File size: 16,368 Bytes
9b9e0ee
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.

# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
# --------------------------------------------------------
# References:
# GLIDE: https://github.com/openai/glide-text2im
# MAE: https://github.com/facebookresearch/mae/blob/main/models_mae.py
# --------------------------------------------------------
import math
import torch
import torch.nn as nn
from timm.models.vision_transformer import Mlp, Attention as Attention_
from einops import rearrange, repeat
import xformers.ops

from .utils import add_decomposed_rel_pos


def modulate(x, shift, scale):
    return x * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1)


def t2i_modulate(x, shift, scale):
    return x * (1 + scale) + shift


class MultiHeadCrossAttention(nn.Module):
    def __init__(self, d_model, num_heads, attn_drop=0., proj_drop=0., **block_kwargs):
        super(MultiHeadCrossAttention, self).__init__()
        assert d_model % num_heads == 0, "d_model must be divisible by num_heads"

        self.d_model = d_model
        self.num_heads = num_heads
        self.head_dim = d_model // num_heads

        self.q_linear = nn.Linear(d_model, d_model)
        self.kv_linear = nn.Linear(d_model, d_model*2)
        self.attn_drop = nn.Dropout(attn_drop)
        self.proj = nn.Linear(d_model, d_model)
        self.proj_drop = nn.Dropout(proj_drop)

    def forward(self, x, cond, mask=None):
        # query/value: img tokens; key: condition; mask: if padding tokens
        B, N, C = x.shape

        q = self.q_linear(x).view(1, -1, self.num_heads, self.head_dim)
        # import pdb
        # pdb.set_trace()
        kv = self.kv_linear(cond).view(1, -1, 2, self.num_heads, self.head_dim)
        k, v = kv.unbind(2)
        attn_bias = None
        assert mask is not None
        # import pdb 
        # pdb.set_trace()
        attn_bias = xformers.ops.fmha.BlockDiagonalMask.from_seqlens([N] * B, mask)

        # attn_bias = torch.zeros([B * self.num_heads, q.shape[1], k.shape[1]], dtype=q.dtype, device=q.device)
        # import pdb 
        # pdb.set_trace()
        # attn_bias.masked_fill_(mask.squeeze(1).repeat(self.num_heads, 1, 1) == 0, float('-inf'))

        x = xformers.ops.memory_efficient_attention(q, k, v, p=self.attn_drop.p, attn_bias=attn_bias)

        x = x.view(B, -1, C)
        x = self.proj(x)
        x = self.proj_drop(x)

        # q = self.q_linear(x).reshape(B, -1, self.num_heads, self.head_dim)
        # kv = self.kv_linear(cond).reshape(B, -1, 2, self.num_heads, self.head_dim)
        # k, v = kv.unbind(2)
        # attn_bias = None
        # if mask is not None:

        

        # x = xformers.ops.memory_efficient_attention(q, k, v, p=self.attn_drop.p, attn_bias=attn_bias)
        # x = x.contiguous().reshape(B, -1, C)
        # x = self.proj(x)
        # x = self.proj_drop(x)

        return x


class WindowAttention(Attention_):
    """Multi-head Attention block with relative position embeddings."""

    def __init__(
        self,
        dim,
        num_heads=8,
        qkv_bias=True,
        use_rel_pos=False,
        rel_pos_zero_init=True,
        input_size=None,
        **block_kwargs,
    ):
        """
        Args:
            dim (int): Number of input channels.
            num_heads (int): Number of attention heads.
            qkv_bias (bool:  If True, add a learnable bias to query, key, value.
            rel_pos (bool): If True, add relative positional embeddings to the attention map.
            rel_pos_zero_init (bool): If True, zero initialize relative positional parameters.
            input_size (int or None): Input resolution for calculating the relative positional
                parameter size.
        """
        super().__init__(dim, num_heads=num_heads, qkv_bias=qkv_bias, **block_kwargs)

        self.use_rel_pos = use_rel_pos
        if self.use_rel_pos:
            # initialize relative positional embeddings
            self.rel_pos_h = nn.Parameter(torch.zeros(2 * input_size[0] - 1, self.head_dim))
            self.rel_pos_w = nn.Parameter(torch.zeros(2 * input_size[1] - 1, self.head_dim))

            if not rel_pos_zero_init:
                nn.init.trunc_normal_(self.rel_pos_h, std=0.02)
                nn.init.trunc_normal_(self.rel_pos_w, std=0.02)

    def forward(self, x, mask=None):
        B, N, C = x.shape
        qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads)
        q, k, v = qkv.unbind(2)
        if use_fp32_attention := getattr(self, 'fp32_attention', False):
            q, k, v = q.float(), k.float(), v.float()

        attn_bias = None
        if mask is not None:
            attn_bias = torch.zeros([B * self.num_heads, q.shape[1], k.shape[1]], dtype=q.dtype, device=q.device)
            attn_bias.masked_fill_(mask.squeeze(1).repeat(self.num_heads, 1, 1) == 0, float('-inf'))
        # import pdb
        # pdb.set_trace()
        x = xformers.ops.memory_efficient_attention(q, k, v, p=self.attn_drop.p, attn_bias=attn_bias)

        x = x.view(B, N, C)
        x = self.proj(x)
        x = self.proj_drop(x)
        return x


#################################################################################
#   AMP attention with fp32 softmax to fix loss NaN problem during training     #
#################################################################################
class Attention(Attention_):
    def forward(self, x):
        B, N, C = x.shape
        qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
        q, k, v = qkv.unbind(0)  # make torchscript happy (cannot use tensor as tuple)
        use_fp32_attention = getattr(self, 'fp32_attention', False)
        if use_fp32_attention:
            q, k = q.float(), k.float()
        with torch.cuda.amp.autocast(enabled=not use_fp32_attention):
            attn = (q @ k.transpose(-2, -1)) * self.scale
            attn = attn.softmax(dim=-1)

        attn = self.attn_drop(attn)

        x = (attn @ v).transpose(1, 2).reshape(B, N, C)
        x = self.proj(x)
        x = self.proj_drop(x)
        return x


class FinalLayer(nn.Module):
    """
    The final layer of PixArt.
    """

    def __init__(self, hidden_size, patch_size, out_channels):
        super().__init__()
        self.norm_final = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
        self.linear = nn.Linear(hidden_size, patch_size * patch_size * out_channels, bias=True)
        self.adaLN_modulation = nn.Sequential(
            nn.SiLU(),
            nn.Linear(hidden_size, 2 * hidden_size, bias=True)
        )

    def forward(self, x, c):
        shift, scale = self.adaLN_modulation(c).chunk(2, dim=1)
        x = modulate(self.norm_final(x), shift, scale)
        x = self.linear(x)
        return x


class T2IFinalLayer(nn.Module):
    """
    The final layer of PixArt.
    """

    def __init__(self, hidden_size, patch_size, out_channels):
        super().__init__()
        self.norm_final = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
        self.linear = nn.Linear(hidden_size, patch_size[0] * patch_size[1] * out_channels, bias=True)
        self.scale_shift_table = nn.Parameter(torch.randn(2, hidden_size) / hidden_size ** 0.5)
        self.out_channels = out_channels

    def forward(self, x, t):
        shift, scale = (self.scale_shift_table[None] + t[:, None]).chunk(2, dim=1)
        x = t2i_modulate(self.norm_final(x), shift, scale)
        x = self.linear(x)
        return x

class MaskFinalLayer(nn.Module):
    """
    The final layer of PixArt.
    """

    def __init__(self, final_hidden_size, c_emb_size, patch_size, out_channels):
        super().__init__()
        self.norm_final = nn.LayerNorm(final_hidden_size, elementwise_affine=False, eps=1e-6)
        self.linear = nn.Linear(final_hidden_size, patch_size * patch_size * out_channels, bias=True)
        self.adaLN_modulation = nn.Sequential(
            nn.SiLU(),
            nn.Linear(c_emb_size, 2 * final_hidden_size, bias=True)
        )
    def forward(self, x, t):
        shift, scale = self.adaLN_modulation(t).chunk(2, dim=1)
        x = modulate(self.norm_final(x), shift, scale)
        x = self.linear(x)
        return x


class DecoderLayer(nn.Module):
    """
    The final layer of PixArt.
    """

    def __init__(self, hidden_size, decoder_hidden_size):
        super().__init__()
        self.norm_decoder = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
        self.linear = nn.Linear(hidden_size, decoder_hidden_size, bias=True)
        self.adaLN_modulation = nn.Sequential(
            nn.SiLU(),
            nn.Linear(hidden_size, 2 * hidden_size, bias=True)
        )
    def forward(self, x, t):
        shift, scale = self.adaLN_modulation(t).chunk(2, dim=1)
        x = modulate(self.norm_decoder(x), shift, scale)
        x = self.linear(x)
        return x


#################################################################################
#               Embedding Layers for Timesteps and Class Labels                 #
#################################################################################
class TimestepEmbedder(nn.Module):
    """
    Embeds scalar timesteps into vector representations.
    """

    def __init__(self, hidden_size, frequency_embedding_size=256):
        super().__init__()
        self.mlp = nn.Sequential(
            nn.Linear(frequency_embedding_size, hidden_size, bias=True),
            nn.SiLU(),
            nn.Linear(hidden_size, hidden_size, bias=True),
        )
        self.frequency_embedding_size = frequency_embedding_size

    @staticmethod
    def timestep_embedding(t, dim, max_period=10000):
        """
        Create sinusoidal timestep embeddings.
        :param t: a 1-D Tensor of N indices, one per batch element.
                          These may be fractional.
        :param dim: the dimension of the output.
        :param max_period: controls the minimum frequency of the embeddings.
        :return: an (N, D) Tensor of positional embeddings.
        """
        # https://github.com/openai/glide-text2im/blob/main/glide_text2im/nn.py
        half = dim // 2
        freqs = torch.exp(
            -math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32, device=t.device) / half)
        args = t[:, None].float() * freqs[None]
        embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
        if dim % 2:
            embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
        return embedding

    def forward(self, t):
        t_freq = self.timestep_embedding(t, self.frequency_embedding_size).to(self.dtype)
        return self.mlp(t_freq)

    @property
    def dtype(self):
        # 返回模型参数的数据类型
        return next(self.parameters()).dtype


class SizeEmbedder(TimestepEmbedder):
    """
    Embeds scalar timesteps into vector representations.
    """

    def __init__(self, hidden_size, frequency_embedding_size=256):
        super().__init__(hidden_size=hidden_size, frequency_embedding_size=frequency_embedding_size)
        self.mlp = nn.Sequential(
            nn.Linear(frequency_embedding_size, hidden_size, bias=True),
            nn.SiLU(),
            nn.Linear(hidden_size, hidden_size, bias=True),
        )
        self.frequency_embedding_size = frequency_embedding_size
        self.outdim = hidden_size

    def forward(self, s, bs):
        if s.ndim == 1:
            s = s[:, None]
        assert s.ndim == 2
        if s.shape[0] != bs:
            s = s.repeat(bs//s.shape[0], 1)
            assert s.shape[0] == bs
        b, dims = s.shape[0], s.shape[1]
        s = rearrange(s, "b d -> (b d)")
        s_freq = self.timestep_embedding(s, self.frequency_embedding_size).to(self.dtype)
        s_emb = self.mlp(s_freq)
        s_emb = rearrange(s_emb, "(b d) d2 -> b (d d2)", b=b, d=dims, d2=self.outdim)
        return s_emb

    @property
    def dtype(self):
        # 返回模型参数的数据类型
        return next(self.parameters()).dtype


class LabelEmbedder(nn.Module):
    """
    Embeds class labels into vector representations. Also handles label dropout for classifier-free guidance.
    """

    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]).cuda() < self.dropout_prob
        else:
            drop_ids = force_drop_ids == 1
        labels = torch.where(drop_ids, self.num_classes, labels)
        return labels

    def forward(self, labels, train, force_drop_ids=None):
        use_dropout = self.dropout_prob > 0
        if (train and use_dropout) or (force_drop_ids is not None):
            labels = self.token_drop(labels, force_drop_ids)
        return self.embedding_table(labels)


class CaptionEmbedder(nn.Module):
    """
    Embeds class labels into vector representations. Also handles label dropout for classifier-free guidance.
    """

    def __init__(self, in_channels, hidden_size, uncond_prob, act_layer=nn.GELU(approximate='tanh'), token_num=120):
        super().__init__()
        self.y_proj = Mlp(in_features=in_channels, hidden_features=hidden_size, out_features=hidden_size, act_layer=act_layer, drop=0)
        self.register_buffer("y_embedding", nn.Parameter(torch.randn(token_num, in_channels) / in_channels ** 0.5))
        self.uncond_prob = uncond_prob

    def token_drop(self, caption, force_drop_ids=None):
        """
        Drops labels to enable classifier-free guidance.
        """
        if force_drop_ids is None:
            drop_ids = torch.rand(caption.shape[0]).cuda() < self.uncond_prob
        else:
            drop_ids = force_drop_ids == 1
        caption = torch.where(drop_ids[:, None, None, None], self.y_embedding, caption)
        return caption

    def forward(self, caption, train, force_drop_ids=None):
        if train:
            assert caption.shape[2:] == self.y_embedding.shape
        use_dropout = self.uncond_prob > 0
        if (train and use_dropout) or (force_drop_ids is not None):
            caption = self.token_drop(caption, force_drop_ids)
        caption = self.y_proj(caption)
        return caption


class CaptionEmbedderDoubleBr(nn.Module):
    """
    Embeds class labels into vector representations. Also handles label dropout for classifier-free guidance.
    """

    def __init__(self, in_channels, hidden_size, uncond_prob, act_layer=nn.GELU(approximate='tanh'), token_num=120):
        super().__init__()
        self.proj = Mlp(in_features=in_channels, hidden_features=hidden_size, out_features=hidden_size, act_layer=act_layer, drop=0)
        self.embedding = nn.Parameter(torch.randn(1, in_channels) / 10 ** 0.5)
        self.y_embedding = nn.Parameter(torch.randn(token_num, in_channels) / 10 ** 0.5)
        self.uncond_prob = uncond_prob

    def token_drop(self, global_caption, caption, force_drop_ids=None):
        """
        Drops labels to enable classifier-free guidance.
        """
        if force_drop_ids is None:
            drop_ids = torch.rand(global_caption.shape[0]).cuda() < self.uncond_prob
        else:
            drop_ids = force_drop_ids == 1
        global_caption = torch.where(drop_ids[:, None], self.embedding, global_caption)
        caption = torch.where(drop_ids[:, None, None, None], self.y_embedding, caption)
        return global_caption, caption

    def forward(self, caption, train, force_drop_ids=None):
        assert caption.shape[2: ] == self.y_embedding.shape
        global_caption = caption.mean(dim=2).squeeze()
        use_dropout = self.uncond_prob > 0
        if (train and use_dropout) or (force_drop_ids is not None):
            global_caption, caption = self.token_drop(global_caption, caption, force_drop_ids)
        y_embed = self.proj(global_caption)
        return y_embed, caption