birefnet.py
89.7 KB · 2249 lines · python Raw
1 ### config.py
2
3 import os
4 import math
5 from transformers import PretrainedConfig
6
7
8 class Config(PretrainedConfig):
9 def __init__(self) -> None:
10 # Compatible with the latest version of transformers
11 super().__init__()
12
13 # PATH settings
14 self.sys_home_dir = os.path.expanduser('~') # Make up your file system as: SYS_HOME_DIR/codes/dis/BiRefNet, SYS_HOME_DIR/datasets/dis/xx, SYS_HOME_DIR/weights/xx
15
16 # TASK settings
17 self.task = ['DIS5K', 'COD', 'HRSOD', 'DIS5K+HRSOD+HRS10K', 'P3M-10k'][0]
18 self.training_set = {
19 'DIS5K': ['DIS-TR', 'DIS-TR+DIS-TE1+DIS-TE2+DIS-TE3+DIS-TE4'][0],
20 'COD': 'TR-COD10K+TR-CAMO',
21 'HRSOD': ['TR-DUTS', 'TR-HRSOD', 'TR-UHRSD', 'TR-DUTS+TR-HRSOD', 'TR-DUTS+TR-UHRSD', 'TR-HRSOD+TR-UHRSD', 'TR-DUTS+TR-HRSOD+TR-UHRSD'][5],
22 'DIS5K+HRSOD+HRS10K': 'DIS-TE1+DIS-TE2+DIS-TE3+DIS-TE4+DIS-TR+TE-HRS10K+TE-HRSOD+TE-UHRSD+TR-HRS10K+TR-HRSOD+TR-UHRSD', # leave DIS-VD for evaluation.
23 'P3M-10k': 'TR-P3M-10k',
24 }[self.task]
25 self.prompt4loc = ['dense', 'sparse'][0]
26
27 # Faster-Training settings
28 self.load_all = True
29 self.compile = True # 1. Trigger CPU memory leak in some extend, which is an inherent problem of PyTorch.
30 # Machines with > 70GB CPU memory can run the whole training on DIS5K with default setting.
31 # 2. Higher PyTorch version may fix it: https://github.com/pytorch/pytorch/issues/119607.
32 # 3. But compile in Pytorch > 2.0.1 seems to bring no acceleration for training.
33 self.precisionHigh = True
34
35 # MODEL settings
36 self.ms_supervision = True
37 self.out_ref = self.ms_supervision and True
38 self.dec_ipt = True
39 self.dec_ipt_split = True
40 self.cxt_num = [0, 3][1] # multi-scale skip connections from encoder
41 self.mul_scl_ipt = ['', 'add', 'cat'][2]
42 self.dec_att = ['', 'ASPP', 'ASPPDeformable'][2]
43 self.squeeze_block = ['', 'BasicDecBlk_x1', 'ResBlk_x4', 'ASPP_x3', 'ASPPDeformable_x3'][1]
44 self.dec_blk = ['BasicDecBlk', 'ResBlk', 'HierarAttDecBlk'][0]
45
46 # TRAINING settings
47 self.batch_size = 4
48 self.IoU_finetune_last_epochs = [
49 0,
50 {
51 'DIS5K': -50,
52 'COD': -20,
53 'HRSOD': -20,
54 'DIS5K+HRSOD+HRS10K': -20,
55 'P3M-10k': -20,
56 }[self.task]
57 ][1] # choose 0 to skip
58 self.lr = (1e-4 if 'DIS5K' in self.task else 1e-5) * math.sqrt(self.batch_size / 4) # DIS needs high lr to converge faster. Adapt the lr linearly
59 self.size = 1024
60 self.num_workers = max(4, self.batch_size) # will be decrease to min(it, batch_size) at the initialization of the data_loader
61
62 # Backbone settings
63 self.bb = [
64 'vgg16', 'vgg16bn', 'resnet50', # 0, 1, 2
65 'swin_v1_t', 'swin_v1_s', # 3, 4
66 'swin_v1_b', 'swin_v1_l', # 5-bs9, 6-bs4
67 'pvt_v2_b0', 'pvt_v2_b1', # 7, 8
68 'pvt_v2_b2', 'pvt_v2_b5', # 9-bs10, 10-bs5
69 ][6]
70 self.lateral_channels_in_collection = {
71 'vgg16': [512, 256, 128, 64], 'vgg16bn': [512, 256, 128, 64], 'resnet50': [1024, 512, 256, 64],
72 'pvt_v2_b2': [512, 320, 128, 64], 'pvt_v2_b5': [512, 320, 128, 64],
73 'swin_v1_b': [1024, 512, 256, 128], 'swin_v1_l': [1536, 768, 384, 192],
74 'swin_v1_t': [768, 384, 192, 96], 'swin_v1_s': [768, 384, 192, 96],
75 'pvt_v2_b0': [256, 160, 64, 32], 'pvt_v2_b1': [512, 320, 128, 64],
76 }[self.bb]
77 if self.mul_scl_ipt == 'cat':
78 self.lateral_channels_in_collection = [channel * 2 for channel in self.lateral_channels_in_collection]
79 self.cxt = self.lateral_channels_in_collection[1:][::-1][-self.cxt_num:] if self.cxt_num else []
80
81 # MODEL settings - inactive
82 self.lat_blk = ['BasicLatBlk'][0]
83 self.dec_channels_inter = ['fixed', 'adap'][0]
84 self.refine = ['', 'itself', 'RefUNet', 'Refiner', 'RefinerPVTInChannels4'][0]
85 self.progressive_ref = self.refine and True
86 self.ender = self.progressive_ref and False
87 self.scale = self.progressive_ref and 2
88 self.auxiliary_classification = False # Only for DIS5K, where class labels are saved in `dataset.py`.
89 self.refine_iteration = 1
90 self.freeze_bb = False
91 self.model = [
92 'BiRefNet',
93 ][0]
94 if self.dec_blk == 'HierarAttDecBlk':
95 self.batch_size = 2 ** [0, 1, 2, 3, 4][2]
96
97 # TRAINING settings - inactive
98 self.preproc_methods = ['flip', 'enhance', 'rotate', 'pepper', 'crop'][:4]
99 self.optimizer = ['Adam', 'AdamW'][1]
100 self.lr_decay_epochs = [1e5] # Set to negative N to decay the lr in the last N-th epoch.
101 self.lr_decay_rate = 0.5
102 # Loss
103 self.lambdas_pix_last = {
104 # not 0 means opening this loss
105 # original rate -- 1 : 30 : 1.5 : 0.2, bce x 30
106 'bce': 30 * 1, # high performance
107 'iou': 0.5 * 1, # 0 / 255
108 'iou_patch': 0.5 * 0, # 0 / 255, win_size = (64, 64)
109 'mse': 150 * 0, # can smooth the saliency map
110 'triplet': 3 * 0,
111 'reg': 100 * 0,
112 'ssim': 10 * 1, # help contours,
113 'cnt': 5 * 0, # help contours
114 'structure': 5 * 0, # structure loss from codes of MVANet. A little improvement on DIS-TE[1,2,3], a bit more decrease on DIS-TE4.
115 }
116 self.lambdas_cls = {
117 'ce': 5.0
118 }
119 # Adv
120 self.lambda_adv_g = 10. * 0 # turn to 0 to avoid adv training
121 self.lambda_adv_d = 3. * (self.lambda_adv_g > 0)
122
123 # PATH settings - inactive
124 self.data_root_dir = os.path.join(self.sys_home_dir, 'datasets/dis')
125 self.weights_root_dir = os.path.join(self.sys_home_dir, 'weights')
126 self.weights = {
127 'pvt_v2_b2': os.path.join(self.weights_root_dir, 'pvt_v2_b2.pth'),
128 'pvt_v2_b5': os.path.join(self.weights_root_dir, ['pvt_v2_b5.pth', 'pvt_v2_b5_22k.pth'][0]),
129 'swin_v1_b': os.path.join(self.weights_root_dir, ['swin_base_patch4_window12_384_22kto1k.pth', 'swin_base_patch4_window12_384_22k.pth'][0]),
130 'swin_v1_l': os.path.join(self.weights_root_dir, ['swin_large_patch4_window12_384_22kto1k.pth', 'swin_large_patch4_window12_384_22k.pth'][0]),
131 'swin_v1_t': os.path.join(self.weights_root_dir, ['swin_tiny_patch4_window7_224_22kto1k_finetune.pth'][0]),
132 'swin_v1_s': os.path.join(self.weights_root_dir, ['swin_small_patch4_window7_224_22kto1k_finetune.pth'][0]),
133 'pvt_v2_b0': os.path.join(self.weights_root_dir, ['pvt_v2_b0.pth'][0]),
134 'pvt_v2_b1': os.path.join(self.weights_root_dir, ['pvt_v2_b1.pth'][0]),
135 }
136
137 # Callbacks - inactive
138 self.verbose_eval = True
139 self.only_S_MAE = False
140 self.use_fp16 = False # Bugs. It may cause nan in training.
141 self.SDPA_enabled = False # Bugs. Slower and errors occur in multi-GPUs
142
143 # others
144 self.device = [0, 'cpu'][0] # .to(0) == .to('cuda:0')
145
146 self.batch_size_valid = 1
147 self.rand_seed = 7
148 # run_sh_file = [f for f in os.listdir('.') if 'train.sh' == f] + [os.path.join('..', f) for f in os.listdir('..') if 'train.sh' == f]
149 # with open(run_sh_file[0], 'r') as f:
150 # lines = f.readlines()
151 # self.save_last = int([l.strip() for l in lines if '"{}")'.format(self.task) in l and 'val_last=' in l][0].split('val_last=')[-1].split()[0])
152 # self.save_step = int([l.strip() for l in lines if '"{}")'.format(self.task) in l and 'step=' in l][0].split('step=')[-1].split()[0])
153 # self.val_step = [0, self.save_step][0]
154
155 def print_task(self) -> None:
156 # Return task for choosing settings in shell scripts.
157 print(self.task)
158
159
160
161 ### models/backbones/pvt_v2.py
162
163 import torch
164 import torch.nn as nn
165 from functools import partial
166
167 from timm.layers import DropPath, to_2tuple, trunc_normal_
168
169
170 import math
171
172 # from config import Config
173
174 # config = Config()
175
176 class Mlp(nn.Module):
177 def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):
178 super().__init__()
179 out_features = out_features or in_features
180 hidden_features = hidden_features or in_features
181 self.fc1 = nn.Linear(in_features, hidden_features)
182 self.dwconv = DWConv(hidden_features)
183 self.act = act_layer()
184 self.fc2 = nn.Linear(hidden_features, out_features)
185 self.drop = nn.Dropout(drop)
186
187 self.apply(self._init_weights)
188
189 def _init_weights(self, m):
190 if isinstance(m, nn.Linear):
191 trunc_normal_(m.weight, std=.02)
192 if isinstance(m, nn.Linear) and m.bias is not None:
193 nn.init.constant_(m.bias, 0)
194 elif isinstance(m, nn.LayerNorm):
195 nn.init.constant_(m.bias, 0)
196 nn.init.constant_(m.weight, 1.0)
197 elif isinstance(m, nn.Conv2d):
198 fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
199 fan_out //= m.groups
200 m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
201 if m.bias is not None:
202 m.bias.data.zero_()
203
204 def forward(self, x, H, W):
205 x = self.fc1(x)
206 x = self.dwconv(x, H, W)
207 x = self.act(x)
208 x = self.drop(x)
209 x = self.fc2(x)
210 x = self.drop(x)
211 return x
212
213
214 class Attention(nn.Module):
215 def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0., sr_ratio=1):
216 super().__init__()
217 assert dim % num_heads == 0, f"dim {dim} should be divided by num_heads {num_heads}."
218
219 self.dim = dim
220 self.num_heads = num_heads
221 head_dim = dim // num_heads
222 self.scale = qk_scale or head_dim ** -0.5
223
224 self.q = nn.Linear(dim, dim, bias=qkv_bias)
225 self.kv = nn.Linear(dim, dim * 2, bias=qkv_bias)
226 self.attn_drop_prob = attn_drop
227 self.attn_drop = nn.Dropout(attn_drop)
228 self.proj = nn.Linear(dim, dim)
229 self.proj_drop = nn.Dropout(proj_drop)
230
231 self.sr_ratio = sr_ratio
232 if sr_ratio > 1:
233 self.sr = nn.Conv2d(dim, dim, kernel_size=sr_ratio, stride=sr_ratio)
234 self.norm = nn.LayerNorm(dim)
235
236 self.apply(self._init_weights)
237
238 def _init_weights(self, m):
239 if isinstance(m, nn.Linear):
240 trunc_normal_(m.weight, std=.02)
241 if isinstance(m, nn.Linear) and m.bias is not None:
242 nn.init.constant_(m.bias, 0)
243 elif isinstance(m, nn.LayerNorm):
244 nn.init.constant_(m.bias, 0)
245 nn.init.constant_(m.weight, 1.0)
246 elif isinstance(m, nn.Conv2d):
247 fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
248 fan_out //= m.groups
249 m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
250 if m.bias is not None:
251 m.bias.data.zero_()
252
253 def forward(self, x, H, W):
254 B, N, C = x.shape
255 q = self.q(x).reshape(B, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)
256
257 if self.sr_ratio > 1:
258 x_ = x.permute(0, 2, 1).reshape(B, C, H, W)
259 x_ = self.sr(x_).reshape(B, C, -1).permute(0, 2, 1)
260 x_ = self.norm(x_)
261 kv = self.kv(x_).reshape(B, -1, 2, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
262 else:
263 kv = self.kv(x).reshape(B, -1, 2, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
264 k, v = kv[0], kv[1]
265
266 if config.SDPA_enabled:
267 x = torch.nn.functional.scaled_dot_product_attention(
268 q, k, v,
269 attn_mask=None, dropout_p=self.attn_drop_prob, is_causal=False
270 ).transpose(1, 2).reshape(B, N, C)
271 else:
272 attn = (q @ k.transpose(-2, -1)) * self.scale
273 attn = attn.softmax(dim=-1)
274 attn = self.attn_drop(attn)
275
276 x = (attn @ v).transpose(1, 2).reshape(B, N, C)
277 x = self.proj(x)
278 x = self.proj_drop(x)
279
280 return x
281
282
283 class Block(nn.Module):
284
285 def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0.,
286 drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm, sr_ratio=1):
287 super().__init__()
288 self.norm1 = norm_layer(dim)
289 self.attn = Attention(
290 dim,
291 num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale,
292 attn_drop=attn_drop, proj_drop=drop, sr_ratio=sr_ratio)
293 # NOTE: drop path for stochastic depth, we shall see if this is better than dropout here
294 self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
295 self.norm2 = norm_layer(dim)
296 mlp_hidden_dim = int(dim * mlp_ratio)
297 self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
298
299 self.apply(self._init_weights)
300
301 def _init_weights(self, m):
302 if isinstance(m, nn.Linear):
303 trunc_normal_(m.weight, std=.02)
304 if isinstance(m, nn.Linear) and m.bias is not None:
305 nn.init.constant_(m.bias, 0)
306 elif isinstance(m, nn.LayerNorm):
307 nn.init.constant_(m.bias, 0)
308 nn.init.constant_(m.weight, 1.0)
309 elif isinstance(m, nn.Conv2d):
310 fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
311 fan_out //= m.groups
312 m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
313 if m.bias is not None:
314 m.bias.data.zero_()
315
316 def forward(self, x, H, W):
317 x = x + self.drop_path(self.attn(self.norm1(x), H, W))
318 x = x + self.drop_path(self.mlp(self.norm2(x), H, W))
319
320 return x
321
322
323 class OverlapPatchEmbed(nn.Module):
324 """ Image to Patch Embedding
325 """
326
327 def __init__(self, img_size=224, patch_size=7, stride=4, in_channels=3, embed_dim=768):
328 super().__init__()
329 img_size = to_2tuple(img_size)
330 patch_size = to_2tuple(patch_size)
331
332 self.img_size = img_size
333 self.patch_size = patch_size
334 self.H, self.W = img_size[0] // patch_size[0], img_size[1] // patch_size[1]
335 self.num_patches = self.H * self.W
336 self.proj = nn.Conv2d(in_channels, embed_dim, kernel_size=patch_size, stride=stride,
337 padding=(patch_size[0] // 2, patch_size[1] // 2))
338 self.norm = nn.LayerNorm(embed_dim)
339
340 self.apply(self._init_weights)
341
342 def _init_weights(self, m):
343 if isinstance(m, nn.Linear):
344 trunc_normal_(m.weight, std=.02)
345 if isinstance(m, nn.Linear) and m.bias is not None:
346 nn.init.constant_(m.bias, 0)
347 elif isinstance(m, nn.LayerNorm):
348 nn.init.constant_(m.bias, 0)
349 nn.init.constant_(m.weight, 1.0)
350 elif isinstance(m, nn.Conv2d):
351 fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
352 fan_out //= m.groups
353 m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
354 if m.bias is not None:
355 m.bias.data.zero_()
356
357 def forward(self, x):
358 x = self.proj(x)
359 _, _, H, W = x.shape
360 x = x.flatten(2).transpose(1, 2)
361 x = self.norm(x)
362
363 return x, H, W
364
365
366 class PyramidVisionTransformerImpr(nn.Module):
367 def __init__(self, img_size=224, patch_size=16, in_channels=3, num_classes=1000, embed_dims=[64, 128, 256, 512],
368 num_heads=[1, 2, 4, 8], mlp_ratios=[4, 4, 4, 4], qkv_bias=False, qk_scale=None, drop_rate=0.,
369 attn_drop_rate=0., drop_path_rate=0., norm_layer=nn.LayerNorm,
370 depths=[3, 4, 6, 3], sr_ratios=[8, 4, 2, 1]):
371 super().__init__()
372 self.num_classes = num_classes
373 self.depths = depths
374
375 # patch_embed
376 self.patch_embed1 = OverlapPatchEmbed(img_size=img_size, patch_size=7, stride=4, in_channels=in_channels,
377 embed_dim=embed_dims[0])
378 self.patch_embed2 = OverlapPatchEmbed(img_size=img_size // 4, patch_size=3, stride=2, in_channels=embed_dims[0],
379 embed_dim=embed_dims[1])
380 self.patch_embed3 = OverlapPatchEmbed(img_size=img_size // 8, patch_size=3, stride=2, in_channels=embed_dims[1],
381 embed_dim=embed_dims[2])
382 self.patch_embed4 = OverlapPatchEmbed(img_size=img_size // 16, patch_size=3, stride=2, in_channels=embed_dims[2],
383 embed_dim=embed_dims[3])
384
385 # transformer encoder
386 dpr = np.linspace(0, drop_path_rate, sum(depths)).tolist() # stochastic depth decay rule
387 cur = 0
388 self.block1 = nn.ModuleList([Block(
389 dim=embed_dims[0], num_heads=num_heads[0], mlp_ratio=mlp_ratios[0], qkv_bias=qkv_bias, qk_scale=qk_scale,
390 drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[cur + i], norm_layer=norm_layer,
391 sr_ratio=sr_ratios[0])
392 for i in range(depths[0])])
393 self.norm1 = norm_layer(embed_dims[0])
394
395 cur += depths[0]
396 self.block2 = nn.ModuleList([Block(
397 dim=embed_dims[1], num_heads=num_heads[1], mlp_ratio=mlp_ratios[1], qkv_bias=qkv_bias, qk_scale=qk_scale,
398 drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[cur + i], norm_layer=norm_layer,
399 sr_ratio=sr_ratios[1])
400 for i in range(depths[1])])
401 self.norm2 = norm_layer(embed_dims[1])
402
403 cur += depths[1]
404 self.block3 = nn.ModuleList([Block(
405 dim=embed_dims[2], num_heads=num_heads[2], mlp_ratio=mlp_ratios[2], qkv_bias=qkv_bias, qk_scale=qk_scale,
406 drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[cur + i], norm_layer=norm_layer,
407 sr_ratio=sr_ratios[2])
408 for i in range(depths[2])])
409 self.norm3 = norm_layer(embed_dims[2])
410
411 cur += depths[2]
412 self.block4 = nn.ModuleList([Block(
413 dim=embed_dims[3], num_heads=num_heads[3], mlp_ratio=mlp_ratios[3], qkv_bias=qkv_bias, qk_scale=qk_scale,
414 drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[cur + i], norm_layer=norm_layer,
415 sr_ratio=sr_ratios[3])
416 for i in range(depths[3])])
417 self.norm4 = norm_layer(embed_dims[3])
418
419 # classification head
420 # self.head = nn.Linear(embed_dims[3], num_classes) if num_classes > 0 else nn.Identity()
421
422 self.apply(self._init_weights)
423
424 def _init_weights(self, m):
425 if isinstance(m, nn.Linear):
426 trunc_normal_(m.weight, std=.02)
427 if isinstance(m, nn.Linear) and m.bias is not None:
428 nn.init.constant_(m.bias, 0)
429 elif isinstance(m, nn.LayerNorm):
430 nn.init.constant_(m.bias, 0)
431 nn.init.constant_(m.weight, 1.0)
432 elif isinstance(m, nn.Conv2d):
433 fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
434 fan_out //= m.groups
435 m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
436 if m.bias is not None:
437 m.bias.data.zero_()
438
439 def init_weights(self, pretrained=None):
440 if isinstance(pretrained, str):
441 logger = 1
442 #load_checkpoint(self, pretrained, map_location='cpu', strict=False, logger=logger)
443
444 def reset_drop_path(self, drop_path_rate):
445 dpr = np.linspace(0, drop_path_rate, sum(self.depths)).tolist()
446 cur = 0
447 for i in range(self.depths[0]):
448 self.block1[i].drop_path.drop_prob = dpr[cur + i]
449
450 cur += self.depths[0]
451 for i in range(self.depths[1]):
452 self.block2[i].drop_path.drop_prob = dpr[cur + i]
453
454 cur += self.depths[1]
455 for i in range(self.depths[2]):
456 self.block3[i].drop_path.drop_prob = dpr[cur + i]
457
458 cur += self.depths[2]
459 for i in range(self.depths[3]):
460 self.block4[i].drop_path.drop_prob = dpr[cur + i]
461
462 def freeze_patch_emb(self):
463 self.patch_embed1.requires_grad = False
464
465 @torch.jit.ignore
466 def no_weight_decay(self):
467 return {'pos_embed1', 'pos_embed2', 'pos_embed3', 'pos_embed4', 'cls_token'} # has pos_embed may be better
468
469 def get_classifier(self):
470 return self.head
471
472 def reset_classifier(self, num_classes, global_pool=''):
473 self.num_classes = num_classes
474 self.head = nn.Linear(self.embed_dim, num_classes) if num_classes > 0 else nn.Identity()
475
476 def forward_features(self, x):
477 B = x.shape[0]
478 outs = []
479
480 # stage 1
481 x, H, W = self.patch_embed1(x)
482 for i, blk in enumerate(self.block1):
483 x = blk(x, H, W)
484 x = self.norm1(x)
485 x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous()
486 outs.append(x)
487
488 # stage 2
489 x, H, W = self.patch_embed2(x)
490 for i, blk in enumerate(self.block2):
491 x = blk(x, H, W)
492 x = self.norm2(x)
493 x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous()
494 outs.append(x)
495
496 # stage 3
497 x, H, W = self.patch_embed3(x)
498 for i, blk in enumerate(self.block3):
499 x = blk(x, H, W)
500 x = self.norm3(x)
501 x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous()
502 outs.append(x)
503
504 # stage 4
505 x, H, W = self.patch_embed4(x)
506 for i, blk in enumerate(self.block4):
507 x = blk(x, H, W)
508 x = self.norm4(x)
509 x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous()
510 outs.append(x)
511
512 return outs
513
514 # return x.mean(dim=1)
515
516 def forward(self, x):
517 x = self.forward_features(x)
518 # x = self.head(x)
519
520 return x
521
522
523 class DWConv(nn.Module):
524 def __init__(self, dim=768):
525 super(DWConv, self).__init__()
526 self.dwconv = nn.Conv2d(dim, dim, 3, 1, 1, bias=True, groups=dim)
527
528 def forward(self, x, H, W):
529 B, N, C = x.shape
530 x = x.transpose(1, 2).view(B, C, H, W).contiguous()
531 x = self.dwconv(x)
532 x = x.flatten(2).transpose(1, 2)
533
534 return x
535
536
537 def _conv_filter(state_dict, patch_size=16):
538 """ convert patch embedding weight from manual patchify + linear proj to conv"""
539 out_dict = {}
540 for k, v in state_dict.items():
541 if 'patch_embed.proj.weight' in k:
542 v = v.reshape((v.shape[0], 3, patch_size, patch_size))
543 out_dict[k] = v
544
545 return out_dict
546
547
548 class pvt_v2_b0(PyramidVisionTransformerImpr):
549 def __init__(self, **kwargs):
550 super(pvt_v2_b0, self).__init__(
551 patch_size=4, embed_dims=[32, 64, 160, 256], num_heads=[1, 2, 5, 8], mlp_ratios=[8, 8, 4, 4],
552 qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[2, 2, 2, 2], sr_ratios=[8, 4, 2, 1],
553 drop_rate=0.0, drop_path_rate=0.1)
554
555
556
557 class pvt_v2_b1(PyramidVisionTransformerImpr):
558 def __init__(self, **kwargs):
559 super(pvt_v2_b1, self).__init__(
560 patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[8, 8, 4, 4],
561 qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[2, 2, 2, 2], sr_ratios=[8, 4, 2, 1],
562 drop_rate=0.0, drop_path_rate=0.1)
563
564 class pvt_v2_b2(PyramidVisionTransformerImpr):
565 def __init__(self, in_channels=3, **kwargs):
566 super(pvt_v2_b2, self).__init__(
567 patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[8, 8, 4, 4],
568 qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[3, 4, 6, 3], sr_ratios=[8, 4, 2, 1],
569 drop_rate=0.0, drop_path_rate=0.1, in_channels=in_channels)
570
571 class pvt_v2_b3(PyramidVisionTransformerImpr):
572 def __init__(self, **kwargs):
573 super(pvt_v2_b3, self).__init__(
574 patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[8, 8, 4, 4],
575 qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[3, 4, 18, 3], sr_ratios=[8, 4, 2, 1],
576 drop_rate=0.0, drop_path_rate=0.1)
577
578 class pvt_v2_b4(PyramidVisionTransformerImpr):
579 def __init__(self, **kwargs):
580 super(pvt_v2_b4, self).__init__(
581 patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[8, 8, 4, 4],
582 qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[3, 8, 27, 3], sr_ratios=[8, 4, 2, 1],
583 drop_rate=0.0, drop_path_rate=0.1)
584
585
586 class pvt_v2_b5(PyramidVisionTransformerImpr):
587 def __init__(self, **kwargs):
588 super(pvt_v2_b5, self).__init__(
589 patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[4, 4, 4, 4],
590 qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[3, 6, 40, 3], sr_ratios=[8, 4, 2, 1],
591 drop_rate=0.0, drop_path_rate=0.1)
592
593
594
595 ### models/backbones/swin_v1.py
596
597 # --------------------------------------------------------
598 # Swin Transformer
599 # Copyright (c) 2021 Microsoft
600 # Licensed under The MIT License [see LICENSE for details]
601 # Written by Ze Liu, Yutong Lin, Yixuan Wei
602 # --------------------------------------------------------
603
604 import torch
605 import torch.nn as nn
606 import torch.nn.functional as F
607 import torch.utils.checkpoint as checkpoint
608 import numpy as np
609 from timm.layers import DropPath, to_2tuple, trunc_normal_
610
611 # from config import Config
612
613
614 # config = Config()
615
616
617 class Mlp(nn.Module):
618 """ Multilayer perceptron."""
619
620 def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):
621 super().__init__()
622 out_features = out_features or in_features
623 hidden_features = hidden_features or in_features
624 self.fc1 = nn.Linear(in_features, hidden_features)
625 self.act = act_layer()
626 self.fc2 = nn.Linear(hidden_features, out_features)
627 self.drop = nn.Dropout(drop)
628
629 def forward(self, x):
630 x = self.fc1(x)
631 x = self.act(x)
632 x = self.drop(x)
633 x = self.fc2(x)
634 x = self.drop(x)
635 return x
636
637
638 def window_partition(x, window_size):
639 """
640 Args:
641 x: (B, H, W, C)
642 window_size (int): window size
643
644 Returns:
645 windows: (num_windows*B, window_size, window_size, C)
646 """
647 B, H, W, C = x.shape
648 x = x.view(B, H // window_size, window_size, W // window_size, window_size, C)
649 windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C)
650 return windows
651
652
653 def window_reverse(windows, window_size, H, W):
654 """
655 Args:
656 windows: (num_windows*B, window_size, window_size, C)
657 window_size (int): Window size
658 H (int): Height of image
659 W (int): Width of image
660
661 Returns:
662 x: (B, H, W, C)
663 """
664 B = int(windows.shape[0] / (H * W / window_size / window_size))
665 x = windows.view(B, H // window_size, W // window_size, window_size, window_size, -1)
666 x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, -1)
667 return x
668
669
670 class WindowAttention(nn.Module):
671 """ Window based multi-head self attention (W-MSA) module with relative position bias.
672 It supports both of shifted and non-shifted window.
673
674 Args:
675 dim (int): Number of input channels.
676 window_size (tuple[int]): The height and width of the window.
677 num_heads (int): Number of attention heads.
678 qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
679 qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set
680 attn_drop (float, optional): Dropout ratio of attention weight. Default: 0.0
681 proj_drop (float, optional): Dropout ratio of output. Default: 0.0
682 """
683
684 def __init__(self, dim, window_size, num_heads, qkv_bias=True, qk_scale=None, attn_drop=0., proj_drop=0.):
685
686 super().__init__()
687 self.dim = dim
688 self.window_size = window_size # Wh, Ww
689 self.num_heads = num_heads
690 head_dim = dim // num_heads
691 self.scale = qk_scale or head_dim ** -0.5
692
693 # define a parameter table of relative position bias
694 self.relative_position_bias_table = nn.Parameter(
695 torch.zeros((2 * window_size[0] - 1) * (2 * window_size[1] - 1), num_heads)) # 2*Wh-1 * 2*Ww-1, nH
696
697 # get pair-wise relative position index for each token inside the window
698 coords_h = torch.arange(self.window_size[0])
699 coords_w = torch.arange(self.window_size[1])
700 coords = torch.stack(torch.meshgrid([coords_h, coords_w], indexing='ij')) # 2, Wh, Ww
701 coords_flatten = torch.flatten(coords, 1) # 2, Wh*Ww
702 relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :] # 2, Wh*Ww, Wh*Ww
703 relative_coords = relative_coords.permute(1, 2, 0).contiguous() # Wh*Ww, Wh*Ww, 2
704 relative_coords[:, :, 0] += self.window_size[0] - 1 # shift to start from 0
705 relative_coords[:, :, 1] += self.window_size[1] - 1
706 relative_coords[:, :, 0] *= 2 * self.window_size[1] - 1
707 relative_position_index = relative_coords.sum(-1) # Wh*Ww, Wh*Ww
708 self.register_buffer("relative_position_index", relative_position_index)
709
710 self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
711 self.attn_drop_prob = attn_drop
712 self.attn_drop = nn.Dropout(attn_drop)
713 self.proj = nn.Linear(dim, dim)
714 self.proj_drop = nn.Dropout(proj_drop)
715
716 trunc_normal_(self.relative_position_bias_table, std=.02)
717 self.softmax = nn.Softmax(dim=-1)
718
719 def forward(self, x, mask=None):
720 """ Forward function.
721
722 Args:
723 x: input features with shape of (num_windows*B, N, C)
724 mask: (0/-inf) mask with shape of (num_windows, Wh*Ww, Wh*Ww) or None
725 """
726 B_, N, C = x.shape
727 qkv = self.qkv(x).reshape(B_, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
728 q, k, v = qkv[0], qkv[1], qkv[2] # make torchscript happy (cannot use tensor as tuple)
729
730 q = q * self.scale
731
732 if config.SDPA_enabled:
733 x = torch.nn.functional.scaled_dot_product_attention(
734 q, k, v,
735 attn_mask=None, dropout_p=self.attn_drop_prob, is_causal=False
736 ).transpose(1, 2).reshape(B_, N, C)
737 else:
738 attn = (q @ k.transpose(-2, -1))
739
740 relative_position_bias = self.relative_position_bias_table[self.relative_position_index.view(-1)].view(
741 self.window_size[0] * self.window_size[1], self.window_size[0] * self.window_size[1], -1
742 ) # Wh*Ww, Wh*Ww, nH
743 relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous() # nH, Wh*Ww, Wh*Ww
744 attn = attn + relative_position_bias.unsqueeze(0)
745
746 if mask is not None:
747 nW = mask.shape[0]
748 attn = attn.view(B_ // nW, nW, self.num_heads, N, N) + mask.unsqueeze(1).unsqueeze(0)
749 attn = attn.view(-1, self.num_heads, N, N)
750 attn = self.softmax(attn)
751 else:
752 attn = self.softmax(attn)
753
754 attn = self.attn_drop(attn)
755
756 x = (attn @ v).transpose(1, 2).reshape(B_, N, C)
757 x = self.proj(x)
758 x = self.proj_drop(x)
759 return x
760
761
762 class SwinTransformerBlock(nn.Module):
763 """ Swin Transformer Block.
764
765 Args:
766 dim (int): Number of input channels.
767 num_heads (int): Number of attention heads.
768 window_size (int): Window size.
769 shift_size (int): Shift size for SW-MSA.
770 mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
771 qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
772 qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
773 drop (float, optional): Dropout rate. Default: 0.0
774 attn_drop (float, optional): Attention dropout rate. Default: 0.0
775 drop_path (float, optional): Stochastic depth rate. Default: 0.0
776 act_layer (nn.Module, optional): Activation layer. Default: nn.GELU
777 norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
778 """
779
780 def __init__(self, dim, num_heads, window_size=7, shift_size=0,
781 mlp_ratio=4., qkv_bias=True, qk_scale=None, drop=0., attn_drop=0., drop_path=0.,
782 act_layer=nn.GELU, norm_layer=nn.LayerNorm):
783 super().__init__()
784 self.dim = dim
785 self.num_heads = num_heads
786 self.window_size = window_size
787 self.shift_size = shift_size
788 self.mlp_ratio = mlp_ratio
789 assert 0 <= self.shift_size < self.window_size, "shift_size must in 0-window_size"
790
791 self.norm1 = norm_layer(dim)
792 self.attn = WindowAttention(
793 dim, window_size=to_2tuple(self.window_size), num_heads=num_heads,
794 qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop)
795
796 self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
797 self.norm2 = norm_layer(dim)
798 mlp_hidden_dim = int(dim * mlp_ratio)
799 self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
800
801 self.H = None
802 self.W = None
803
804 def forward(self, x, mask_matrix):
805 """ Forward function.
806
807 Args:
808 x: Input feature, tensor size (B, H*W, C).
809 H, W: Spatial resolution of the input feature.
810 mask_matrix: Attention mask for cyclic shift.
811 """
812 B, L, C = x.shape
813 H, W = self.H, self.W
814 assert L == H * W, "input feature has wrong size"
815
816 shortcut = x
817 x = self.norm1(x)
818 x = x.view(B, H, W, C)
819
820 # pad feature maps to multiples of window size
821 pad_l = pad_t = 0
822 pad_r = (self.window_size - W % self.window_size) % self.window_size
823 pad_b = (self.window_size - H % self.window_size) % self.window_size
824 x = F.pad(x, (0, 0, pad_l, pad_r, pad_t, pad_b))
825 _, Hp, Wp, _ = x.shape
826
827 # cyclic shift
828 if self.shift_size > 0:
829 shifted_x = torch.roll(x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2))
830 attn_mask = mask_matrix
831 else:
832 shifted_x = x
833 attn_mask = None
834
835 # partition windows
836 x_windows = window_partition(shifted_x, self.window_size) # nW*B, window_size, window_size, C
837 x_windows = x_windows.view(-1, self.window_size * self.window_size, C) # nW*B, window_size*window_size, C
838
839 # W-MSA/SW-MSA
840 attn_windows = self.attn(x_windows, mask=attn_mask) # nW*B, window_size*window_size, C
841
842 # merge windows
843 attn_windows = attn_windows.view(-1, self.window_size, self.window_size, C)
844 shifted_x = window_reverse(attn_windows, self.window_size, Hp, Wp) # B H' W' C
845
846 # reverse cyclic shift
847 if self.shift_size > 0:
848 x = torch.roll(shifted_x, shifts=(self.shift_size, self.shift_size), dims=(1, 2))
849 else:
850 x = shifted_x
851
852 if pad_r > 0 or pad_b > 0:
853 x = x[:, :H, :W, :].contiguous()
854
855 x = x.view(B, H * W, C)
856
857 # FFN
858 x = shortcut + self.drop_path(x)
859 x = x + self.drop_path(self.mlp(self.norm2(x)))
860
861 return x
862
863
864 class PatchMerging(nn.Module):
865 """ Patch Merging Layer
866
867 Args:
868 dim (int): Number of input channels.
869 norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
870 """
871 def __init__(self, dim, norm_layer=nn.LayerNorm):
872 super().__init__()
873 self.dim = dim
874 self.reduction = nn.Linear(4 * dim, 2 * dim, bias=False)
875 self.norm = norm_layer(4 * dim)
876
877 def forward(self, x, H, W):
878 """ Forward function.
879
880 Args:
881 x: Input feature, tensor size (B, H*W, C).
882 H, W: Spatial resolution of the input feature.
883 """
884 B, L, C = x.shape
885 assert L == H * W, "input feature has wrong size"
886
887 x = x.view(B, H, W, C)
888
889 # padding
890 pad_input = (H % 2 == 1) or (W % 2 == 1)
891 if pad_input:
892 x = F.pad(x, (0, 0, 0, W % 2, 0, H % 2))
893
894 x0 = x[:, 0::2, 0::2, :] # B H/2 W/2 C
895 x1 = x[:, 1::2, 0::2, :] # B H/2 W/2 C
896 x2 = x[:, 0::2, 1::2, :] # B H/2 W/2 C
897 x3 = x[:, 1::2, 1::2, :] # B H/2 W/2 C
898 x = torch.cat([x0, x1, x2, x3], -1) # B H/2 W/2 4*C
899 x = x.view(B, -1, 4 * C) # B H/2*W/2 4*C
900
901 x = self.norm(x)
902 x = self.reduction(x)
903
904 return x
905
906
907 class BasicLayer(nn.Module):
908 """ A basic Swin Transformer layer for one stage.
909
910 Args:
911 dim (int): Number of feature channels
912 depth (int): Depths of this stage.
913 num_heads (int): Number of attention head.
914 window_size (int): Local window size. Default: 7.
915 mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4.
916 qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
917 qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
918 drop (float, optional): Dropout rate. Default: 0.0
919 attn_drop (float, optional): Attention dropout rate. Default: 0.0
920 drop_path (float | tuple[float], optional): Stochastic depth rate. Default: 0.0
921 norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
922 downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default: None
923 use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False.
924 """
925
926 def __init__(self,
927 dim,
928 depth,
929 num_heads,
930 window_size=7,
931 mlp_ratio=4.,
932 qkv_bias=True,
933 qk_scale=None,
934 drop=0.,
935 attn_drop=0.,
936 drop_path=0.,
937 norm_layer=nn.LayerNorm,
938 downsample=None,
939 use_checkpoint=False):
940 super().__init__()
941 self.window_size = window_size
942 self.shift_size = window_size // 2
943 self.depth = depth
944 self.use_checkpoint = use_checkpoint
945
946 # build blocks
947 self.blocks = nn.ModuleList([
948 SwinTransformerBlock(
949 dim=dim,
950 num_heads=num_heads,
951 window_size=window_size,
952 shift_size=0 if (i % 2 == 0) else window_size // 2,
953 mlp_ratio=mlp_ratio,
954 qkv_bias=qkv_bias,
955 qk_scale=qk_scale,
956 drop=drop,
957 attn_drop=attn_drop,
958 drop_path=drop_path[i] if isinstance(drop_path, list) else drop_path,
959 norm_layer=norm_layer)
960 for i in range(depth)])
961
962 # patch merging layer
963 if downsample is not None:
964 self.downsample = downsample(dim=dim, norm_layer=norm_layer)
965 else:
966 self.downsample = None
967
968 def forward(self, x, H, W):
969 """ Forward function.
970
971 Args:
972 x: Input feature, tensor size (B, H*W, C).
973 H, W: Spatial resolution of the input feature.
974 """
975
976 # calculate attention mask for SW-MSA
977 # Turn int to torch.tensor for the compatiability with torch.compile in PyTorch 2.5.
978 Hp = torch.ceil(torch.tensor(H) / self.window_size).to(torch.int64) * self.window_size
979 Wp = torch.ceil(torch.tensor(W) / self.window_size).to(torch.int64) * self.window_size
980 img_mask = torch.zeros((1, Hp, Wp, 1), device=x.device) # 1 Hp Wp 1
981 h_slices = (slice(0, -self.window_size),
982 slice(-self.window_size, -self.shift_size),
983 slice(-self.shift_size, None))
984 w_slices = (slice(0, -self.window_size),
985 slice(-self.window_size, -self.shift_size),
986 slice(-self.shift_size, None))
987 cnt = 0
988 for h in h_slices:
989 for w in w_slices:
990 img_mask[:, h, w, :] = cnt
991 cnt += 1
992
993 mask_windows = window_partition(img_mask, self.window_size) # nW, window_size, window_size, 1
994 mask_windows = mask_windows.view(-1, self.window_size * self.window_size)
995 attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2)
996 attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(attn_mask == 0, float(0.0)).to(x.dtype)
997
998 for blk in self.blocks:
999 blk.H, blk.W = H, W
1000 if self.use_checkpoint:
1001 x = checkpoint.checkpoint(blk, x, attn_mask)
1002 else:
1003 x = blk(x, attn_mask)
1004 if self.downsample is not None:
1005 x_down = self.downsample(x, H, W)
1006 Wh, Ww = (H + 1) // 2, (W + 1) // 2
1007 return x, H, W, x_down, Wh, Ww
1008 else:
1009 return x, H, W, x, H, W
1010
1011
1012 class PatchEmbed(nn.Module):
1013 """ Image to Patch Embedding
1014
1015 Args:
1016 patch_size (int): Patch token size. Default: 4.
1017 in_channels (int): Number of input image channels. Default: 3.
1018 embed_dim (int): Number of linear projection output channels. Default: 96.
1019 norm_layer (nn.Module, optional): Normalization layer. Default: None
1020 """
1021
1022 def __init__(self, patch_size=4, in_channels=3, embed_dim=96, norm_layer=None):
1023 super().__init__()
1024 patch_size = to_2tuple(patch_size)
1025 self.patch_size = patch_size
1026
1027 self.in_channels = in_channels
1028 self.embed_dim = embed_dim
1029
1030 self.proj = nn.Conv2d(in_channels, embed_dim, kernel_size=patch_size, stride=patch_size)
1031 if norm_layer is not None:
1032 self.norm = norm_layer(embed_dim)
1033 else:
1034 self.norm = None
1035
1036 def forward(self, x):
1037 """Forward function."""
1038 # padding
1039 _, _, H, W = x.size()
1040 if W % self.patch_size[1] != 0:
1041 x = F.pad(x, (0, self.patch_size[1] - W % self.patch_size[1]))
1042 if H % self.patch_size[0] != 0:
1043 x = F.pad(x, (0, 0, 0, self.patch_size[0] - H % self.patch_size[0]))
1044
1045 x = self.proj(x) # B C Wh Ww
1046 if self.norm is not None:
1047 Wh, Ww = x.size(2), x.size(3)
1048 x = x.flatten(2).transpose(1, 2)
1049 x = self.norm(x)
1050 x = x.transpose(1, 2).view(-1, self.embed_dim, Wh, Ww)
1051
1052 return x
1053
1054
1055 class SwinTransformer(nn.Module):
1056 """ Swin Transformer backbone.
1057 A PyTorch impl of : `Swin Transformer: Hierarchical Vision Transformer using Shifted Windows` -
1058 https://arxiv.org/pdf/2103.14030
1059
1060 Args:
1061 pretrain_img_size (int): Input image size for training the pretrained model,
1062 used in absolute postion embedding. Default 224.
1063 patch_size (int | tuple(int)): Patch size. Default: 4.
1064 in_channels (int): Number of input image channels. Default: 3.
1065 embed_dim (int): Number of linear projection output channels. Default: 96.
1066 depths (tuple[int]): Depths of each Swin Transformer stage.
1067 num_heads (tuple[int]): Number of attention head of each stage.
1068 window_size (int): Window size. Default: 7.
1069 mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4.
1070 qkv_bias (bool): If True, add a learnable bias to query, key, value. Default: True
1071 qk_scale (float): Override default qk scale of head_dim ** -0.5 if set.
1072 drop_rate (float): Dropout rate.
1073 attn_drop_rate (float): Attention dropout rate. Default: 0.
1074 drop_path_rate (float): Stochastic depth rate. Default: 0.2.
1075 norm_layer (nn.Module): Normalization layer. Default: nn.LayerNorm.
1076 ape (bool): If True, add absolute position embedding to the patch embedding. Default: False.
1077 patch_norm (bool): If True, add normalization after patch embedding. Default: True.
1078 out_indices (Sequence[int]): Output from which stages.
1079 frozen_stages (int): Stages to be frozen (stop grad and set eval mode).
1080 -1 means not freezing any parameters.
1081 use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False.
1082 """
1083
1084 def __init__(self,
1085 pretrain_img_size=224,
1086 patch_size=4,
1087 in_channels=3,
1088 embed_dim=96,
1089 depths=[2, 2, 6, 2],
1090 num_heads=[3, 6, 12, 24],
1091 window_size=7,
1092 mlp_ratio=4.,
1093 qkv_bias=True,
1094 qk_scale=None,
1095 drop_rate=0.,
1096 attn_drop_rate=0.,
1097 drop_path_rate=0.2,
1098 norm_layer=nn.LayerNorm,
1099 ape=False,
1100 patch_norm=True,
1101 out_indices=(0, 1, 2, 3),
1102 frozen_stages=-1,
1103 use_checkpoint=False):
1104 super().__init__()
1105
1106 self.pretrain_img_size = pretrain_img_size
1107 self.num_layers = len(depths)
1108 self.embed_dim = embed_dim
1109 self.ape = ape
1110 self.patch_norm = patch_norm
1111 self.out_indices = out_indices
1112 self.frozen_stages = frozen_stages
1113
1114 # split image into non-overlapping patches
1115 self.patch_embed = PatchEmbed(
1116 patch_size=patch_size, in_channels=in_channels, embed_dim=embed_dim,
1117 norm_layer=norm_layer if self.patch_norm else None)
1118
1119 # absolute position embedding
1120 if self.ape:
1121 pretrain_img_size = to_2tuple(pretrain_img_size)
1122 patch_size = to_2tuple(patch_size)
1123 patches_resolution = [pretrain_img_size[0] // patch_size[0], pretrain_img_size[1] // patch_size[1]]
1124
1125 self.absolute_pos_embed = nn.Parameter(torch.zeros(1, embed_dim, patches_resolution[0], patches_resolution[1]))
1126 trunc_normal_(self.absolute_pos_embed, std=.02)
1127
1128 self.pos_drop = nn.Dropout(p=drop_rate)
1129
1130 # stochastic depth
1131 dpr = np.linspace(0, drop_path_rate, sum(depths)).tolist() # stochastic depth decay rule
1132
1133 # build layers
1134 self.layers = nn.ModuleList()
1135 for i_layer in range(self.num_layers):
1136 layer = BasicLayer(
1137 dim=int(embed_dim * 2 ** i_layer),
1138 depth=depths[i_layer],
1139 num_heads=num_heads[i_layer],
1140 window_size=window_size,
1141 mlp_ratio=mlp_ratio,
1142 qkv_bias=qkv_bias,
1143 qk_scale=qk_scale,
1144 drop=drop_rate,
1145 attn_drop=attn_drop_rate,
1146 drop_path=dpr[sum(depths[:i_layer]):sum(depths[:i_layer + 1])],
1147 norm_layer=norm_layer,
1148 downsample=PatchMerging if (i_layer < self.num_layers - 1) else None,
1149 use_checkpoint=use_checkpoint)
1150 self.layers.append(layer)
1151
1152 num_features = [int(embed_dim * 2 ** i) for i in range(self.num_layers)]
1153 self.num_features = num_features
1154
1155 # add a norm layer for each output
1156 for i_layer in out_indices:
1157 layer = norm_layer(num_features[i_layer])
1158 layer_name = f'norm{i_layer}'
1159 self.add_module(layer_name, layer)
1160
1161 self._freeze_stages()
1162
1163 def _freeze_stages(self):
1164 if self.frozen_stages >= 0:
1165 self.patch_embed.eval()
1166 for param in self.patch_embed.parameters():
1167 param.requires_grad = False
1168
1169 if self.frozen_stages >= 1 and self.ape:
1170 self.absolute_pos_embed.requires_grad = False
1171
1172 if self.frozen_stages >= 2:
1173 self.pos_drop.eval()
1174 for i in range(0, self.frozen_stages - 1):
1175 m = self.layers[i]
1176 m.eval()
1177 for param in m.parameters():
1178 param.requires_grad = False
1179
1180
1181 def forward(self, x):
1182 """Forward function."""
1183 x = self.patch_embed(x)
1184
1185 Wh, Ww = x.size(2), x.size(3)
1186 if self.ape:
1187 # interpolate the position embedding to the corresponding size
1188 absolute_pos_embed = F.interpolate(self.absolute_pos_embed, size=(Wh, Ww), mode='bicubic')
1189 x = (x + absolute_pos_embed) # B Wh*Ww C
1190
1191 outs = []#x.contiguous()]
1192 x = x.flatten(2).transpose(1, 2)
1193 x = self.pos_drop(x)
1194 for i in range(self.num_layers):
1195 layer = self.layers[i]
1196 x_out, H, W, x, Wh, Ww = layer(x, Wh, Ww)
1197
1198 if i in self.out_indices:
1199 norm_layer = getattr(self, f'norm{i}')
1200 x_out = norm_layer(x_out)
1201
1202 out = x_out.view(-1, H, W, self.num_features[i]).permute(0, 3, 1, 2).contiguous()
1203 outs.append(out)
1204
1205 return tuple(outs)
1206
1207 def train(self, mode=True):
1208 """Convert the model into training mode while keep layers freezed."""
1209 super(SwinTransformer, self).train(mode)
1210 self._freeze_stages()
1211
1212 def swin_v1_t():
1213 model = SwinTransformer(embed_dim=96, depths=[2, 2, 6, 2], num_heads=[3, 6, 12, 24], window_size=7)
1214 return model
1215
1216 def swin_v1_s():
1217 model = SwinTransformer(embed_dim=96, depths=[2, 2, 18, 2], num_heads=[3, 6, 12, 24], window_size=7)
1218 return model
1219
1220 def swin_v1_b():
1221 model = SwinTransformer(embed_dim=128, depths=[2, 2, 18, 2], num_heads=[4, 8, 16, 32], window_size=12)
1222 return model
1223
1224 def swin_v1_l():
1225 model = SwinTransformer(embed_dim=192, depths=[2, 2, 18, 2], num_heads=[6, 12, 24, 48], window_size=12)
1226 return model
1227
1228
1229
1230 ### models/modules/deform_conv.py
1231
1232 import torch
1233 import torch.nn as nn
1234 from torchvision.ops import deform_conv2d
1235
1236
1237 class DeformableConv2d(nn.Module):
1238 def __init__(self,
1239 in_channels,
1240 out_channels,
1241 kernel_size=3,
1242 stride=1,
1243 padding=1,
1244 bias=False):
1245
1246 super(DeformableConv2d, self).__init__()
1247
1248 assert type(kernel_size) == tuple or type(kernel_size) == int
1249
1250 kernel_size = kernel_size if type(kernel_size) == tuple else (kernel_size, kernel_size)
1251 self.stride = stride if type(stride) == tuple else (stride, stride)
1252 self.padding = padding
1253
1254 self.offset_conv = nn.Conv2d(in_channels,
1255 2 * kernel_size[0] * kernel_size[1],
1256 kernel_size=kernel_size,
1257 stride=stride,
1258 padding=self.padding,
1259 bias=True)
1260
1261 nn.init.constant_(self.offset_conv.weight, 0.)
1262 nn.init.constant_(self.offset_conv.bias, 0.)
1263
1264 self.modulator_conv = nn.Conv2d(in_channels,
1265 1 * kernel_size[0] * kernel_size[1],
1266 kernel_size=kernel_size,
1267 stride=stride,
1268 padding=self.padding,
1269 bias=True)
1270
1271 nn.init.constant_(self.modulator_conv.weight, 0.)
1272 nn.init.constant_(self.modulator_conv.bias, 0.)
1273
1274 self.regular_conv = nn.Conv2d(in_channels,
1275 out_channels=out_channels,
1276 kernel_size=kernel_size,
1277 stride=stride,
1278 padding=self.padding,
1279 bias=bias)
1280
1281 def forward(self, x):
1282 #h, w = x.shape[2:]
1283 #max_offset = max(h, w)/4.
1284
1285 offset = self.offset_conv(x)#.clamp(-max_offset, max_offset)
1286 modulator = 2. * torch.sigmoid(self.modulator_conv(x))
1287
1288 x = deform_conv2d(
1289 input=x,
1290 offset=offset,
1291 weight=self.regular_conv.weight,
1292 bias=self.regular_conv.bias,
1293 padding=self.padding,
1294 mask=modulator,
1295 stride=self.stride,
1296 )
1297 return x
1298
1299
1300
1301
1302 ### utils.py
1303
1304 import torch.nn as nn
1305
1306
1307 def build_act_layer(act_layer):
1308 if act_layer == 'ReLU':
1309 return nn.ReLU(inplace=True)
1310 elif act_layer == 'SiLU':
1311 return nn.SiLU(inplace=True)
1312 elif act_layer == 'GELU':
1313 return nn.GELU()
1314
1315 raise NotImplementedError(f'build_act_layer does not support {act_layer}')
1316
1317
1318 def build_norm_layer(dim,
1319 norm_layer,
1320 in_format='channels_last',
1321 out_format='channels_last',
1322 eps=1e-6):
1323 layers = []
1324 if norm_layer == 'BN':
1325 if in_format == 'channels_last':
1326 layers.append(to_channels_first())
1327 layers.append(nn.BatchNorm2d(dim))
1328 if out_format == 'channels_last':
1329 layers.append(to_channels_last())
1330 elif norm_layer == 'LN':
1331 if in_format == 'channels_first':
1332 layers.append(to_channels_last())
1333 layers.append(nn.LayerNorm(dim, eps=eps))
1334 if out_format == 'channels_first':
1335 layers.append(to_channels_first())
1336 else:
1337 raise NotImplementedError(
1338 f'build_norm_layer does not support {norm_layer}')
1339 return nn.Sequential(*layers)
1340
1341
1342 class to_channels_first(nn.Module):
1343
1344 def __init__(self):
1345 super().__init__()
1346
1347 def forward(self, x):
1348 return x.permute(0, 3, 1, 2)
1349
1350
1351 class to_channels_last(nn.Module):
1352
1353 def __init__(self):
1354 super().__init__()
1355
1356 def forward(self, x):
1357 return x.permute(0, 2, 3, 1)
1358
1359
1360
1361 ### dataset.py
1362
1363 _class_labels_TR_sorted = (
1364 'Airplane, Ant, Antenna, Archery, Axe, BabyCarriage, Bag, BalanceBeam, Balcony, Balloon, Basket, BasketballHoop, Beatle, Bed, Bee, Bench, Bicycle, '
1365 'BicycleFrame, BicycleStand, Boat, Bonsai, BoomLift, Bridge, BunkBed, Butterfly, Button, Cable, CableLift, Cage, Camcorder, Cannon, Canoe, Car, '
1366 'CarParkDropArm, Carriage, Cart, Caterpillar, CeilingLamp, Centipede, Chair, Clip, Clock, Clothes, CoatHanger, Comb, ConcretePumpTruck, Crack, Crane, '
1367 'Cup, DentalChair, Desk, DeskChair, Diagram, DishRack, DoorHandle, Dragonfish, Dragonfly, Drum, Earphone, Easel, ElectricIron, Excavator, Eyeglasses, '
1368 'Fan, Fence, Fencing, FerrisWheel, FireExtinguisher, Fishing, Flag, FloorLamp, Forklift, GasStation, Gate, Gear, Goal, Golf, GymEquipment, Hammock, '
1369 'Handcart, Handcraft, Handrail, HangGlider, Harp, Harvester, Headset, Helicopter, Helmet, Hook, HorizontalBar, Hydrovalve, IroningTable, Jewelry, Key, '
1370 'KidsPlayground, Kitchenware, Kite, Knife, Ladder, LaundryRack, Lightning, Lobster, Locust, Machine, MachineGun, MagazineRack, Mantis, Medal, MemorialArchway, '
1371 'Microphone, Missile, MobileHolder, Monitor, Mosquito, Motorcycle, MovingTrolley, Mower, MusicPlayer, MusicStand, ObservationTower, Octopus, OilWell, '
1372 'OlympicLogo, OperatingTable, OutdoorFitnessEquipment, Parachute, Pavilion, Piano, Pipe, PlowHarrow, PoleVault, Punchbag, Rack, Racket, Rifle, Ring, Robot, '
1373 'RockClimbing, Rope, Sailboat, Satellite, Scaffold, Scale, Scissor, Scooter, Sculpture, Seadragon, Seahorse, Seal, SewingMachine, Ship, Shoe, ShoppingCart, '
1374 'ShoppingTrolley, Shower, Shrimp, Signboard, Skateboarding, Skeleton, Skiing, Spade, SpeedBoat, Spider, Spoon, Stair, Stand, Stationary, SteeringWheel, '
1375 'Stethoscope, Stool, Stove, StreetLamp, SweetStand, Swing, Sword, TV, Table, TableChair, TableLamp, TableTennis, Tank, Tapeline, Teapot, Telescope, Tent, '
1376 'TobaccoPipe, Toy, Tractor, TrafficLight, TrafficSign, Trampoline, TransmissionTower, Tree, Tricycle, TrimmerCover, Tripod, Trombone, Truck, Trumpet, Tuba, '
1377 'UAV, Umbrella, UnevenBars, UtilityPole, VacuumCleaner, Violin, Wakesurfing, Watch, WaterTower, WateringPot, Well, WellLid, Wheel, Wheelchair, WindTurbine, Windmill, WineGlass, WireWhisk, Yacht'
1378 )
1379 class_labels_TR_sorted = _class_labels_TR_sorted.split(', ')
1380
1381
1382 ### models/backbones/build_backbones.py
1383
1384 import torch
1385 import torch.nn as nn
1386 from collections import OrderedDict
1387 from torchvision.models import vgg16, vgg16_bn, VGG16_Weights, VGG16_BN_Weights, resnet50, ResNet50_Weights
1388 # from models.pvt_v2 import pvt_v2_b0, pvt_v2_b1, pvt_v2_b2, pvt_v2_b5
1389 # from models.swin_v1 import swin_v1_t, swin_v1_s, swin_v1_b, swin_v1_l
1390 # from config import Config
1391
1392
1393 config = Config()
1394
1395 def build_backbone(bb_name, pretrained=True, params_settings=''):
1396 if bb_name == 'vgg16':
1397 bb_net = list(vgg16(pretrained=VGG16_Weights.DEFAULT if pretrained else None).children())[0]
1398 bb = nn.Sequential(OrderedDict({'conv1': bb_net[:4], 'conv2': bb_net[4:9], 'conv3': bb_net[9:16], 'conv4': bb_net[16:23]}))
1399 elif bb_name == 'vgg16bn':
1400 bb_net = list(vgg16_bn(pretrained=VGG16_BN_Weights.DEFAULT if pretrained else None).children())[0]
1401 bb = nn.Sequential(OrderedDict({'conv1': bb_net[:6], 'conv2': bb_net[6:13], 'conv3': bb_net[13:23], 'conv4': bb_net[23:33]}))
1402 elif bb_name == 'resnet50':
1403 bb_net = list(resnet50(pretrained=ResNet50_Weights.DEFAULT if pretrained else None).children())
1404 bb = nn.Sequential(OrderedDict({'conv1': nn.Sequential(*bb_net[0:3]), 'conv2': bb_net[4], 'conv3': bb_net[5], 'conv4': bb_net[6]}))
1405 else:
1406 bb = eval('{}({})'.format(bb_name, params_settings))
1407 if pretrained:
1408 bb = load_weights(bb, bb_name)
1409 return bb
1410
1411 def load_weights(model, model_name):
1412 save_model = torch.load(config.weights[model_name], map_location='cpu')
1413 model_dict = model.state_dict()
1414 state_dict = {k: v if v.size() == model_dict[k].size() else model_dict[k] for k, v in save_model.items() if k in model_dict.keys()}
1415 # to ignore the weights with mismatched size when I modify the backbone itself.
1416 if not state_dict:
1417 save_model_keys = list(save_model.keys())
1418 sub_item = save_model_keys[0] if len(save_model_keys) == 1 else None
1419 state_dict = {k: v if v.size() == model_dict[k].size() else model_dict[k] for k, v in save_model[sub_item].items() if k in model_dict.keys()}
1420 if not state_dict or not sub_item:
1421 print('Weights are not successully loaded. Check the state dict of weights file.')
1422 return None
1423 else:
1424 print('Found correct weights in the "{}" item of loaded state_dict.'.format(sub_item))
1425 model_dict.update(state_dict)
1426 model.load_state_dict(model_dict)
1427 return model
1428
1429
1430
1431 ### models/modules/decoder_blocks.py
1432
1433 import torch
1434 import torch.nn as nn
1435 # from models.aspp import ASPP, ASPPDeformable
1436 # from config import Config
1437
1438
1439 # config = Config()
1440
1441
1442 class BasicDecBlk(nn.Module):
1443 def __init__(self, in_channels=64, out_channels=64, inter_channels=64):
1444 super(BasicDecBlk, self).__init__()
1445 inter_channels = in_channels // 4 if config.dec_channels_inter == 'adap' else 64
1446 self.conv_in = nn.Conv2d(in_channels, inter_channels, 3, 1, padding=1)
1447 self.relu_in = nn.ReLU(inplace=True)
1448 if config.dec_att == 'ASPP':
1449 self.dec_att = ASPP(in_channels=inter_channels)
1450 elif config.dec_att == 'ASPPDeformable':
1451 self.dec_att = ASPPDeformable(in_channels=inter_channels)
1452 self.conv_out = nn.Conv2d(inter_channels, out_channels, 3, 1, padding=1)
1453 self.bn_in = nn.BatchNorm2d(inter_channels) if config.batch_size > 1 else nn.Identity()
1454 self.bn_out = nn.BatchNorm2d(out_channels) if config.batch_size > 1 else nn.Identity()
1455
1456 def forward(self, x):
1457 x = self.conv_in(x)
1458 x = self.bn_in(x)
1459 x = self.relu_in(x)
1460 if hasattr(self, 'dec_att'):
1461 x = self.dec_att(x)
1462 x = self.conv_out(x)
1463 x = self.bn_out(x)
1464 return x
1465
1466
1467 class ResBlk(nn.Module):
1468 def __init__(self, in_channels=64, out_channels=None, inter_channels=64):
1469 super(ResBlk, self).__init__()
1470 if out_channels is None:
1471 out_channels = in_channels
1472 inter_channels = in_channels // 4 if config.dec_channels_inter == 'adap' else 64
1473
1474 self.conv_in = nn.Conv2d(in_channels, inter_channels, 3, 1, padding=1)
1475 self.bn_in = nn.BatchNorm2d(inter_channels) if config.batch_size > 1 else nn.Identity()
1476 self.relu_in = nn.ReLU(inplace=True)
1477
1478 if config.dec_att == 'ASPP':
1479 self.dec_att = ASPP(in_channels=inter_channels)
1480 elif config.dec_att == 'ASPPDeformable':
1481 self.dec_att = ASPPDeformable(in_channels=inter_channels)
1482
1483 self.conv_out = nn.Conv2d(inter_channels, out_channels, 3, 1, padding=1)
1484 self.bn_out = nn.BatchNorm2d(out_channels) if config.batch_size > 1 else nn.Identity()
1485
1486 self.conv_resi = nn.Conv2d(in_channels, out_channels, 1, 1, 0)
1487
1488 def forward(self, x):
1489 _x = self.conv_resi(x)
1490 x = self.conv_in(x)
1491 x = self.bn_in(x)
1492 x = self.relu_in(x)
1493 if hasattr(self, 'dec_att'):
1494 x = self.dec_att(x)
1495 x = self.conv_out(x)
1496 x = self.bn_out(x)
1497 return x + _x
1498
1499
1500
1501 ### models/modules/lateral_blocks.py
1502
1503 import numpy as np
1504 import torch
1505 import torch.nn as nn
1506 import torch.nn.functional as F
1507 from functools import partial
1508
1509 # from config import Config
1510
1511
1512 # config = Config()
1513
1514
1515 class BasicLatBlk(nn.Module):
1516 def __init__(self, in_channels=64, out_channels=64, inter_channels=64):
1517 super(BasicLatBlk, self).__init__()
1518 inter_channels = in_channels // 4 if config.dec_channels_inter == 'adap' else 64
1519 self.conv = nn.Conv2d(in_channels, out_channels, 1, 1, 0)
1520
1521 def forward(self, x):
1522 x = self.conv(x)
1523 return x
1524
1525
1526
1527 ### models/modules/aspp.py
1528
1529 import torch
1530 import torch.nn as nn
1531 import torch.nn.functional as F
1532 # from models.deform_conv import DeformableConv2d
1533 # from config import Config
1534
1535
1536 # config = Config()
1537
1538
1539 class _ASPPModule(nn.Module):
1540 def __init__(self, in_channels, planes, kernel_size, padding, dilation):
1541 super(_ASPPModule, self).__init__()
1542 self.atrous_conv = nn.Conv2d(in_channels, planes, kernel_size=kernel_size,
1543 stride=1, padding=padding, dilation=dilation, bias=False)
1544 self.bn = nn.BatchNorm2d(planes) if config.batch_size > 1 else nn.Identity()
1545 self.relu = nn.ReLU(inplace=True)
1546
1547 def forward(self, x):
1548 x = self.atrous_conv(x)
1549 x = self.bn(x)
1550
1551 return self.relu(x)
1552
1553
1554 class ASPP(nn.Module):
1555 def __init__(self, in_channels=64, out_channels=None, output_stride=16):
1556 super(ASPP, self).__init__()
1557 self.down_scale = 1
1558 if out_channels is None:
1559 out_channels = in_channels
1560 self.in_channelster = 256 // self.down_scale
1561 if output_stride == 16:
1562 dilations = [1, 6, 12, 18]
1563 elif output_stride == 8:
1564 dilations = [1, 12, 24, 36]
1565 else:
1566 raise NotImplementedError
1567
1568 self.aspp1 = _ASPPModule(in_channels, self.in_channelster, 1, padding=0, dilation=dilations[0])
1569 self.aspp2 = _ASPPModule(in_channels, self.in_channelster, 3, padding=dilations[1], dilation=dilations[1])
1570 self.aspp3 = _ASPPModule(in_channels, self.in_channelster, 3, padding=dilations[2], dilation=dilations[2])
1571 self.aspp4 = _ASPPModule(in_channels, self.in_channelster, 3, padding=dilations[3], dilation=dilations[3])
1572
1573 self.global_avg_pool = nn.Sequential(nn.AdaptiveAvgPool2d((1, 1)),
1574 nn.Conv2d(in_channels, self.in_channelster, 1, stride=1, bias=False),
1575 nn.BatchNorm2d(self.in_channelster) if config.batch_size > 1 else nn.Identity(),
1576 nn.ReLU(inplace=True))
1577 self.conv1 = nn.Conv2d(self.in_channelster * 5, out_channels, 1, bias=False)
1578 self.bn1 = nn.BatchNorm2d(out_channels) if config.batch_size > 1 else nn.Identity()
1579 self.relu = nn.ReLU(inplace=True)
1580 self.dropout = nn.Dropout(0.5)
1581
1582 def forward(self, x):
1583 x1 = self.aspp1(x)
1584 x2 = self.aspp2(x)
1585 x3 = self.aspp3(x)
1586 x4 = self.aspp4(x)
1587 x5 = self.global_avg_pool(x)
1588 x5 = F.interpolate(x5, size=x1.size()[2:], mode='bilinear', align_corners=True)
1589 x = torch.cat((x1, x2, x3, x4, x5), dim=1)
1590
1591 x = self.conv1(x)
1592 x = self.bn1(x)
1593 x = self.relu(x)
1594
1595 return self.dropout(x)
1596
1597
1598 ##################### Deformable
1599 class _ASPPModuleDeformable(nn.Module):
1600 def __init__(self, in_channels, planes, kernel_size, padding):
1601 super(_ASPPModuleDeformable, self).__init__()
1602 self.atrous_conv = DeformableConv2d(in_channels, planes, kernel_size=kernel_size,
1603 stride=1, padding=padding, bias=False)
1604 self.bn = nn.BatchNorm2d(planes) if config.batch_size > 1 else nn.Identity()
1605 self.relu = nn.ReLU(inplace=True)
1606
1607 def forward(self, x):
1608 x = self.atrous_conv(x)
1609 x = self.bn(x)
1610
1611 return self.relu(x)
1612
1613
1614 class ASPPDeformable(nn.Module):
1615 def __init__(self, in_channels, out_channels=None, parallel_block_sizes=[1, 3, 7]):
1616 super(ASPPDeformable, self).__init__()
1617 self.down_scale = 1
1618 if out_channels is None:
1619 out_channels = in_channels
1620 self.in_channelster = 256 // self.down_scale
1621
1622 self.aspp1 = _ASPPModuleDeformable(in_channels, self.in_channelster, 1, padding=0)
1623 self.aspp_deforms = nn.ModuleList([
1624 _ASPPModuleDeformable(in_channels, self.in_channelster, conv_size, padding=int(conv_size//2)) for conv_size in parallel_block_sizes
1625 ])
1626
1627 self.global_avg_pool = nn.Sequential(nn.AdaptiveAvgPool2d((1, 1)),
1628 nn.Conv2d(in_channels, self.in_channelster, 1, stride=1, bias=False),
1629 nn.BatchNorm2d(self.in_channelster) if config.batch_size > 1 else nn.Identity(),
1630 nn.ReLU(inplace=True))
1631 self.conv1 = nn.Conv2d(self.in_channelster * (2 + len(self.aspp_deforms)), out_channels, 1, bias=False)
1632 self.bn1 = nn.BatchNorm2d(out_channels) if config.batch_size > 1 else nn.Identity()
1633 self.relu = nn.ReLU(inplace=True)
1634 self.dropout = nn.Dropout(0.5)
1635
1636 def forward(self, x):
1637 x1 = self.aspp1(x)
1638 x_aspp_deforms = [aspp_deform(x) for aspp_deform in self.aspp_deforms]
1639 x5 = self.global_avg_pool(x)
1640 x5 = F.interpolate(x5, size=x1.size()[2:], mode='bilinear', align_corners=True)
1641 x = torch.cat((x1, *x_aspp_deforms, x5), dim=1)
1642
1643 x = self.conv1(x)
1644 x = self.bn1(x)
1645 x = self.relu(x)
1646
1647 return self.dropout(x)
1648
1649
1650
1651 ### models/refinement/refiner.py
1652
1653 import torch
1654 import torch.nn as nn
1655 from collections import OrderedDict
1656 import torch
1657 import torch.nn as nn
1658 import torch.nn.functional as F
1659 from torchvision.models import vgg16, vgg16_bn
1660 from torchvision.models import resnet50
1661
1662 # from config import Config
1663 # from dataset import class_labels_TR_sorted
1664 # from models.build_backbone import build_backbone
1665 # from models.decoder_blocks import BasicDecBlk
1666 # from models.lateral_blocks import BasicLatBlk
1667 # from models.ing import *
1668 # from models.stem_layer import StemLayer
1669
1670
1671 class RefinerPVTInChannels4(nn.Module):
1672 def __init__(self, in_channels=3+1):
1673 super(RefinerPVTInChannels4, self).__init__()
1674 self.config = Config()
1675 self.epoch = 1
1676 self.bb = build_backbone(self.config.bb, params_settings='in_channels=4')
1677
1678 lateral_channels_in_collection = {
1679 'vgg16': [512, 256, 128, 64], 'vgg16bn': [512, 256, 128, 64], 'resnet50': [1024, 512, 256, 64],
1680 'pvt_v2_b2': [512, 320, 128, 64], 'pvt_v2_b5': [512, 320, 128, 64],
1681 'swin_v1_b': [1024, 512, 256, 128], 'swin_v1_l': [1536, 768, 384, 192],
1682 }
1683 channels = lateral_channels_in_collection[self.config.bb]
1684 self.squeeze_module = BasicDecBlk(channels[0], channels[0])
1685
1686 self.decoder = Decoder(channels)
1687
1688 if 0:
1689 for key, value in self.named_parameters():
1690 if 'bb.' in key:
1691 value.requires_grad = False
1692
1693 def forward(self, x):
1694 if isinstance(x, list):
1695 x = torch.cat(x, dim=1)
1696 ########## Encoder ##########
1697 if self.config.bb in ['vgg16', 'vgg16bn', 'resnet50']:
1698 x1 = self.bb.conv1(x)
1699 x2 = self.bb.conv2(x1)
1700 x3 = self.bb.conv3(x2)
1701 x4 = self.bb.conv4(x3)
1702 else:
1703 x1, x2, x3, x4 = self.bb(x)
1704
1705 x4 = self.squeeze_module(x4)
1706
1707 ########## Decoder ##########
1708
1709 features = [x, x1, x2, x3, x4]
1710 scaled_preds = self.decoder(features)
1711
1712 return scaled_preds
1713
1714
1715 class Refiner(nn.Module):
1716 def __init__(self, in_channels=3+1):
1717 super(Refiner, self).__init__()
1718 self.config = Config()
1719 self.epoch = 1
1720 self.stem_layer = StemLayer(in_channels=in_channels, inter_channels=48, out_channels=3, norm_layer='BN' if self.config.batch_size > 1 else 'LN')
1721 self.bb = build_backbone(self.config.bb)
1722
1723 lateral_channels_in_collection = {
1724 'vgg16': [512, 256, 128, 64], 'vgg16bn': [512, 256, 128, 64], 'resnet50': [1024, 512, 256, 64],
1725 'pvt_v2_b2': [512, 320, 128, 64], 'pvt_v2_b5': [512, 320, 128, 64],
1726 'swin_v1_b': [1024, 512, 256, 128], 'swin_v1_l': [1536, 768, 384, 192],
1727 }
1728 channels = lateral_channels_in_collection[self.config.bb]
1729 self.squeeze_module = BasicDecBlk(channels[0], channels[0])
1730
1731 self.decoder = Decoder(channels)
1732
1733 if 0:
1734 for key, value in self.named_parameters():
1735 if 'bb.' in key:
1736 value.requires_grad = False
1737
1738 def forward(self, x):
1739 if isinstance(x, list):
1740 x = torch.cat(x, dim=1)
1741 x = self.stem_layer(x)
1742 ########## Encoder ##########
1743 if self.config.bb in ['vgg16', 'vgg16bn', 'resnet50']:
1744 x1 = self.bb.conv1(x)
1745 x2 = self.bb.conv2(x1)
1746 x3 = self.bb.conv3(x2)
1747 x4 = self.bb.conv4(x3)
1748 else:
1749 x1, x2, x3, x4 = self.bb(x)
1750
1751 x4 = self.squeeze_module(x4)
1752
1753 ########## Decoder ##########
1754
1755 features = [x, x1, x2, x3, x4]
1756 scaled_preds = self.decoder(features)
1757
1758 return scaled_preds
1759
1760
1761 class Decoder(nn.Module):
1762 def __init__(self, channels):
1763 super(Decoder, self).__init__()
1764 self.config = Config()
1765 DecoderBlock = eval('BasicDecBlk')
1766 LateralBlock = eval('BasicLatBlk')
1767
1768 self.decoder_block4 = DecoderBlock(channels[0], channels[1])
1769 self.decoder_block3 = DecoderBlock(channels[1], channels[2])
1770 self.decoder_block2 = DecoderBlock(channels[2], channels[3])
1771 self.decoder_block1 = DecoderBlock(channels[3], channels[3]//2)
1772
1773 self.lateral_block4 = LateralBlock(channels[1], channels[1])
1774 self.lateral_block3 = LateralBlock(channels[2], channels[2])
1775 self.lateral_block2 = LateralBlock(channels[3], channels[3])
1776
1777 if self.config.ms_supervision:
1778 self.conv_ms_spvn_4 = nn.Conv2d(channels[1], 1, 1, 1, 0)
1779 self.conv_ms_spvn_3 = nn.Conv2d(channels[2], 1, 1, 1, 0)
1780 self.conv_ms_spvn_2 = nn.Conv2d(channels[3], 1, 1, 1, 0)
1781 self.conv_out1 = nn.Sequential(nn.Conv2d(channels[3]//2, 1, 1, 1, 0))
1782
1783 def forward(self, features):
1784 x, x1, x2, x3, x4 = features
1785 outs = []
1786 p4 = self.decoder_block4(x4)
1787 _p4 = F.interpolate(p4, size=x3.shape[2:], mode='bilinear', align_corners=True)
1788 _p3 = _p4 + self.lateral_block4(x3)
1789
1790 p3 = self.decoder_block3(_p3)
1791 _p3 = F.interpolate(p3, size=x2.shape[2:], mode='bilinear', align_corners=True)
1792 _p2 = _p3 + self.lateral_block3(x2)
1793
1794 p2 = self.decoder_block2(_p2)
1795 _p2 = F.interpolate(p2, size=x1.shape[2:], mode='bilinear', align_corners=True)
1796 _p1 = _p2 + self.lateral_block2(x1)
1797
1798 _p1 = self.decoder_block1(_p1)
1799 _p1 = F.interpolate(_p1, size=x.shape[2:], mode='bilinear', align_corners=True)
1800 p1_out = self.conv_out1(_p1)
1801
1802 if self.config.ms_supervision:
1803 outs.append(self.conv_ms_spvn_4(p4))
1804 outs.append(self.conv_ms_spvn_3(p3))
1805 outs.append(self.conv_ms_spvn_2(p2))
1806 outs.append(p1_out)
1807 return outs
1808
1809
1810 class RefUNet(nn.Module):
1811 # Refinement
1812 def __init__(self, in_channels=3+1):
1813 super(RefUNet, self).__init__()
1814 self.encoder_1 = nn.Sequential(
1815 nn.Conv2d(in_channels, 64, 3, 1, 1),
1816 nn.Conv2d(64, 64, 3, 1, 1),
1817 nn.BatchNorm2d(64),
1818 nn.ReLU(inplace=True)
1819 )
1820
1821 self.encoder_2 = nn.Sequential(
1822 nn.MaxPool2d(2, 2, ceil_mode=True),
1823 nn.Conv2d(64, 64, 3, 1, 1),
1824 nn.BatchNorm2d(64),
1825 nn.ReLU(inplace=True)
1826 )
1827
1828 self.encoder_3 = nn.Sequential(
1829 nn.MaxPool2d(2, 2, ceil_mode=True),
1830 nn.Conv2d(64, 64, 3, 1, 1),
1831 nn.BatchNorm2d(64),
1832 nn.ReLU(inplace=True)
1833 )
1834
1835 self.encoder_4 = nn.Sequential(
1836 nn.MaxPool2d(2, 2, ceil_mode=True),
1837 nn.Conv2d(64, 64, 3, 1, 1),
1838 nn.BatchNorm2d(64),
1839 nn.ReLU(inplace=True)
1840 )
1841
1842 self.pool4 = nn.MaxPool2d(2, 2, ceil_mode=True)
1843 #####
1844 self.decoder_5 = nn.Sequential(
1845 nn.Conv2d(64, 64, 3, 1, 1),
1846 nn.BatchNorm2d(64),
1847 nn.ReLU(inplace=True)
1848 )
1849 #####
1850 self.decoder_4 = nn.Sequential(
1851 nn.Conv2d(128, 64, 3, 1, 1),
1852 nn.BatchNorm2d(64),
1853 nn.ReLU(inplace=True)
1854 )
1855
1856 self.decoder_3 = nn.Sequential(
1857 nn.Conv2d(128, 64, 3, 1, 1),
1858 nn.BatchNorm2d(64),
1859 nn.ReLU(inplace=True)
1860 )
1861
1862 self.decoder_2 = nn.Sequential(
1863 nn.Conv2d(128, 64, 3, 1, 1),
1864 nn.BatchNorm2d(64),
1865 nn.ReLU(inplace=True)
1866 )
1867
1868 self.decoder_1 = nn.Sequential(
1869 nn.Conv2d(128, 64, 3, 1, 1),
1870 nn.BatchNorm2d(64),
1871 nn.ReLU(inplace=True)
1872 )
1873
1874 self.conv_d0 = nn.Conv2d(64, 1, 3, 1, 1)
1875
1876 self.upscore2 = nn.Upsample(scale_factor=2, mode='bilinear', align_corners=True)
1877
1878 def forward(self, x):
1879 outs = []
1880 if isinstance(x, list):
1881 x = torch.cat(x, dim=1)
1882 hx = x
1883
1884 hx1 = self.encoder_1(hx)
1885 hx2 = self.encoder_2(hx1)
1886 hx3 = self.encoder_3(hx2)
1887 hx4 = self.encoder_4(hx3)
1888
1889 hx = self.decoder_5(self.pool4(hx4))
1890 hx = torch.cat((self.upscore2(hx), hx4), 1)
1891
1892 d4 = self.decoder_4(hx)
1893 hx = torch.cat((self.upscore2(d4), hx3), 1)
1894
1895 d3 = self.decoder_3(hx)
1896 hx = torch.cat((self.upscore2(d3), hx2), 1)
1897
1898 d2 = self.decoder_2(hx)
1899 hx = torch.cat((self.upscore2(d2), hx1), 1)
1900
1901 d1 = self.decoder_1(hx)
1902
1903 x = self.conv_d0(d1)
1904 outs.append(x)
1905 return outs
1906
1907
1908
1909 ### models/stem_layer.py
1910
1911 import torch.nn as nn
1912 # from utils import build_act_layer, build_norm_layer
1913
1914
1915 class StemLayer(nn.Module):
1916 r""" Stem layer of InternImage
1917 Args:
1918 in_channels (int): number of input channels
1919 out_channels (int): number of output channels
1920 act_layer (str): activation layer
1921 norm_layer (str): normalization layer
1922 """
1923
1924 def __init__(self,
1925 in_channels=3+1,
1926 inter_channels=48,
1927 out_channels=96,
1928 act_layer='GELU',
1929 norm_layer='BN'):
1930 super().__init__()
1931 self.conv1 = nn.Conv2d(in_channels,
1932 inter_channels,
1933 kernel_size=3,
1934 stride=1,
1935 padding=1)
1936 self.norm1 = build_norm_layer(
1937 inter_channels, norm_layer, 'channels_first', 'channels_first'
1938 )
1939 self.act = build_act_layer(act_layer)
1940 self.conv2 = nn.Conv2d(inter_channels,
1941 out_channels,
1942 kernel_size=3,
1943 stride=1,
1944 padding=1)
1945 self.norm2 = build_norm_layer(
1946 out_channels, norm_layer, 'channels_first', 'channels_first'
1947 )
1948
1949 def forward(self, x):
1950 x = self.conv1(x)
1951 x = self.norm1(x)
1952 x = self.act(x)
1953 x = self.conv2(x)
1954 x = self.norm2(x)
1955 return x
1956
1957
1958 ### models/birefnet.py
1959
1960 import torch
1961 import torch.nn as nn
1962 import torch.nn.functional as F
1963 from kornia.filters import laplacian
1964 from transformers import PreTrainedModel
1965 from einops import rearrange
1966
1967 # from config import Config
1968 # from dataset import class_labels_TR_sorted
1969 # from models.build_backbone import build_backbone
1970 # from models.decoder_blocks import BasicDecBlk, ResBlk, HierarAttDecBlk
1971 # from models.lateral_blocks import BasicLatBlk
1972 # from models.aspp import ASPP, ASPPDeformable
1973 # from models.ing import *
1974 # from models.refiner import Refiner, RefinerPVTInChannels4, RefUNet
1975 # from models.stem_layer import StemLayer
1976 from .BiRefNet_config import BiRefNetConfig
1977
1978
1979 def image2patches(image, grid_h=2, grid_w=2, patch_ref=None, transformation='b c (hg h) (wg w) -> (b hg wg) c h w'):
1980 if patch_ref is not None:
1981 grid_h, grid_w = image.shape[-2] // patch_ref.shape[-2], image.shape[-1] // patch_ref.shape[-1]
1982 patches = rearrange(image, transformation, hg=grid_h, wg=grid_w)
1983 return patches
1984
1985 def patches2image(patches, grid_h=2, grid_w=2, patch_ref=None, transformation='(b hg wg) c h w -> b c (hg h) (wg w)'):
1986 if patch_ref is not None:
1987 grid_h, grid_w = patch_ref.shape[-2] // patches[0].shape[-2], patch_ref.shape[-1] // patches[0].shape[-1]
1988 image = rearrange(patches, transformation, hg=grid_h, wg=grid_w)
1989 return image
1990
1991 class BiRefNet(
1992 PreTrainedModel
1993 ):
1994 config_class = BiRefNetConfig
1995 def __init__(self, bb_pretrained=True, config=BiRefNetConfig()):
1996 super(BiRefNet, self).__init__(config)
1997 bb_pretrained = config.bb_pretrained
1998 self.config = Config()
1999 self.epoch = 1
2000 self.bb = build_backbone(self.config.bb, pretrained=bb_pretrained)
2001
2002 channels = self.config.lateral_channels_in_collection
2003
2004 if self.config.auxiliary_classification:
2005 self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
2006 self.cls_head = nn.Sequential(
2007 nn.Linear(channels[0], len(class_labels_TR_sorted))
2008 )
2009
2010 if self.config.squeeze_block:
2011 self.squeeze_module = nn.Sequential(*[
2012 eval(self.config.squeeze_block.split('_x')[0])(channels[0]+sum(self.config.cxt), channels[0])
2013 for _ in range(eval(self.config.squeeze_block.split('_x')[1]))
2014 ])
2015
2016 self.decoder = Decoder(channels)
2017
2018 if self.config.ender:
2019 self.dec_end = nn.Sequential(
2020 nn.Conv2d(1, 16, 3, 1, 1),
2021 nn.Conv2d(16, 1, 3, 1, 1),
2022 nn.ReLU(inplace=True),
2023 )
2024
2025 # refine patch-level segmentation
2026 if self.config.refine:
2027 if self.config.refine == 'itself':
2028 self.stem_layer = StemLayer(in_channels=3+1, inter_channels=48, out_channels=3, norm_layer='BN' if self.config.batch_size > 1 else 'LN')
2029 else:
2030 self.refiner = eval('{}({})'.format(self.config.refine, 'in_channels=3+1'))
2031
2032 if self.config.freeze_bb:
2033 # Freeze the backbone...
2034 print(self.named_parameters())
2035 for key, value in self.named_parameters():
2036 if 'bb.' in key and 'refiner.' not in key:
2037 value.requires_grad = False
2038
2039 self.post_init()
2040
2041 def forward_enc(self, x):
2042 if self.config.bb in ['vgg16', 'vgg16bn', 'resnet50']:
2043 x1 = self.bb.conv1(x); x2 = self.bb.conv2(x1); x3 = self.bb.conv3(x2); x4 = self.bb.conv4(x3)
2044 else:
2045 x1, x2, x3, x4 = self.bb(x)
2046 if self.config.mul_scl_ipt == 'cat':
2047 B, C, H, W = x.shape
2048 x1_, x2_, x3_, x4_ = self.bb(F.interpolate(x, size=(H//2, W//2), mode='bilinear', align_corners=True))
2049 x1 = torch.cat([x1, F.interpolate(x1_, size=x1.shape[2:], mode='bilinear', align_corners=True)], dim=1)
2050 x2 = torch.cat([x2, F.interpolate(x2_, size=x2.shape[2:], mode='bilinear', align_corners=True)], dim=1)
2051 x3 = torch.cat([x3, F.interpolate(x3_, size=x3.shape[2:], mode='bilinear', align_corners=True)], dim=1)
2052 x4 = torch.cat([x4, F.interpolate(x4_, size=x4.shape[2:], mode='bilinear', align_corners=True)], dim=1)
2053 elif self.config.mul_scl_ipt == 'add':
2054 B, C, H, W = x.shape
2055 x1_, x2_, x3_, x4_ = self.bb(F.interpolate(x, size=(H//2, W//2), mode='bilinear', align_corners=True))
2056 x1 = x1 + F.interpolate(x1_, size=x1.shape[2:], mode='bilinear', align_corners=True)
2057 x2 = x2 + F.interpolate(x2_, size=x2.shape[2:], mode='bilinear', align_corners=True)
2058 x3 = x3 + F.interpolate(x3_, size=x3.shape[2:], mode='bilinear', align_corners=True)
2059 x4 = x4 + F.interpolate(x4_, size=x4.shape[2:], mode='bilinear', align_corners=True)
2060 class_preds = self.cls_head(self.avgpool(x4).view(x4.shape[0], -1)) if self.training and self.config.auxiliary_classification else None
2061 if self.config.cxt:
2062 x4 = torch.cat(
2063 (
2064 *[
2065 F.interpolate(x1, size=x4.shape[2:], mode='bilinear', align_corners=True),
2066 F.interpolate(x2, size=x4.shape[2:], mode='bilinear', align_corners=True),
2067 F.interpolate(x3, size=x4.shape[2:], mode='bilinear', align_corners=True),
2068 ][-len(self.config.cxt):],
2069 x4
2070 ),
2071 dim=1
2072 )
2073 return (x1, x2, x3, x4), class_preds
2074
2075 def forward_ori(self, x):
2076 ########## Encoder ##########
2077 (x1, x2, x3, x4), class_preds = self.forward_enc(x)
2078 if self.config.squeeze_block:
2079 x4 = self.squeeze_module(x4)
2080 ########## Decoder ##########
2081 features = [x, x1, x2, x3, x4]
2082 if self.training and self.config.out_ref:
2083 features.append(laplacian(torch.mean(x, dim=1).unsqueeze(1), kernel_size=5))
2084 scaled_preds = self.decoder(features)
2085 return scaled_preds, class_preds
2086
2087 def forward(self, x):
2088 scaled_preds, class_preds = self.forward_ori(x)
2089 class_preds_lst = [class_preds]
2090 return [scaled_preds, class_preds_lst] if self.training else scaled_preds
2091
2092
2093 class Decoder(nn.Module):
2094 def __init__(self, channels):
2095 super(Decoder, self).__init__()
2096 self.config = Config()
2097 DecoderBlock = eval(self.config.dec_blk)
2098 LateralBlock = eval(self.config.lat_blk)
2099
2100 if self.config.dec_ipt:
2101 self.split = self.config.dec_ipt_split
2102 N_dec_ipt = 64
2103 DBlock = SimpleConvs
2104 ic = 64
2105 ipt_cha_opt = 1
2106 self.ipt_blk5 = DBlock(2**10*3 if self.split else 3, [N_dec_ipt, channels[0]//8][ipt_cha_opt], inter_channels=ic)
2107 self.ipt_blk4 = DBlock(2**8*3 if self.split else 3, [N_dec_ipt, channels[0]//8][ipt_cha_opt], inter_channels=ic)
2108 self.ipt_blk3 = DBlock(2**6*3 if self.split else 3, [N_dec_ipt, channels[1]//8][ipt_cha_opt], inter_channels=ic)
2109 self.ipt_blk2 = DBlock(2**4*3 if self.split else 3, [N_dec_ipt, channels[2]//8][ipt_cha_opt], inter_channels=ic)
2110 self.ipt_blk1 = DBlock(2**0*3 if self.split else 3, [N_dec_ipt, channels[3]//8][ipt_cha_opt], inter_channels=ic)
2111 else:
2112 self.split = None
2113
2114 self.decoder_block4 = DecoderBlock(channels[0]+([N_dec_ipt, channels[0]//8][ipt_cha_opt] if self.config.dec_ipt else 0), channels[1])
2115 self.decoder_block3 = DecoderBlock(channels[1]+([N_dec_ipt, channels[0]//8][ipt_cha_opt] if self.config.dec_ipt else 0), channels[2])
2116 self.decoder_block2 = DecoderBlock(channels[2]+([N_dec_ipt, channels[1]//8][ipt_cha_opt] if self.config.dec_ipt else 0), channels[3])
2117 self.decoder_block1 = DecoderBlock(channels[3]+([N_dec_ipt, channels[2]//8][ipt_cha_opt] if self.config.dec_ipt else 0), channels[3]//2)
2118 self.conv_out1 = nn.Sequential(nn.Conv2d(channels[3]//2+([N_dec_ipt, channels[3]//8][ipt_cha_opt] if self.config.dec_ipt else 0), 1, 1, 1, 0))
2119
2120 self.lateral_block4 = LateralBlock(channels[1], channels[1])
2121 self.lateral_block3 = LateralBlock(channels[2], channels[2])
2122 self.lateral_block2 = LateralBlock(channels[3], channels[3])
2123
2124 if self.config.ms_supervision:
2125 self.conv_ms_spvn_4 = nn.Conv2d(channels[1], 1, 1, 1, 0)
2126 self.conv_ms_spvn_3 = nn.Conv2d(channels[2], 1, 1, 1, 0)
2127 self.conv_ms_spvn_2 = nn.Conv2d(channels[3], 1, 1, 1, 0)
2128
2129 if self.config.out_ref:
2130 _N = 16
2131 self.gdt_convs_4 = nn.Sequential(nn.Conv2d(channels[1], _N, 3, 1, 1), nn.BatchNorm2d(_N) if self.config.batch_size > 1 else nn.Identity(), nn.ReLU(inplace=True))
2132 self.gdt_convs_3 = nn.Sequential(nn.Conv2d(channels[2], _N, 3, 1, 1), nn.BatchNorm2d(_N) if self.config.batch_size > 1 else nn.Identity(), nn.ReLU(inplace=True))
2133 self.gdt_convs_2 = nn.Sequential(nn.Conv2d(channels[3], _N, 3, 1, 1), nn.BatchNorm2d(_N) if self.config.batch_size > 1 else nn.Identity(), nn.ReLU(inplace=True))
2134
2135 self.gdt_convs_pred_4 = nn.Sequential(nn.Conv2d(_N, 1, 1, 1, 0))
2136 self.gdt_convs_pred_3 = nn.Sequential(nn.Conv2d(_N, 1, 1, 1, 0))
2137 self.gdt_convs_pred_2 = nn.Sequential(nn.Conv2d(_N, 1, 1, 1, 0))
2138
2139 self.gdt_convs_attn_4 = nn.Sequential(nn.Conv2d(_N, 1, 1, 1, 0))
2140 self.gdt_convs_attn_3 = nn.Sequential(nn.Conv2d(_N, 1, 1, 1, 0))
2141 self.gdt_convs_attn_2 = nn.Sequential(nn.Conv2d(_N, 1, 1, 1, 0))
2142
2143 def forward(self, features):
2144 if self.training and self.config.out_ref:
2145 outs_gdt_pred = []
2146 outs_gdt_label = []
2147 x, x1, x2, x3, x4, gdt_gt = features
2148 else:
2149 x, x1, x2, x3, x4 = features
2150 outs = []
2151
2152 if self.config.dec_ipt:
2153 patches_batch = image2patches(x, patch_ref=x4, transformation='b c (hg h) (wg w) -> b (c hg wg) h w') if self.split else x
2154 x4 = torch.cat((x4, self.ipt_blk5(F.interpolate(patches_batch, size=x4.shape[2:], mode='bilinear', align_corners=True))), 1)
2155 p4 = self.decoder_block4(x4)
2156 m4 = self.conv_ms_spvn_4(p4) if self.config.ms_supervision and self.training else None
2157 if self.config.out_ref:
2158 p4_gdt = self.gdt_convs_4(p4)
2159 if self.training:
2160 # >> GT:
2161 m4_dia = m4
2162 gdt_label_main_4 = gdt_gt * F.interpolate(m4_dia, size=gdt_gt.shape[2:], mode='bilinear', align_corners=True)
2163 outs_gdt_label.append(gdt_label_main_4)
2164 # >> Pred:
2165 gdt_pred_4 = self.gdt_convs_pred_4(p4_gdt)
2166 outs_gdt_pred.append(gdt_pred_4)
2167 gdt_attn_4 = self.gdt_convs_attn_4(p4_gdt).sigmoid()
2168 # >> Finally:
2169 p4 = p4 * gdt_attn_4
2170 _p4 = F.interpolate(p4, size=x3.shape[2:], mode='bilinear', align_corners=True)
2171 _p3 = _p4 + self.lateral_block4(x3)
2172
2173 if self.config.dec_ipt:
2174 patches_batch = image2patches(x, patch_ref=_p3, transformation='b c (hg h) (wg w) -> b (c hg wg) h w') if self.split else x
2175 _p3 = torch.cat((_p3, self.ipt_blk4(F.interpolate(patches_batch, size=x3.shape[2:], mode='bilinear', align_corners=True))), 1)
2176 p3 = self.decoder_block3(_p3)
2177 m3 = self.conv_ms_spvn_3(p3) if self.config.ms_supervision and self.training else None
2178 if self.config.out_ref:
2179 p3_gdt = self.gdt_convs_3(p3)
2180 if self.training:
2181 # >> GT:
2182 # m3 --dilation--> m3_dia
2183 # G_3^gt * m3_dia --> G_3^m, which is the label of gradient
2184 m3_dia = m3
2185 gdt_label_main_3 = gdt_gt * F.interpolate(m3_dia, size=gdt_gt.shape[2:], mode='bilinear', align_corners=True)
2186 outs_gdt_label.append(gdt_label_main_3)
2187 # >> Pred:
2188 # p3 --conv--BN--> F_3^G, where F_3^G predicts the \hat{G_3} with xx
2189 # F_3^G --sigmoid--> A_3^G
2190 gdt_pred_3 = self.gdt_convs_pred_3(p3_gdt)
2191 outs_gdt_pred.append(gdt_pred_3)
2192 gdt_attn_3 = self.gdt_convs_attn_3(p3_gdt).sigmoid()
2193 # >> Finally:
2194 # p3 = p3 * A_3^G
2195 p3 = p3 * gdt_attn_3
2196 _p3 = F.interpolate(p3, size=x2.shape[2:], mode='bilinear', align_corners=True)
2197 _p2 = _p3 + self.lateral_block3(x2)
2198
2199 if self.config.dec_ipt:
2200 patches_batch = image2patches(x, patch_ref=_p2, transformation='b c (hg h) (wg w) -> b (c hg wg) h w') if self.split else x
2201 _p2 = torch.cat((_p2, self.ipt_blk3(F.interpolate(patches_batch, size=x2.shape[2:], mode='bilinear', align_corners=True))), 1)
2202 p2 = self.decoder_block2(_p2)
2203 m2 = self.conv_ms_spvn_2(p2) if self.config.ms_supervision and self.training else None
2204 if self.config.out_ref:
2205 p2_gdt = self.gdt_convs_2(p2)
2206 if self.training:
2207 # >> GT:
2208 m2_dia = m2
2209 gdt_label_main_2 = gdt_gt * F.interpolate(m2_dia, size=gdt_gt.shape[2:], mode='bilinear', align_corners=True)
2210 outs_gdt_label.append(gdt_label_main_2)
2211 # >> Pred:
2212 gdt_pred_2 = self.gdt_convs_pred_2(p2_gdt)
2213 outs_gdt_pred.append(gdt_pred_2)
2214 gdt_attn_2 = self.gdt_convs_attn_2(p2_gdt).sigmoid()
2215 # >> Finally:
2216 p2 = p2 * gdt_attn_2
2217 _p2 = F.interpolate(p2, size=x1.shape[2:], mode='bilinear', align_corners=True)
2218 _p1 = _p2 + self.lateral_block2(x1)
2219
2220 if self.config.dec_ipt:
2221 patches_batch = image2patches(x, patch_ref=_p1, transformation='b c (hg h) (wg w) -> b (c hg wg) h w') if self.split else x
2222 _p1 = torch.cat((_p1, self.ipt_blk2(F.interpolate(patches_batch, size=x1.shape[2:], mode='bilinear', align_corners=True))), 1)
2223 _p1 = self.decoder_block1(_p1)
2224 _p1 = F.interpolate(_p1, size=x.shape[2:], mode='bilinear', align_corners=True)
2225
2226 if self.config.dec_ipt:
2227 patches_batch = image2patches(x, patch_ref=_p1, transformation='b c (hg h) (wg w) -> b (c hg wg) h w') if self.split else x
2228 _p1 = torch.cat((_p1, self.ipt_blk1(F.interpolate(patches_batch, size=x.shape[2:], mode='bilinear', align_corners=True))), 1)
2229 p1_out = self.conv_out1(_p1)
2230
2231 if self.config.ms_supervision and self.training:
2232 outs.append(m4)
2233 outs.append(m3)
2234 outs.append(m2)
2235 outs.append(p1_out)
2236 return outs if not (self.config.out_ref and self.training) else ([outs_gdt_pred, outs_gdt_label], outs)
2237
2238
2239 class SimpleConvs(nn.Module):
2240 def __init__(
2241 self, in_channels: int, out_channels: int, inter_channels=64
2242 ) -> None:
2243 super().__init__()
2244 self.conv1 = nn.Conv2d(in_channels, inter_channels, 3, 1, 1)
2245 self.conv_out = nn.Conv2d(inter_channels, out_channels, 3, 1, 1)
2246
2247 def forward(self, x):
2248 return self.conv_out(self.conv1(x))
2249