout1(self.attention[0](g_hw_b_c + self.g_pos, pools + self.p_poses, pools)[0]) |
| 789 | g_hw_b_c = self.norm1(g_hw_b_c) |
| 790 | g_hw_b_c = g_hw_b_c + self.dropout2(self.linear2(self.dropout(self.activation(self.linear1(g_hw_b_c)).clone()))) |
| 791 | g_hw_b_c = self.norm2(g_hw_b_c) |
| 792 | |
| 793 | # attention between origin locs (q) & freashed glb (k,v) |
| 794 | l_hw_b_c = rearrange(l, "b c h w -> (h w) b c") |
| 795 | _g_hw_b_c = rearrange(g_hw_b_c, '(h w) b c -> h w b c', h=h, w=w) |
| 796 | _g_hw_b_c = rearrange(_g_hw_b_c, "(ng h) (nw w) b c -> (h w) (ng nw b) c", ng=2, nw=2) |
| 797 | outputs_re = [] |
| 798 | for i, (_l, _g) in enumerate(zip(l_hw_b_c.chunk(4, dim=1), _g_hw_b_c.chunk(4, dim=1))): |
| 799 | outputs_re.append(self.attention[i + 1](_l, _g, _g)[0]) # (h w) 1 c |
| 800 | outputs_re = torch.cat(outputs_re, 1) # (h w) 4 c |
| 801 | |
| 802 | l_hw_b_c = l_hw_b_c + self.dropout1(outputs_re) |
| 803 | l_hw_b_c = self.norm1(l_hw_b_c) |
| 804 | l_hw_b_c = l_hw_b_c + self.dropout2(self.linear4(self.dropout(self.activation(self.linear3(l_hw_b_c)).clone()))) |
| 805 | l_hw_b_c = self.norm2(l_hw_b_c) |
| 806 | |
| 807 | l = torch.cat((l_hw_b_c, g_hw_b_c), 1) # hw,b(5),c |
| 808 | return rearrange(l, "(h w) b c -> b c h w", h=h, w=w) ## (5,c,h*w) |
| 809 | |
| 810 | |
| 811 | |
| 812 | |
| 813 | |
| 814 | |
| 815 | |
| 816 | |
| 817 | |
| 818 | class MCRM(nn.Module): |
| 819 | def __init__(self, d_model, num_heads, pool_ratios=[4, 8, 16], h=None): |
| 820 | super(MCRM, self).__init__() |
| 821 | self.attention = nn.ModuleList([ |
| 822 | nn.MultiheadAttention(d_model, num_heads, dropout=0.1), |
| 823 | nn.MultiheadAttention(d_model, num_heads, dropout=0.1), |
| 824 | nn.MultiheadAttention(d_model, num_heads, dropout=0.1), |
| 825 | nn.MultiheadAttention(d_model, num_heads, dropout=0.1) |
| 826 | ]) |
| 827 | self.linear3 = nn.Linear(d_model, d_model * 2) |
| 828 | self.linear4 = nn.Linear(d_model * 2, d_model) |
| 829 | self.norm1 = nn.LayerNorm(d_model) |
| 830 | self.norm2 = nn.LayerNorm(d_model) |
| 831 | self.dropout = nn.Dropout(0.1) |
| 832 | self.dropout1 = nn.Dropout(0.1) |
| 833 | self.dropout2 = nn.Dropout(0.1) |
| 834 | self.sigmoid = nn.Sigmoid() |
| 835 | self.activation = get_activation_fn('gelu') |
| 836 | self.sal_conv = nn.Conv2d(d_model, 1, 1) |
| 837 | self.pool_ratios = pool_ratios |
| 838 | |
| 839 | def forward(self, x): |
| 840 | device = x.device |
| 841 | b, c, h, w = x.size() |
| 842 | loc, glb = x.split([4, 1], dim=0) # 4,c,h,w; 1,c,h,w |
| 843 | |
| 844 | patched_glb = rearrange(glb, 'b c (hg h) (wg w) -> (hg wg b) c h w', hg=2, wg=2) |
| 845 | |
| 846 | token_attention_map = self.sigmoid(self.sal_conv(glb)) |
| 847 | token_attention_map = F.interpolate(token_attention_map, size=patches2image(loc).shape[-2:], mode='nearest') |
| 848 | loc = loc * rearrange(token_attention_map, 'b c (hg h) (wg w) -> (hg wg b) c h w', hg=2, wg=2) |
| 849 | |
| 850 | pools = [] |
| 851 | for pool_ratio in self.pool_ratios: |
| 852 | tgt_hw = (round(h / pool_ratio), round(w / pool_ratio)) |
| 853 | pool = F.adaptive_avg_pool2d(patched_glb, tgt_hw) |
| 854 | pools.append(rearrange(pool, 'nl c h w -> nl c (h w)')) # nl(4),c,hw |
| 855 | |
| 856 | pools = rearrange(torch.cat(pools, 2), "nl c nphw -> nl nphw 1 c") |
| 857 | loc_ = rearrange(loc, 'nl c h w -> nl (h w) 1 c') |
| 858 | |
| 859 | outputs = [] |
| 860 | for i, q in enumerate(loc_.unbind(dim=0)): # traverse all local patches |
| 861 | v = pools[i] |
| 862 | k = v |
| 863 | outputs.append(self.attention[i](q, k, v)[0]) |
| 864 | |
| 865 | outputs = torch.cat(outputs, 1) |
| 866 | src = loc.view(4, c, -1).permute(2, 0, 1) + self.dropout1(outputs) |
| 867 | src = self.norm1(src) |
| 868 | src = src + self.dropout2(self.linear4(self.dropout(self.activation(self.linear3(src)).clone()))) |
| 869 | src = self.norm2(src) |
| 870 | src = src.permute(1, 2, 0).reshape(4, c, h, w) # freshed loc |
| 871 | glb = glb + F.interpolate(patches2image(src), size=glb.shape[-2:], mode='nearest') # freshed glb |
| 872 | |
| 873 | return torch.cat((src, glb), 0), token_attention_map |
| 874 | |
| 875 | |
| 876 | |
| 877 | class BEN_Base(nn.Module): |
| 878 | def __init__(self): |
| 879 | super().__init__() |
| 880 | |
| 881 | self.backbone = SwinTransformer(embed_dim=128, depths=[2, 2, 18, 2], num_heads=[4, 8, 16, 32], window_size=12) |
| 882 | emb_dim = 128 |
| 883 | self.sideout5 = nn.Sequential(nn.Conv2d(emb_dim, 1, kernel_size=3, padding=1)) |
| 884 | self.sideout4 = nn.Sequential(nn.Conv2d(emb_dim, 1, kernel_size=3, padding=1)) |
| 885 | self.sideout3 = nn.Sequential(nn.Conv2d(emb_dim, 1, kernel_size=3, padding=1)) |
| 886 | self.sideout2 = nn.Sequential(nn.Conv2d(emb_dim, 1, kernel_size=3, padding=1)) |
| 887 | self.sideout1 = nn.Sequential(nn.Conv2d(emb_dim, 1, kernel_size=3, padding=1)) |
| 888 | |
| 889 | self.output5 = make_cbr(1024, emb_dim) |
| 890 | self.output4 = make_cbr(512, emb_dim) |
| 891 | self.output3 = make_cbr(256, emb_dim) |
| 892 | self.output2 = make_cbr(128, emb_dim) |
| 893 | self.output1 = make_cbr(128, emb_dim) |
| 894 | |
| 895 | self.multifieldcrossatt = MCLM(emb_dim, 1, [1, 4, 8]) |
| 896 | self.conv1 = make_cbr(emb_dim, emb_dim) |
| 897 | self.conv2 = make_cbr(emb_dim, emb_dim) |
| 898 | self.conv3 = make_cbr(emb_dim, emb_dim) |
| 899 | self.conv4 = make_cbr(emb_dim, emb_dim) |
| 900 | self.dec_blk1 = MCRM(emb_dim, 1, [2, 4, 8]) |
| 901 | self.dec_blk2 = MCRM(emb_dim, 1, [2, 4, 8]) |
| 902 | self.dec_blk3 = MCRM(emb_dim, 1, [2, 4, 8]) |
| 903 | self.dec_blk4 = MCRM(emb_dim, 1, [2, 4, 8]) |
| 904 | |
| 905 | self.insmask_head = nn.Sequential( |
| 906 | nn.Conv2d(emb_dim, 384, kernel_size=3, padding=1), |
| 907 | nn.InstanceNorm2d(384), |
| 908 | nn.GELU(), |
| 909 | nn.Conv2d(384, 384, kernel_size=3, padding=1), |
| 910 | nn.InstanceNorm2d(384), |
| 911 | nn.GELU(), |
| 912 | nn.Conv2d(384, emb_dim, kernel_size=3, padding=1) |
| 913 | ) |
| 914 | |
| 915 | self.shallow = nn.Sequential(nn.Conv2d(3, emb_dim, kernel_size=3, padding=1)) |
| 916 | self.upsample1 = make_cbg(emb_dim, emb_dim) |
| 917 | self.upsample2 = make_cbg(emb_dim, emb_dim) |
| 918 | self.output = nn.Sequential(nn.Conv2d(emb_dim, 1, kernel_size=3, padding=1)) |
| 919 | |
| 920 | for m in self.modules(): |
| 921 | if isinstance(m, nn.GELU) or isinstance(m, nn.Dropout): |
| 922 | m.inplace = True |
| 923 | |
| 924 | |
| 925 | |
| 926 | @torch.inference_mode() |
| 927 | @torch.autocast(device_type="cuda",dtype=torch.float16) |
| 928 | def forward(self, x): |
| 929 | real_batch = x.size(0) |
| 930 | |
| 931 | shallow_batch = self.shallow(x) |
| 932 | glb_batch = rescale_to(x, scale_factor=0.5, interpolation='bilinear') |
| 933 | |
| 934 | |
| 935 | |
| 936 | final_input = None |
| 937 | for i in range(real_batch): |
| 938 | start = i * 4 |
| 939 | end = (i + 1) * 4 |
| 940 | loc_batch = image2patches(x[i,:,:,:].unsqueeze(dim=0)) |
| 941 | input_ = torch.cat((loc_batch, glb_batch[i,:,:,:].unsqueeze(dim=0)), dim=0) |
| 942 | |
| 943 | |
| 944 | if final_input == None: |
| 945 | final_input= input_ |
| 946 | else: final_input = torch.cat((final_input, input_), dim=0) |
| 947 | |
| 948 | features = self.backbone(final_input) |
| 949 | outputs = [] |
| 950 | |
| 951 | for i in range(real_batch): |
| 952 | |
| 953 | start = i * 5 |
| 954 | end = (i + 1) * 5 |
| 955 | |
| 956 | f4 = features[4][start:end, :, :, :] # shape: [5, C, H, W] |
| 957 | f3 = features[3][start:end, :, :, :] |
| 958 | f2 = features[2][start:end, :, :, :] |
| 959 | f1 = features[1][start:end, :, :, :] |
| 960 | f0 = features[0][start:end, :, :, :] |
| 961 | e5 = self.output5(f4) |
| 962 | e4 = self.output4(f3) |
| 963 | e3 = self.output3(f2) |
| 964 | e2 = self.output2(f1) |
| 965 | e1 = self.output1(f0) |
| 966 | loc_e5, glb_e5 = e5.split([4, 1], dim=0) |
| 967 | e5 = self.multifieldcrossatt(loc_e5, glb_e5) # (4,128,16,16) |
| 968 | |
| 969 | |
| 970 | e4, tokenattmap4 = self.dec_blk4(e4 + resize_as(e5, e4)) |
| 971 | e4 = self.conv4(e4) |
| 972 | e3, tokenattmap3 = self.dec_blk3(e3 + resize_as(e4, e3)) |
| 973 | e3 = self.conv3(e3) |
| 974 | e2, tokenattmap2 = self.dec_blk2(e2 + resize_as(e3, e2)) |
| 975 | e2 = self.conv2(e2) |
| 976 | e1, tokenattmap1 = self.dec_blk1(e1 + resize_as(e2, e1)) |
| 977 | e1 = self.conv1(e1) |
| 978 | |
| 979 | loc_e1, glb_e1 = e1.split([4, 1], dim=0) |
| 980 | |
| 981 | output1_cat = patches2image(loc_e1) # (1,128,256,256) |
| 982 | |
| 983 | # add glb feat in |
| 984 | output1_cat = output1_cat + resize_as(glb_e1, output1_cat) |
| 985 | # merge |
| 986 | final_output = self.insmask_head(output1_cat) # (1,128,256,256) |
| 987 | # shallow feature merge |
| 988 | shallow = shallow_batch[i,:,:,:].unsqueeze(dim=0) |
| 989 | final_output = final_output + resize_as(shallow, final_output) |
| 990 | final_output = self.upsample1(rescale_to(final_output)) |
| 991 | final_output = rescale_to(final_output + resize_as(shallow, final_output)) |
| 992 | final_output = self.upsample2(final_output) |
| 993 | final_output = self.output(final_output) |
| 994 | mask = final_output.sigmoid() |
| 995 | outputs.append(mask) |
| 996 | |
| 997 | return torch.cat(outputs, dim=0) |
| 998 | |
| 999 | |
| 1000 | |
| 1001 | |
| 1002 | def loadcheckpoints(self,model_path): |
| 1003 | model_dict = torch.load(model_path, map_location="cpu", weights_only=True) |
| 1004 | self.load_state_dict(model_dict['model_state_dict'], strict=True) |
| 1005 | del model_path |
| 1006 | |
| 1007 | def inference(self,image,refine_foreground=False): |
| 1008 | |
| 1009 | set_random_seed(9) |
| 1010 | # image = ImageOps.exif_transpose(image) |
| 1011 | if isinstance(image, Image.Image): |
| 1012 | image, h, w,original_image = rgb_loader_refiner(image) |
| 1013 | if torch.cuda.is_available(): |
| 1014 | |
| 1015 | img_tensor = img_transform(image).unsqueeze(0).to(next(self.parameters()).device) |
| 1016 | else: |
| 1017 | img_tensor = img_transform32(image).unsqueeze(0).to(next(self.parameters()).device) |
| 1018 | |
| 1019 | |
| 1020 | with torch.no_grad(): |
| 1021 | res = self.forward(img_tensor) |
| 1022 | |
| 1023 | # Show Results |
| 1024 | if refine_foreground == True: |
| 1025 | |
| 1026 | pred_pil = transforms.ToPILImage()(res.squeeze()) |
| 1027 | image_masked = refine_foreground_process(original_image, pred_pil) |
| 1028 | |
| 1029 | image_masked.putalpha(pred_pil.resize(original_image.size)) |
| 1030 | return image_masked |
| 1031 | |
| 1032 | else: |
| 1033 | alpha = postprocess_image(res, im_size=[w,h]) |
| 1034 | pred_pil = transforms.ToPILImage()(alpha) |
| 1035 | mask = pred_pil.resize(original_image.size) |
| 1036 | original_image.putalpha(mask) |
| 1037 | # mask = Image.fromarray(alpha) |
| 1038 | |
| 1039 | return original_image |
| 1040 | |
| 1041 | |
| 1042 | else: |
| 1043 | foregrounds = [] |
| 1044 | for batch in image: |
| 1045 | image, h, w,original_image = rgb_loader_refiner(batch) |
| 1046 | if torch.cuda.is_available(): |
| 1047 | |
| 1048 | img_tensor = img_transform(image).unsqueeze(0).to(next(self.parameters()).device) |
| 1049 | else: |
| 1050 | img_tensor = img_transform32(image).unsqueeze(0).to(next(self.parameters()).device) |
| 1051 | |
| 1052 | with torch.no_grad(): |
| 1053 | res = self.forward(img_tensor) |
| 1054 | |
| 1055 | if refine_foreground == True: |
| 1056 | |
| 1057 | pred_pil = transforms.ToPILImage()(res.squeeze()) |
| 1058 | image_masked = refine_foreground_process(original_image, pred_pil) |
| 1059 | |
| 1060 | image_masked.putalpha(pred_pil.resize(original_image.size)) |
| 1061 | |
| 1062 | foregrounds.append(image_masked) |
| 1063 | else: |
| 1064 | alpha = postprocess_image(res, im_size=[w,h]) |
| 1065 | pred_pil = transforms.ToPILImage()(alpha) |
| 1066 | mask = pred_pil.resize(original_image.size) |
| 1067 | original_image.putalpha(mask) |
| 1068 | # mask = Image.fromarray(alpha) |
| 1069 | foregrounds.append(original_image) |
| 1070 | |
| 1071 | return foregrounds |
| 1072 | |
| 1073 | |
| 1074 | |
| 1075 | |
| 1076 | def segment_video(self, video_path, output_path="./", fps=0, refine_foreground=False, batch=1, print_frames_processed=True, webm = False, rgb_value= (0, 255, 0)): |
| 1077 | |
| 1078 | """ |
| 1079 | Segments the given video to extract the foreground (with alpha) from each frame |
| 1080 | and saves the result as either a WebM video (with alpha channel) or MP4 (with a |
| 1081 | color background). |
| 1082 | |
| 1083 | Args: |
| 1084 | video_path (str): |
| 1085 | Path to the input video file. |
| 1086 | |
| 1087 | output_path (str, optional): |
| 1088 | Directory (or full path) where the output video and/or files will be saved. |
| 1089 | Defaults to "./". |
| 1090 | |
| 1091 | fps (int, optional): |
| 1092 | The frames per second (FPS) to use for the output video. If 0 (default), the |
| 1093 | original FPS of the input video is used. Otherwise, overrides it. |
| 1094 | |
| 1095 | refine_foreground (bool, optional): |
| 1096 | Whether to run an additional “refine foreground” process on each frame. |
| 1097 | Defaults to False. |
| 1098 | |
| 1099 | batch (int, optional): |
| 1100 | Number of frames to process at once (inference batch size). Large batch sizes |
| 1101 | may require more GPU memory. Defaults to 1. |
| 1102 | |
| 1103 | print_frames_processed (bool, optional): |
| 1104 | If True (default), prints progress (how many frames have been processed) to |
| 1105 | the console. |
| 1106 | |
| 1107 | webm (bool, optional): |
| 1108 | If True (default), exports a WebM video with alpha channel (VP9 / yuva420p). |
| 1109 | If False, exports an MP4 video composited over a solid color background. |
| 1110 | |
| 1111 | rgb_value (tuple, optional): |
| 1112 | The RGB background color (e.g., green screen) used to composite frames when |
| 1113 | saving to MP4. Defaults to (0, 255, 0). |
| 1114 | |
| 1115 | Returns: |
| 1116 | None. Writes the output video(s) to disk in the specified format. |
| 1117 | """ |
| 1118 | |
| 1119 | |
| 1120 | cap = cv2.VideoCapture(video_path) |
| 1121 | if not cap.isOpened(): |
| 1122 | raise IOError(f"Cannot open video: {video_path}") |
| 1123 | |
| 1124 | original_fps = cap.get(cv2.CAP_PROP_FPS) |
| 1125 | original_fps = 30 if original_fps == 0 else original_fps |
| 1126 | fps = original_fps if fps == 0 else fps |
| 1127 | |
| 1128 | ret, first_frame = cap.read() |
| 1129 | if not ret: |
| 1130 | raise ValueError("No frames found in the video.") |
| 1131 | height, width = first_frame.shape[:2] |
| 1132 | cap.set(cv2.CAP_PROP_POS_FRAMES, 0) |
| 1133 | |
| 1134 | foregrounds = [] |
| 1135 | frame_idx = 0 |
| 1136 | processed_count = 0 |
| 1137 | batch_frames = [] |
| 1138 | total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT)) |
| 1139 | |
| 1140 | while True: |
| 1141 | ret, frame = cap.read() |
| 1142 | if not ret: |
| 1143 | if batch_frames: |
| 1144 | batch_results = self.inference(batch_frames, refine_foreground) |
| 1145 | if isinstance(batch_results, Image.Image): |
| 1146 | foregrounds.append(batch_results) |
| 1147 | else: |
| 1148 | foregrounds.extend(batch_results) |
| 1149 | if print_frames_processed: |
| 1150 | print(f"Processed frames {frame_idx-len(batch_frames)+1} to {frame_idx} of {total_frames}") |
| 1151 | break |
| 1152 | |
| 1153 | # Process every frame instead of using intervals |
| 1154 | frame_rgb = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB) |
| 1155 | pil_frame = Image.fromarray(frame_rgb) |
| 1156 | batch_frames.append(pil_frame) |
| 1157 | |
| 1158 | if len(batch_frames) == batch: |
| 1159 | batch_results = self.inference(batch_frames, refine_foreground) |
| 1160 | if isinstance(batch_results, Image.Image): |
| 1161 | foregrounds.append(batch_results) |
| 1162 | else: |
| 1163 | foregrounds.extend(batch_results) |
| 1164 | if print_frames_processed: |
| 1165 | print(f"Processed frames {frame_idx-batch+1} to {frame_idx} of {total_frames}") |
| 1166 | batch_frames = [] |
| 1167 | processed_count += batch |
| 1168 | |
| 1169 | frame_idx += 1 |
| 1170 | |
| 1171 | |
| 1172 | if webm: |
| 1173 | alpha_webm_path = os.path.join(output_path, "foreground.webm") |
| 1174 | pil_images_to_webm_alpha(foregrounds, alpha_webm_path, fps=original_fps) |
| 1175 | |
| 1176 | else: |
| 1177 | cap.release() |
| 1178 | fg_output = os.path.join(output_path, 'foreground.mp4') |
| 1179 | |
| 1180 | pil_images_to_mp4(foregrounds, fg_output, fps=original_fps,rgb_value=rgb_value) |
| 1181 | cv2.destroyAllWindows() |
| 1182 | |
| 1183 | try: |
| 1184 | fg_audio_output = os.path.join(output_path, 'foreground_output_with_audio.mp4') |
| 1185 | add_audio_to_video(fg_output, video_path, fg_audio_output) |
| 1186 | except Exception as e: |
| 1187 | print("No audio found in the original video") |
| 1188 | print(e) |
| 1189 | |
| 1190 | |
| 1191 | |
| 1192 | |
| 1193 | |
| 1194 | def rgb_loader_refiner( original_image): |
| 1195 | h, w = original_image.size |
| 1196 | |
| 1197 | image = original_image |
| 1198 | # Convert to RGB if necessary |
| 1199 | if image.mode != 'RGB': |
| 1200 | image = image.convert('RGB') |
| 1201 | |
| 1202 | # Resize the image |
| 1203 | image = image.resize((1024, 1024), resample=Image.LANCZOS) |
| 1204 | |
| 1205 | return image.convert('RGB'), h, w,original_image |
| 1206 | |
| 1207 | # Define the image transformation |
| 1208 | img_transform = transforms.Compose([ |
| 1209 | transforms.ToTensor(), |
| 1210 | transforms.ConvertImageDtype(torch.float16), |
| 1211 | transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) |
| 1212 | ]) |
| 1213 | |
| 1214 | img_transform32 = transforms.Compose([ |
| 1215 | transforms.ToTensor(), |
| 1216 | transforms.ConvertImageDtype(torch.float32), |
| 1217 | transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) |
| 1218 | ]) |
| 1219 | |
| 1220 | |
| 1221 | |
| 1222 | |
| 1223 | |
| 1224 | def pil_images_to_mp4(images, output_path, fps=24, rgb_value=(0, 255, 0)): |
| 1225 | """ |
| 1226 | Converts an array of PIL images to an MP4 video. |
| 1227 | |
| 1228 | Args: |
| 1229 | images: List of PIL images |
| 1230 | output_path: Path to save the MP4 file |
| 1231 | fps: Frames per second (default: 24) |
| 1232 | rgb_value: Background RGB color tuple (default: green (0, 255, 0)) |
| 1233 | """ |
| 1234 | if not images: |
| 1235 | raise ValueError("No images provided to convert to MP4.") |
| 1236 | |
| 1237 | width, height = images[0].size |
| 1238 | fourcc = cv2.VideoWriter_fourcc(*'mp4v') |
| 1239 | video_writer = cv2.VideoWriter(output_path, fourcc, fps, (width, height)) |
| 1240 | |
| 1241 | for image in images: |
| 1242 | # If image has alpha channel, composite onto the specified background color |
| 1243 | if image.mode == 'RGBA': |
| 1244 | # Create background image with specified RGB color |
| 1245 | background = Image.new('RGB', image.size, rgb_value) |
| 1246 | background = background.convert('RGBA') |
| 1247 | # Composite the image onto the background |
| 1248 | image = Image.alpha_composite(background, image) |
| 1249 | image = image.convert('RGB') |
| 1250 | else: |
| 1251 | # Ensure RGB format for non-alpha images |
| 1252 | image = image.convert('RGB') |
| 1253 | |
| 1254 | # Convert to OpenCV format and write |
| 1255 | open_cv_image = cv2.cvtColor(np.array(image), cv2.COLOR_RGB2BGR) |
| 1256 | video_writer.write(open_cv_image) |
| 1257 | |
| 1258 | video_writer.release() |
| 1259 | |
| 1260 | def pil_images_to_webm_alpha(images, output_path, fps=30): |
| 1261 | """ |
| 1262 | Converts a list of PIL RGBA images to a VP9 .webm video with alpha channel. |
| 1263 | |
| 1264 | NOTE: Not all players will display alpha in WebM. |
| 1265 | Browsers like Chrome/Firefox typically do support VP9 alpha. |
| 1266 | """ |
| 1267 | if not images: |
| 1268 | raise ValueError("No images provided for WebM with alpha.") |
| 1269 | |
| 1270 | # Ensure output directory exists |
| 1271 | os.makedirs(os.path.dirname(output_path), exist_ok=True) |
| 1272 | |
| 1273 | with tempfile.TemporaryDirectory() as tmpdir: |
| 1274 | # Save frames as PNG (with alpha) |
| 1275 | for idx, img in enumerate(images): |
| 1276 | if img.mode != "RGBA": |
| 1277 | img = img.convert("RGBA") |
| 1278 | out_path = os.path.join(tmpdir, f"{idx:06d}.png") |
| 1279 | img.save(out_path, "PNG") |
| 1280 | |
| 1281 | # Construct ffmpeg command |
| 1282 | # -c:v libvpx-vp9 => VP9 encoder |
| 1283 | # -pix_fmt yuva420p => alpha-enabled pixel format |
| 1284 | # -auto-alt-ref 0 => helps preserve alpha frames (libvpx quirk) |
| 1285 | ffmpeg_cmd = [ |
| 1286 | "ffmpeg", "-y", |
| 1287 | "-framerate", str(fps), |
| 1288 | "-i", os.path.join(tmpdir, "%06d.png"), |
| 1289 | "-c:v", "libvpx-vp9", |
| 1290 | "-pix_fmt", "yuva420p", |
| 1291 | "-auto-alt-ref", "0", |
| 1292 | output_path |
| 1293 | ] |
| 1294 | |
| 1295 | subprocess.run(ffmpeg_cmd, check=True) |
| 1296 | |
| 1297 | print(f"WebM with alpha saved to {output_path}") |
| 1298 | |
| 1299 | def add_audio_to_video(video_without_audio_path, original_video_path, output_path): |
| 1300 | """ |
| 1301 | Check if the original video has an audio stream. If yes, add it. If not, skip. |
| 1302 | """ |
| 1303 | # 1) Probe original video for audio streams |
| 1304 | probe_command = [ |
| 1305 | 'ffprobe', '-v', 'error', |
| 1306 | '-select_streams', 'a:0', |
| 1307 | '-show_entries', 'stream=index', |
| 1308 | '-of', 'csv=p=0', |
| 1309 | original_video_path |
| 1310 | ] |
| 1311 | result = subprocess.run(probe_command, stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True) |
| 1312 | |
| 1313 | # result.stdout is empty if no audio stream found |
| 1314 | if not result.stdout.strip(): |
| 1315 | print("No audio track found in original video, skipping audio addition.") |
| 1316 | return |
| 1317 | |
| 1318 | print("Audio track detected; proceeding to mux audio.") |
| 1319 | # 2) If audio found, run ffmpeg to add it |
| 1320 | command = [ |
| 1321 | 'ffmpeg', '-y', |
| 1322 | '-i', video_without_audio_path, |
| 1323 | '-i', original_video_path, |
| 1324 | '-c', 'copy', |
| 1325 | '-map', '0:v:0', |
| 1326 | '-map', '1:a:0', # we know there's an audio track now |
| 1327 | output_path |
| 1328 | ] |
| 1329 | subprocess.run(command, check=True) |
| 1330 | print(f"Audio added successfully => {output_path}") |
| 1331 | |
| 1332 | |
| 1333 | |
| 1334 | |
| 1335 | |
| 1336 | ### Thanks to the source: https://huggingface.co/ZhengPeng7/BiRefNet/blob/main/handler.py |
| 1337 | def refine_foreground_process(image, mask, r=90): |
| 1338 | if mask.size != image.size: |
| 1339 | mask = mask.resize(image.size) |
| 1340 | image = np.array(image) / 255.0 |
| 1341 | mask = np.array(mask) / 255.0 |
| 1342 | estimated_foreground = FB_blur_fusion_foreground_estimator_2(image, mask, r=r) |
| 1343 | image_masked = Image.fromarray((estimated_foreground * 255.0).astype(np.uint8)) |
| 1344 | return image_masked |
| 1345 | |
| 1346 | |
| 1347 | def FB_blur_fusion_foreground_estimator_2(image, alpha, r=90): |
| 1348 | # Thanks to the source: https://github.com/Photoroom/fast-foreground-estimation |
| 1349 | alpha = alpha[:, :, None] |
| 1350 | F, blur_B = FB_blur_fusion_foreground_estimator(image, image, image, alpha, r) |
| 1351 | return FB_blur_fusion_foreground_estimator(image, F, blur_B, alpha, r=6)[0] |
| 1352 | |
| 1353 | |
| 1354 | def FB_blur_fusion_foreground_estimator(image, F, B, alpha, r=90): |
| 1355 | if isinstance(image, Image.Image): |
| 1356 | image = np.array(image) / 255.0 |
| 1357 | blurred_alpha = cv2.blur(alpha, (r, r))[:, :, None] |
| 1358 | |
| 1359 | blurred_FA = cv2.blur(F * alpha, (r, r)) |
| 1360 | blurred_F = blurred_FA / (blurred_alpha + 1e-5) |
| 1361 | |
| 1362 | blurred_B1A = cv2.blur(B * (1 - alpha), (r, r)) |
| 1363 | blurred_B = blurred_B1A / ((1 - blurred_alpha) + 1e-5) |
| 1364 | F = blurred_F + alpha * \ |
| 1365 | (image - alpha * blurred_F - (1 - alpha) * blurred_B) |
| 1366 | F = np.clip(F, 0, 1) |
| 1367 | return F, blurred_B |
| 1368 | |
| 1369 | |
| 1370 | |
| 1371 | def postprocess_image(result: torch.Tensor, im_size: list) -> np.ndarray: |
| 1372 | result = torch.squeeze(F.interpolate(result, size=im_size, mode='bilinear'), 0) |
| 1373 | ma = torch.max(result) |
| 1374 | mi = torch.min(result) |
| 1375 | result = (result - mi) / (ma - mi) |
| 1376 | im_array = (result * 255).permute(1, 2, 0).cpu().data.numpy().astype(np.uint8) |
| 1377 | im_array = np.squeeze(im_array) |
| 1378 | return im_array |
| 1379 | |
| 1380 | |
| 1381 | |
| 1382 | |
| 1383 | def rgb_loader_refiner( original_image): |
| 1384 | h, w = original_image.size |
| 1385 | # # Apply EXIF orientation |
| 1386 | |
| 1387 | image = ImageOps.exif_transpose(original_image) |
| 1388 | |
| 1389 | if original_image.mode != 'RGB': |
| 1390 | original_image = original_image.convert('RGB') |
| 1391 | |
| 1392 | image = original_image |
| 1393 | # Convert to RGB if necessary |
| 1394 | |
| 1395 | # Resize the image |
| 1396 | image = image.resize((1024, 1024), resample=Image.LANCZOS) |
| 1397 | |
| 1398 | return image, h, w,original_image |
| 1399 | |
| 1400 | |
| 1401 | |
| 1402 | |