forked from huggingface/diffusers
-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathembeddings.py
More file actions
655 lines (509 loc) 路 23 KB
/
Copy pathembeddings.py
File metadata and controls
655 lines (509 loc) 路 23 KB
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
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
# Copyright 2023 The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import math
from typing import Optional
import numpy as np
import torch
from torch import nn
from .activations import get_activation
def get_timestep_embedding(
timesteps: torch.Tensor,
embedding_dim: int,
flip_sin_to_cos: bool = False,
downscale_freq_shift: float = 1,
scale: float = 1,
max_period: int = 10000,
):
"""
This matches the implementation in Denoising Diffusion Probabilistic Models: Create sinusoidal timestep embeddings.
:param timesteps: a 1-D Tensor of N indices, one per batch element.
These may be fractional.
:param embedding_dim: the dimension of the output. :param max_period: controls the minimum frequency of the
embeddings. :return: an [N x dim] Tensor of positional embeddings.
"""
assert len(timesteps.shape) == 1, "Timesteps should be a 1d-array"
half_dim = embedding_dim // 2
exponent = -math.log(max_period) * torch.arange(
start=0, end=half_dim, dtype=torch.float32, device=timesteps.device
)
exponent = exponent / (half_dim - downscale_freq_shift)
emb = torch.exp(exponent)
emb = timesteps[:, None].float() * emb[None, :]
# scale embeddings
emb = scale * emb
# concat sine and cosine embeddings
emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=-1)
# flip sine and cosine embeddings
if flip_sin_to_cos:
emb = torch.cat([emb[:, half_dim:], emb[:, :half_dim]], dim=-1)
# zero pad
if embedding_dim % 2 == 1:
emb = torch.nn.functional.pad(emb, (0, 1, 0, 0))
return emb
def get_2d_sincos_pos_embed(embed_dim, grid_size, cls_token=False, extra_tokens=0):
"""
grid_size: int of the grid height and width return: pos_embed: [grid_size*grid_size, embed_dim] or
[1+grid_size*grid_size, embed_dim] (w/ or w/o cls_token)
"""
grid_h = np.arange(grid_size, dtype=np.float32)
grid_w = np.arange(grid_size, dtype=np.float32)
grid = np.meshgrid(grid_w, grid_h) # here w goes first
grid = np.stack(grid, axis=0)
grid = grid.reshape([2, 1, grid_size, grid_size])
pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid)
if cls_token and extra_tokens > 0:
pos_embed = np.concatenate([np.zeros([extra_tokens, embed_dim]), pos_embed], axis=0)
return pos_embed
def get_2d_sincos_pos_embed_from_grid(embed_dim, grid):
if embed_dim % 2 != 0:
raise ValueError("embed_dim must be divisible by 2")
# use half of dimensions to encode grid_h
emb_h = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[0]) # (H*W, D/2)
emb_w = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[1]) # (H*W, D/2)
emb = np.concatenate([emb_h, emb_w], axis=1) # (H*W, D)
return emb
def get_1d_sincos_pos_embed_from_grid(embed_dim, pos):
"""
embed_dim: output dimension for each position pos: a list of positions to be encoded: size (M,) out: (M, D)
"""
if embed_dim % 2 != 0:
raise ValueError("embed_dim must be divisible by 2")
omega = np.arange(embed_dim // 2, dtype=np.float64)
omega /= embed_dim / 2.0
omega = 1.0 / 10000**omega # (D/2,)
pos = pos.reshape(-1) # (M,)
out = np.einsum("m,d->md", pos, omega) # (M, D/2), outer product
emb_sin = np.sin(out) # (M, D/2)
emb_cos = np.cos(out) # (M, D/2)
emb = np.concatenate([emb_sin, emb_cos], axis=1) # (M, D)
return emb
class PatchEmbed(nn.Module):
"""2D Image to Patch Embedding"""
def __init__(
self,
height=224,
width=224,
patch_size=16,
in_channels=3,
embed_dim=768,
layer_norm=False,
flatten=True,
bias=True,
):
super().__init__()
num_patches = (height // patch_size) * (width // patch_size)
self.flatten = flatten
self.layer_norm = layer_norm
self.proj = nn.Conv2d(
in_channels, embed_dim, kernel_size=(patch_size, patch_size), stride=patch_size, bias=bias
)
if layer_norm:
self.norm = nn.LayerNorm(embed_dim, elementwise_affine=False, eps=1e-6)
else:
self.norm = None
pos_embed = get_2d_sincos_pos_embed(embed_dim, int(num_patches**0.5))
self.register_buffer("pos_embed", torch.from_numpy(pos_embed).float().unsqueeze(0), persistent=False)
def forward(self, latent):
latent = self.proj(latent)
if self.flatten:
latent = latent.flatten(2).transpose(1, 2) # BCHW -> BNC
if self.layer_norm:
latent = self.norm(latent)
return latent + self.pos_embed
class TimestepEmbedding(nn.Module):
def __init__(
self,
in_channels: int,
time_embed_dim: int,
act_fn: str = "silu",
out_dim: int = None,
post_act_fn: Optional[str] = None,
cond_proj_dim=None,
):
super().__init__()
self.linear_1 = nn.Linear(in_channels, time_embed_dim)
if cond_proj_dim is not None:
self.cond_proj = nn.Linear(cond_proj_dim, in_channels, bias=False)
else:
self.cond_proj = None
self.act = get_activation(act_fn)
if out_dim is not None:
time_embed_dim_out = out_dim
else:
time_embed_dim_out = time_embed_dim
self.linear_2 = nn.Linear(time_embed_dim, time_embed_dim_out)
if post_act_fn is None:
self.post_act = None
else:
self.post_act = get_activation(post_act_fn)
def forward(self, sample, condition=None):
if condition is not None:
sample = sample + self.cond_proj(condition)
sample = self.linear_1(sample)
if self.act is not None:
sample = self.act(sample)
sample = self.linear_2(sample)
if self.post_act is not None:
sample = self.post_act(sample)
return sample
class Timesteps(nn.Module):
def __init__(self, num_channels: int, flip_sin_to_cos: bool, downscale_freq_shift: float):
super().__init__()
self.num_channels = num_channels
self.flip_sin_to_cos = flip_sin_to_cos
self.downscale_freq_shift = downscale_freq_shift
def forward(self, timesteps):
t_emb = get_timestep_embedding(
timesteps,
self.num_channels,
flip_sin_to_cos=self.flip_sin_to_cos,
downscale_freq_shift=self.downscale_freq_shift,
)
return t_emb
class GaussianFourierProjection(nn.Module):
"""Gaussian Fourier embeddings for noise levels."""
def __init__(
self, embedding_size: int = 256, scale: float = 1.0, set_W_to_weight=True, log=True, flip_sin_to_cos=False
):
super().__init__()
self.weight = nn.Parameter(torch.randn(embedding_size) * scale, requires_grad=False)
self.log = log
self.flip_sin_to_cos = flip_sin_to_cos
if set_W_to_weight:
# to delete later
self.W = nn.Parameter(torch.randn(embedding_size) * scale, requires_grad=False)
self.weight = self.W
def forward(self, x):
if self.log:
x = torch.log(x)
x_proj = x[:, None] * self.weight[None, :] * 2 * np.pi
if self.flip_sin_to_cos:
out = torch.cat([torch.cos(x_proj), torch.sin(x_proj)], dim=-1)
else:
out = torch.cat([torch.sin(x_proj), torch.cos(x_proj)], dim=-1)
return out
class ImagePositionalEmbeddings(nn.Module):
"""
Converts latent image classes into vector embeddings. Sums the vector embeddings with positional embeddings for the
height and width of the latent space.
For more details, see figure 10 of the dall-e paper: https://arxiv.org/abs/2102.12092
For VQ-diffusion:
Output vector embeddings are used as input for the transformer.
Note that the vector embeddings for the transformer are different than the vector embeddings from the VQVAE.
Args:
num_embed (`int`):
Number of embeddings for the latent pixels embeddings.
height (`int`):
Height of the latent image i.e. the number of height embeddings.
width (`int`):
Width of the latent image i.e. the number of width embeddings.
embed_dim (`int`):
Dimension of the produced vector embeddings. Used for the latent pixel, height, and width embeddings.
"""
def __init__(
self,
num_embed: int,
height: int,
width: int,
embed_dim: int,
):
super().__init__()
self.height = height
self.width = width
self.num_embed = num_embed
self.embed_dim = embed_dim
self.emb = nn.Embedding(self.num_embed, embed_dim)
self.height_emb = nn.Embedding(self.height, embed_dim)
self.width_emb = nn.Embedding(self.width, embed_dim)
def forward(self, index):
emb = self.emb(index)
height_emb = self.height_emb(torch.arange(self.height, device=index.device).view(1, self.height))
# 1 x H x D -> 1 x H x 1 x D
height_emb = height_emb.unsqueeze(2)
width_emb = self.width_emb(torch.arange(self.width, device=index.device).view(1, self.width))
# 1 x W x D -> 1 x 1 x W x D
width_emb = width_emb.unsqueeze(1)
pos_emb = height_emb + width_emb
# 1 x H x W x D -> 1 x L xD
pos_emb = pos_emb.view(1, self.height * self.width, -1)
emb = emb + pos_emb[:, : emb.shape[1], :]
return emb
class LabelEmbedding(nn.Module):
"""
Embeds class labels into vector representations. Also handles label dropout for classifier-free guidance.
Args:
num_classes (`int`): The number of classes.
hidden_size (`int`): The size of the vector embeddings.
dropout_prob (`float`): The probability of dropping a label.
"""
def __init__(self, num_classes, hidden_size, dropout_prob):
super().__init__()
use_cfg_embedding = dropout_prob > 0
self.embedding_table = nn.Embedding(num_classes + use_cfg_embedding, hidden_size)
self.num_classes = num_classes
self.dropout_prob = dropout_prob
def token_drop(self, labels, force_drop_ids=None):
"""
Drops labels to enable classifier-free guidance.
"""
if force_drop_ids is None:
drop_ids = torch.rand(labels.shape[0], device=labels.device) < self.dropout_prob
else:
drop_ids = torch.tensor(force_drop_ids == 1)
labels = torch.where(drop_ids, self.num_classes, labels)
return labels
def forward(self, labels: torch.LongTensor, force_drop_ids=None):
use_dropout = self.dropout_prob > 0
if (self.training and use_dropout) or (force_drop_ids is not None):
labels = self.token_drop(labels, force_drop_ids)
embeddings = self.embedding_table(labels)
return embeddings
class TextImageProjection(nn.Module):
def __init__(
self,
text_embed_dim: int = 1024,
image_embed_dim: int = 768,
cross_attention_dim: int = 768,
num_image_text_embeds: int = 10,
):
super().__init__()
self.num_image_text_embeds = num_image_text_embeds
self.image_embeds = nn.Linear(image_embed_dim, self.num_image_text_embeds * cross_attention_dim)
self.text_proj = nn.Linear(text_embed_dim, cross_attention_dim)
def forward(self, text_embeds: torch.FloatTensor, image_embeds: torch.FloatTensor):
batch_size = text_embeds.shape[0]
# image
image_text_embeds = self.image_embeds(image_embeds)
image_text_embeds = image_text_embeds.reshape(batch_size, self.num_image_text_embeds, -1)
# text
text_embeds = self.text_proj(text_embeds)
return torch.cat([image_text_embeds, text_embeds], dim=1)
class ImageProjection(nn.Module):
def __init__(
self,
image_embed_dim: int = 768,
cross_attention_dim: int = 768,
num_image_text_embeds: int = 32,
):
super().__init__()
self.num_image_text_embeds = num_image_text_embeds
self.image_embeds = nn.Linear(image_embed_dim, self.num_image_text_embeds * cross_attention_dim)
self.norm = nn.LayerNorm(cross_attention_dim)
def forward(self, image_embeds: torch.FloatTensor):
batch_size = image_embeds.shape[0]
# image
image_embeds = self.image_embeds(image_embeds)
image_embeds = image_embeds.reshape(batch_size, self.num_image_text_embeds, -1)
image_embeds = self.norm(image_embeds)
return image_embeds
class CombinedTimestepLabelEmbeddings(nn.Module):
def __init__(self, num_classes, embedding_dim, class_dropout_prob=0.1):
super().__init__()
self.time_proj = Timesteps(num_channels=256, flip_sin_to_cos=True, downscale_freq_shift=1)
self.timestep_embedder = TimestepEmbedding(in_channels=256, time_embed_dim=embedding_dim)
self.class_embedder = LabelEmbedding(num_classes, embedding_dim, class_dropout_prob)
def forward(self, timestep, class_labels, hidden_dtype=None):
timesteps_proj = self.time_proj(timestep)
timesteps_emb = self.timestep_embedder(timesteps_proj.to(dtype=hidden_dtype)) # (N, D)
class_labels = self.class_embedder(class_labels) # (N, D)
conditioning = timesteps_emb + class_labels # (N, D)
return conditioning
class TextTimeEmbedding(nn.Module):
def __init__(self, encoder_dim: int, time_embed_dim: int, num_heads: int = 64):
super().__init__()
self.norm1 = nn.LayerNorm(encoder_dim)
self.pool = AttentionPooling(num_heads, encoder_dim)
self.proj = nn.Linear(encoder_dim, time_embed_dim)
self.norm2 = nn.LayerNorm(time_embed_dim)
def forward(self, hidden_states):
hidden_states = self.norm1(hidden_states)
hidden_states = self.pool(hidden_states)
hidden_states = self.proj(hidden_states)
hidden_states = self.norm2(hidden_states)
return hidden_states
class TextImageTimeEmbedding(nn.Module):
def __init__(self, text_embed_dim: int = 768, image_embed_dim: int = 768, time_embed_dim: int = 1536):
super().__init__()
self.text_proj = nn.Linear(text_embed_dim, time_embed_dim)
self.text_norm = nn.LayerNorm(time_embed_dim)
self.image_proj = nn.Linear(image_embed_dim, time_embed_dim)
def forward(self, text_embeds: torch.FloatTensor, image_embeds: torch.FloatTensor):
# text
time_text_embeds = self.text_proj(text_embeds)
time_text_embeds = self.text_norm(time_text_embeds)
# image
time_image_embeds = self.image_proj(image_embeds)
return time_image_embeds + time_text_embeds
class ImageTimeEmbedding(nn.Module):
def __init__(self, image_embed_dim: int = 768, time_embed_dim: int = 1536):
super().__init__()
self.image_proj = nn.Linear(image_embed_dim, time_embed_dim)
self.image_norm = nn.LayerNorm(time_embed_dim)
def forward(self, image_embeds: torch.FloatTensor):
# image
time_image_embeds = self.image_proj(image_embeds)
time_image_embeds = self.image_norm(time_image_embeds)
return time_image_embeds
class ImageHintTimeEmbedding(nn.Module):
def __init__(self, image_embed_dim: int = 768, time_embed_dim: int = 1536):
super().__init__()
self.image_proj = nn.Linear(image_embed_dim, time_embed_dim)
self.image_norm = nn.LayerNorm(time_embed_dim)
self.input_hint_block = nn.Sequential(
nn.Conv2d(3, 16, 3, padding=1),
nn.SiLU(),
nn.Conv2d(16, 16, 3, padding=1),
nn.SiLU(),
nn.Conv2d(16, 32, 3, padding=1, stride=2),
nn.SiLU(),
nn.Conv2d(32, 32, 3, padding=1),
nn.SiLU(),
nn.Conv2d(32, 96, 3, padding=1, stride=2),
nn.SiLU(),
nn.Conv2d(96, 96, 3, padding=1),
nn.SiLU(),
nn.Conv2d(96, 256, 3, padding=1, stride=2),
nn.SiLU(),
nn.Conv2d(256, 4, 3, padding=1),
)
def forward(self, image_embeds: torch.FloatTensor, hint: torch.FloatTensor):
# image
time_image_embeds = self.image_proj(image_embeds)
time_image_embeds = self.image_norm(time_image_embeds)
hint = self.input_hint_block(hint)
return time_image_embeds, hint
class AttentionPooling(nn.Module):
# Copied from https://github.com/deep-floyd/IF/blob/2f91391f27dd3c468bf174be5805b4cc92980c0b/deepfloyd_if/model/nn.py#L54
def __init__(self, num_heads, embed_dim, dtype=None):
super().__init__()
self.dtype = dtype
self.positional_embedding = nn.Parameter(torch.randn(1, embed_dim) / embed_dim**0.5)
self.k_proj = nn.Linear(embed_dim, embed_dim, dtype=self.dtype)
self.q_proj = nn.Linear(embed_dim, embed_dim, dtype=self.dtype)
self.v_proj = nn.Linear(embed_dim, embed_dim, dtype=self.dtype)
self.num_heads = num_heads
self.dim_per_head = embed_dim // self.num_heads
def forward(self, x):
bs, length, width = x.size()
def shape(x):
# (bs, length, width) --> (bs, length, n_heads, dim_per_head)
x = x.view(bs, -1, self.num_heads, self.dim_per_head)
# (bs, length, n_heads, dim_per_head) --> (bs, n_heads, length, dim_per_head)
x = x.transpose(1, 2)
# (bs, n_heads, length, dim_per_head) --> (bs*n_heads, length, dim_per_head)
x = x.reshape(bs * self.num_heads, -1, self.dim_per_head)
# (bs*n_heads, length, dim_per_head) --> (bs*n_heads, dim_per_head, length)
x = x.transpose(1, 2)
return x
class_token = x.mean(dim=1, keepdim=True) + self.positional_embedding.to(x.dtype)
x = torch.cat([class_token, x], dim=1) # (bs, length+1, width)
# (bs*n_heads, class_token_length, dim_per_head)
q = shape(self.q_proj(class_token))
# (bs*n_heads, length+class_token_length, dim_per_head)
k = shape(self.k_proj(x))
v = shape(self.v_proj(x))
# (bs*n_heads, class_token_length, length+class_token_length):
scale = 1 / math.sqrt(math.sqrt(self.dim_per_head))
weight = torch.einsum("bct,bcs->bts", q * scale, k * scale) # More stable with f16 than dividing afterwards
weight = torch.softmax(weight.float(), dim=-1).type(weight.dtype)
# (bs*n_heads, dim_per_head, class_token_length)
a = torch.einsum("bts,bcs->bct", weight, v)
# (bs, length+1, width)
a = a.reshape(bs, -1, 1).transpose(1, 2)
return a[:, 0, :] # cls_token
class FourierEmbedder(nn.Module):
def __init__(self, num_freqs=64, temperature=100):
super().__init__()
self.num_freqs = num_freqs
self.temperature = temperature
freq_bands = temperature ** (torch.arange(num_freqs) / num_freqs)
freq_bands = freq_bands[None, None, None]
self.register_buffer("freq_bands", freq_bands, persistent=False)
def __call__(self, x):
x = self.freq_bands * x.unsqueeze(-1)
return torch.stack((x.sin(), x.cos()), dim=-1).permute(0, 1, 3, 4, 2).reshape(*x.shape[:2], -1)
class PositionNet(nn.Module):
def __init__(self, positive_len, out_dim, feature_type="text-only", fourier_freqs=8):
super().__init__()
self.positive_len = positive_len
self.out_dim = out_dim
self.fourier_embedder = FourierEmbedder(num_freqs=fourier_freqs)
self.position_dim = fourier_freqs * 2 * 4 # 2: sin/cos, 4: xyxy
if isinstance(out_dim, tuple):
out_dim = out_dim[0]
if feature_type == "text-only":
self.linears = nn.Sequential(
nn.Linear(self.positive_len + self.position_dim, 512),
nn.SiLU(),
nn.Linear(512, 512),
nn.SiLU(),
nn.Linear(512, out_dim),
)
self.null_positive_feature = torch.nn.Parameter(torch.zeros([self.positive_len]))
elif feature_type == "text-image":
self.linears_text = nn.Sequential(
nn.Linear(self.positive_len + self.position_dim, 512),
nn.SiLU(),
nn.Linear(512, 512),
nn.SiLU(),
nn.Linear(512, out_dim),
)
self.linears_image = nn.Sequential(
nn.Linear(self.positive_len + self.position_dim, 512),
nn.SiLU(),
nn.Linear(512, 512),
nn.SiLU(),
nn.Linear(512, out_dim),
)
self.null_text_feature = torch.nn.Parameter(torch.zeros([self.positive_len]))
self.null_image_feature = torch.nn.Parameter(torch.zeros([self.positive_len]))
self.null_position_feature = torch.nn.Parameter(torch.zeros([self.position_dim]))
def forward(
self,
boxes,
masks,
positive_embeddings=None,
phrases_masks=None,
image_masks=None,
phrases_embeddings=None,
image_embeddings=None,
):
masks = masks.unsqueeze(-1)
# embedding position (it may includes padding as placeholder)
xyxy_embedding = self.fourier_embedder(boxes) # B*N*4 -> B*N*C
# learnable null embedding
xyxy_null = self.null_position_feature.view(1, 1, -1)
# replace padding with learnable null embedding
xyxy_embedding = xyxy_embedding * masks + (1 - masks) * xyxy_null
# positionet with text only information
if positive_embeddings is not None:
# learnable null embedding
positive_null = self.null_positive_feature.view(1, 1, -1)
# replace padding with learnable null embedding
positive_embeddings = positive_embeddings * masks + (1 - masks) * positive_null
objs = self.linears(torch.cat([positive_embeddings, xyxy_embedding], dim=-1))
# positionet with text and image infomation
else:
phrases_masks = phrases_masks.unsqueeze(-1)
image_masks = image_masks.unsqueeze(-1)
# learnable null embedding
text_null = self.null_text_feature.view(1, 1, -1)
image_null = self.null_image_feature.view(1, 1, -1)
# replace padding with learnable null embedding
phrases_embeddings = phrases_embeddings * phrases_masks + (1 - phrases_masks) * text_null
image_embeddings = image_embeddings * image_masks + (1 - image_masks) * image_null
objs_text = self.linears_text(torch.cat([phrases_embeddings, xyxy_embedding], dim=-1))
objs_image = self.linears_image(torch.cat([image_embeddings, xyxy_embedding], dim=-1))
objs = torch.cat([objs_text, objs_image], dim=1)
return objs