File size: 13,202 Bytes
2d07fab |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 |
import torch
import torch.nn.functional as F
from torch import nn
from torchvision.ops import box_convert
import embeddings as emb
import encoders as enc
from encoders import weight_init
def conv3x3(in_channels, out_channels, num_groups=0):
return nn.Sequential(
# Conv2d w/o bias since BatchNorm2d/GroupNorm already accounts for it (affine=True)
nn.Conv2d(in_channels, out_channels, (3, 3), 1, 1, bias=False),
nn.BatchNorm2d(out_channels) if num_groups < 1 else nn.GroupNorm(num_groups, out_channels),
nn.ReLU(inplace=True),
)
class IntuitionKillingMachine(nn.Module):
def __init__(self,
backbone='resnet50', pretrained=True, embedding_size=256,
num_heads=8, num_layers=6, num_conv=4, dropout_p=0.1,
segmentation_head=True, mask_pooling=True):
super().__init__()
if backbone.endswith('+tr'):
self.vis_enc = enc.TransformerImageEncoder(
backbone=backbone.rstrip('+tr'),
out_channels=embedding_size,
pretrained=pretrained,
)
elif backbone.endswith('+fpn'):
self.vis_enc = enc.FPNImageEncoder(
backbone=backbone.rstrip('+fpn'),
out_channels=embedding_size,
pretrained=pretrained,
with_pos=False
)
else:
self.vis_enc = enc.ImageEncoder(
backbone=backbone,
out_channels=embedding_size,
pretrained=pretrained,
with_pos=False
)
# freeze ResNet stem
if 'resnet' in backbone:
self.vis_enc.backbone.conv1.requires_grad = False
self.vis_enc.backbone.conv1.eval()
self.vis_pos_emb = emb.LearnedPositionEmbedding2D(
embedding_dim=embedding_size
)
self.lan_enc = enc.LanguageEncoder(
out_features=embedding_size,
global_pooling=False,
dropout_p=dropout_p
)
self.lan_pos_emb = emb.LearnedPositionEmbedding1D(
embedding_dim=embedding_size
)
from transformers_pos import (
XTransformerEncoder,
TransformerEncoder,
TransformerEncoderLayer,
)
self.encoder = TransformerEncoder(
TransformerEncoderLayer(
d_model=embedding_size,
nhead=num_heads,
dropout=dropout_p,
batch_first=True
),
num_layers=num_layers
)
# ---
# CONV PRE-HEAD (NECK?)
if num_conv > 0:
self.pre_head = nn.Sequential(*[
conv3x3(embedding_size, embedding_size) for _ in range(num_conv)
])
self.pre_head.apply(weight_init)
else:
self.pre_head = nn.Identity()
# ---
# OUTPUT HEADS
# box prediction
self.head = nn.Sequential(
nn.Linear(embedding_size, 4, bias=True),
nn.Sigmoid()
)
self.head.apply(weight_init)
# box segmentation mask
self.segm_head = None
if segmentation_head:
self.segm_head = nn.Sequential(
nn.Conv2d(embedding_size, 1, (3, 3), 1, 1, bias=True),
#nn.Sigmoid()
)
self.segm_head.apply(weight_init)
# ---
self.mask_pooling = bool(mask_pooling)
if self.mask_pooling and self.segm_head is None:
raise RuntimeError('mask pooling w/o a segmentation head does not makes sense')
self.embedding_size = embedding_size
# def slow_param_ids(self, **kwargs):
# return []
def slow_param_ids(self, slow_visual_backbone=True, slow_language_backbone=True):
ids = []
if slow_visual_backbone:
ids += [id(p) for p in self.vis_enc.backbone.parameters()]
if hasattr(self.vis_enc, 'encoder'): # +tr
ids += [id(p) for p in self.vis_enc.encoder.parameters()]
if slow_language_backbone:
if isinstance(self.lan_enc, enc.LanguageEncoder):
ids += [id(p) for p in self.lan_enc.language_model.parameters()]
else:
ids += [id(p) for p in self.lan_enc.embeddings.parameters()]
return ids
def flatten(self, x):
N, D, H, W = x.size()
x = x.to(memory_format=torch.channels_last)
x = x.permute(0, 2, 3, 1).view(N, H*W, D)
return x # NxHWxD
def unflatten(self, x, size):
N, R, D = x.size()
H, W = size
assert R == H*W, 'wrong tensor size'
x = x.permute(0, 2, 1).to(memory_format=torch.contiguous_format)
x = x.view(N, D, H, W)
return x # NxDxHxW
def forward(self, input):
img, mask, tok = input['image'], input['mask'], input['tok']
# ---
# VISUAL EMBEDDINGS
x, x_mask = self.vis_enc(img, mask) # NxDxHxW, NxHxW
x_pos = self.vis_pos_emb(x, x_mask)
N, D, H, W = x.size() # save dims before flatten
x = self.flatten(x) # NxRxD
x_mask = self.flatten(x_mask).squeeze(-1) # NxR
x_pos = self.flatten(x_pos) # NxRxD
# ---
# LANGUAGE EMBEDDINGS
z, z_mask = self.lan_enc(tok) # NxTxD, NxT
z_pos = self.lan_pos_emb(z) # NxTxD
# ---
# V+L TRANSFORMER
# [...visual...]+[[CLS]...language tokens...[SEP]]
xz = torch.cat([x, z], dim=1)
xz_mask = torch.cat([x_mask, z_mask], dim=1)
xz_pos = torch.cat([x_pos, z_pos], dim=1)
xz = self.encoder(xz, src_key_padding_mask=(xz_mask==0), pos=xz_pos) #, size=(H,W))
# restore spatiality of visual embeddings after cross-modal encoding
xz_vis = xz[:, :H*W, ...]
xz_vis = self.unflatten(xz_vis, (H, W))
x_mask = self.unflatten(x_mask.unsqueeze(-1), (H, W))
# ---
# convolutional pre-head
xz_vis = self.pre_head(xz_vis)
# ---
# segmentation head w/ (opt.) pooling
segm_mask, pooled_feat = None, None
if self.segm_head is not None:
segm_mask = torch.sigmoid(self.segm_head(xz_vis)) * x_mask
if self.mask_pooling: # box mask guided pooling
pooled_feat = (segm_mask * xz_vis).sum((2, 3)) / segm_mask.sum((2, 3))
segm_mask = F.interpolate(segm_mask, img.size()[2:], mode='bilinear', align_corners=True)
# if not mask_pooling, do the pooling using all visual feats (equiv. to a uniform mask)
if pooled_feat is None:
pooled_feat = (x_mask * xz_vis).sum((2, 3)) / x_mask.sum((2, 3))
# bbox prediction
pred = self.head(pooled_feat)
pred = box_convert(pred, 'cxcywh', 'xyxy')
return pred, segm_mask
class HeadlessMachine(nn.Module):
def __init__(self,
backbone='resnet50', pretrained=True, embedding_size=256,
num_heads=8, num_layers=6, num_conv=4, dropout_p=0.1,
segmentation_head=True, mask_pooling=True):
super().__init__()
if backbone.endswith('+tr'):
self.vis_enc = enc.TransformerImageEncoder(
backbone=backbone.rstrip('+tr'),
out_channels=embedding_size,
pretrained=pretrained,
)
elif backbone.endswith('+fpn'):
self.vis_enc = enc.FPNImageEncoder(
backbone=backbone.rstrip('+fpn'),
out_channels=embedding_size,
pretrained=pretrained,
with_pos=False
)
else:
self.vis_enc = enc.ImageEncoder(
backbone=backbone,
out_channels=embedding_size,
pretrained=pretrained,
with_pos=False
)
# freeze ResNet stem
if 'resnet' in backbone:
self.vis_enc.backbone.conv1.requires_grad = False
self.vis_enc.backbone.conv1.eval()
self.vis_pos_emb = emb.LearnedPositionEmbedding2D(
embedding_dim=embedding_size
)
self.lan_enc = enc.LanguageEncoder(
out_features=embedding_size,
global_pooling=False,
dropout_p=dropout_p
)
self.lan_pos_emb = emb.LearnedPositionEmbedding1D(
embedding_dim=embedding_size
)
from transformers_pos import (
XTransformerEncoder,
TransformerEncoder,
TransformerEncoderLayer,
)
self.encoder = TransformerEncoder(
TransformerEncoderLayer(
d_model=embedding_size,
nhead=num_heads,
dropout=dropout_p,
batch_first=True
),
num_layers=num_layers
)
# ---
# CONV PRE-HEAD (NECK?)
if num_conv > 0:
self.pre_head = nn.Sequential(*[
conv3x3(embedding_size, embedding_size) for _ in range(num_conv)
])
self.pre_head.apply(weight_init)
else:
self.pre_head = nn.Identity()
# ---
# OUTPUT HEADS
# box prediction
self.head = nn.Sequential(
nn.Linear(embedding_size, 4, bias=True),
nn.Sigmoid()
)
self.head.apply(weight_init)
# box segmentation mask
self.segm_head = None
if segmentation_head:
self.segm_head = nn.Sequential(
nn.Conv2d(embedding_size, 1, (3, 3), 1, 1, bias=True),
#nn.Sigmoid()
)
self.segm_head.apply(weight_init)
# ---
self.mask_pooling = bool(mask_pooling)
if self.mask_pooling and self.segm_head is None:
raise RuntimeError('mask pooling w/o a segmentation head does not makes sense')
self.embedding_size = embedding_size
# def slow_param_ids(self, **kwargs):
# return []
def slow_param_ids(self, slow_visual_backbone=True, slow_language_backbone=True):
ids = []
if slow_visual_backbone:
ids += [id(p) for p in self.vis_enc.backbone.parameters()]
if hasattr(self.vis_enc, 'encoder'): # +tr
ids += [id(p) for p in self.vis_enc.encoder.parameters()]
if slow_language_backbone:
if isinstance(self.lan_enc, enc.LanguageEncoder):
ids += [id(p) for p in self.lan_enc.language_model.parameters()]
else:
ids += [id(p) for p in self.lan_enc.embeddings.parameters()]
return ids
def flatten(self, x):
N, D, H, W = x.size()
x = x.to(memory_format=torch.channels_last)
x = x.permute(0, 2, 3, 1).view(N, H*W, D)
return x # NxHWxD
def unflatten(self, x, size):
N, R, D = x.size()
H, W = size
assert R == H*W, 'wrong tensor size'
x = x.permute(0, 2, 1).to(memory_format=torch.contiguous_format)
x = x.view(N, D, H, W)
return x # NxDxHxW
def forward(self, input):
img, mask, tok = input['image'], input['mask'], input['tok']
# ---
# VISUAL EMBEDDINGS
x, x_mask = self.vis_enc(img, mask) # NxDxHxW, NxHxW
x_pos = self.vis_pos_emb(x, x_mask)
N, D, H, W = x.size() # save dims before flatten
x = self.flatten(x) # NxRxD
x_mask = self.flatten(x_mask).squeeze(-1) # NxR
x_pos = self.flatten(x_pos) # NxRxD
# ---
# LANGUAGE EMBEDDINGS
z, z_mask = self.lan_enc(tok) # NxTxD, NxT
z_pos = self.lan_pos_emb(z) # NxTxD
# ---
# V+L TRANSFORMER
# [...visual...]+[[CLS]...language tokens...[SEP]]
xz = torch.cat([x, z], dim=1)
xz_mask = torch.cat([x_mask, z_mask], dim=1)
xz_pos = torch.cat([x_pos, z_pos], dim=1)
xz = self.encoder(xz, src_key_padding_mask=(xz_mask==0), pos=xz_pos) #, size=(H,W))
# restore spatiality of visual embeddings after cross-modal encoding
xz_vis = xz[:, :H*W, ...]
xz_vis = self.unflatten(xz_vis, (H, W))
x_mask = self.unflatten(x_mask.unsqueeze(-1), (H, W))
# ---
# convolutional pre-head
xz_vis = self.pre_head(xz_vis)
# ---
# segmentation head w/ (opt.) pooling
segm_mask, pooled_feat = None, None
if self.segm_head is not None:
segm_mask = torch.sigmoid(self.segm_head(xz_vis)) * x_mask
if self.mask_pooling: # box mask guided pooling
pooled_feat = (segm_mask * xz_vis).sum((2, 3)) / segm_mask.sum((2, 3))
segm_mask = F.interpolate(segm_mask, img.size()[2:], mode='bilinear', align_corners=True)
# if not mask_pooling, do the pooling using all visual feats (equiv. to a uniform mask)
if pooled_feat is None:
pooled_feat = (x_mask * xz_vis).sum((2, 3)) / x_mask.sum((2, 3))
# bbox prediction
pred = self.head(pooled_feat)
pred = box_convert(pred, 'cxcywh', 'xyxy')
return pred, segm_mask
|