forked from huggingface/diffusers
-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathattention.py
More file actions
402 lines (334 loc) · 14.7 KB
/
attention.py
File metadata and controls
402 lines (334 loc) · 14.7 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
# Copyright 2022 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 torch
import torch.nn.functional as F
from torch import nn
from ..utils.import_utils import is_xformers_available
from .cross_attention import CrossAttention
if is_xformers_available():
import xformers
import xformers.ops
else:
xformers = None
class AttentionBlock(nn.Module):
"""
An attention block that allows spatial positions to attend to each other. Originally ported from here, but adapted
to the N-d case.
https://github.com/hojonathanho/diffusion/blob/1e0dceb3b3495bbe19116a5e1b3596cd0706c543/diffusion_tf/models/unet.py#L66.
Uses three q, k, v linear layers to compute attention.
Parameters:
channels (`int`): The number of channels in the input and output.
num_head_channels (`int`, *optional*):
The number of channels in each head. If None, then `num_heads` = 1.
norm_num_groups (`int`, *optional*, defaults to 32): The number of groups to use for group norm.
rescale_output_factor (`float`, *optional*, defaults to 1.0): The factor to rescale the output by.
eps (`float`, *optional*, defaults to 1e-5): The epsilon value to use for group norm.
"""
# IMPORTANT;TODO(Patrick, William) - this class will be deprecated soon. Do not use it anymore
def __init__(
self,
channels: int,
num_head_channels: Optional[int] = None,
norm_num_groups: int = 32,
rescale_output_factor: float = 1.0,
eps: float = 1e-5,
):
super().__init__()
self.channels = channels
self.num_heads = channels // num_head_channels if num_head_channels is not None else 1
self.num_head_size = num_head_channels
self.group_norm = nn.GroupNorm(num_channels=channels, num_groups=norm_num_groups, eps=eps, affine=True)
# define q,k,v as linear layers
self.query = nn.Linear(channels, channels)
self.key = nn.Linear(channels, channels)
self.value = nn.Linear(channels, channels)
self.rescale_output_factor = rescale_output_factor
self.proj_attn = nn.Linear(channels, channels, 1)
self._use_memory_efficient_attention_xformers = False
def reshape_heads_to_batch_dim(self, tensor):
batch_size, seq_len, dim = tensor.shape
head_size = self.num_heads
tensor = tensor.reshape(batch_size, seq_len, head_size, dim // head_size)
tensor = tensor.permute(0, 2, 1, 3).reshape(batch_size * head_size, seq_len, dim // head_size)
return tensor
def reshape_batch_dim_to_heads(self, tensor):
batch_size, seq_len, dim = tensor.shape
head_size = self.num_heads
tensor = tensor.reshape(batch_size // head_size, head_size, seq_len, dim)
tensor = tensor.permute(0, 2, 1, 3).reshape(batch_size // head_size, seq_len, dim * head_size)
return tensor
def set_use_memory_efficient_attention_xformers(self, use_memory_efficient_attention_xformers: bool):
if use_memory_efficient_attention_xformers:
if not is_xformers_available():
raise ModuleNotFoundError(
"Refer to https://github.com/facebookresearch/xformers for more information on how to install"
" xformers",
name="xformers",
)
elif not torch.cuda.is_available():
raise ValueError(
"torch.cuda.is_available() should be True but is False. xformers' memory efficient attention is"
" only available for GPU "
)
else:
try:
# Make sure we can run the memory efficient attention
_ = xformers.ops.memory_efficient_attention(
torch.randn((1, 2, 40), device="cuda"),
torch.randn((1, 2, 40), device="cuda"),
torch.randn((1, 2, 40), device="cuda"),
)
except Exception as e:
raise e
self._use_memory_efficient_attention_xformers = use_memory_efficient_attention_xformers
def forward(self, hidden_states):
residual = hidden_states
batch, channel, height, width = hidden_states.shape
# norm
hidden_states = self.group_norm(hidden_states)
hidden_states = hidden_states.view(batch, channel, height * width).transpose(1, 2)
# proj to q, k, v
query_proj = self.query(hidden_states)
key_proj = self.key(hidden_states)
value_proj = self.value(hidden_states)
scale = 1 / math.sqrt(self.channels / self.num_heads)
query_proj = self.reshape_heads_to_batch_dim(query_proj)
key_proj = self.reshape_heads_to_batch_dim(key_proj)
value_proj = self.reshape_heads_to_batch_dim(value_proj)
if self._use_memory_efficient_attention_xformers:
# Memory efficient attention
hidden_states = xformers.ops.memory_efficient_attention(query_proj, key_proj, value_proj, attn_bias=None)
hidden_states = hidden_states.to(query_proj.dtype)
else:
attention_scores = torch.baddbmm(
torch.empty(
query_proj.shape[0],
query_proj.shape[1],
key_proj.shape[1],
dtype=query_proj.dtype,
device=query_proj.device,
),
query_proj,
key_proj.transpose(-1, -2),
beta=0,
alpha=scale,
)
attention_probs = torch.softmax(attention_scores.float(), dim=-1).type(attention_scores.dtype)
hidden_states = torch.bmm(attention_probs, value_proj)
# reshape hidden_states
hidden_states = self.reshape_batch_dim_to_heads(hidden_states)
# compute next hidden_states
hidden_states = self.proj_attn(hidden_states)
hidden_states = hidden_states.transpose(-1, -2).reshape(batch, channel, height, width)
# res connect and rescale
hidden_states = (hidden_states + residual) / self.rescale_output_factor
return hidden_states
class BasicTransformerBlock(nn.Module):
r"""
A basic Transformer block.
Parameters:
dim (`int`): The number of channels in the input and output.
num_attention_heads (`int`): The number of heads to use for multi-head attention.
attention_head_dim (`int`): The number of channels in each head.
dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention.
activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
num_embeds_ada_norm (:
obj: `int`, *optional*): The number of diffusion steps used during training. See `Transformer2DModel`.
attention_bias (:
obj: `bool`, *optional*, defaults to `False`): Configure if the attentions should contain a bias parameter.
"""
def __init__(
self,
dim: int,
num_attention_heads: int,
attention_head_dim: int,
dropout=0.0,
cross_attention_dim: Optional[int] = None,
activation_fn: str = "geglu",
num_embeds_ada_norm: Optional[int] = None,
attention_bias: bool = False,
only_cross_attention: bool = False,
upcast_attention: bool = False,
):
super().__init__()
self.only_cross_attention = only_cross_attention
self.use_ada_layer_norm = num_embeds_ada_norm is not None
# 1. Self-Attn
self.attn1 = CrossAttention(
query_dim=dim,
heads=num_attention_heads,
dim_head=attention_head_dim,
dropout=dropout,
bias=attention_bias,
cross_attention_dim=cross_attention_dim if only_cross_attention else None,
upcast_attention=upcast_attention,
)
self.ff = FeedForward(dim, dropout=dropout, activation_fn=activation_fn)
# 2. Cross-Attn
if cross_attention_dim is not None:
self.attn2 = CrossAttention(
query_dim=dim,
cross_attention_dim=cross_attention_dim,
heads=num_attention_heads,
dim_head=attention_head_dim,
dropout=dropout,
bias=attention_bias,
upcast_attention=upcast_attention,
) # is self-attn if encoder_hidden_states is none
else:
self.attn2 = None
self.norm1 = AdaLayerNorm(dim, num_embeds_ada_norm) if self.use_ada_layer_norm else nn.LayerNorm(dim)
if cross_attention_dim is not None:
self.norm2 = AdaLayerNorm(dim, num_embeds_ada_norm) if self.use_ada_layer_norm else nn.LayerNorm(dim)
else:
self.norm2 = None
# 3. Feed-forward
self.norm3 = nn.LayerNorm(dim)
def forward(
self,
hidden_states,
encoder_hidden_states=None,
timestep=None,
attention_mask=None,
cross_attention_kwargs=None,
):
# 1. Self-Attention
norm_hidden_states = (
self.norm1(hidden_states, timestep) if self.use_ada_layer_norm else self.norm1(hidden_states)
)
cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {}
attn_output = self.attn1(
norm_hidden_states,
encoder_hidden_states=encoder_hidden_states if self.only_cross_attention else None,
attention_mask=attention_mask,
**cross_attention_kwargs,
)
hidden_states = attn_output + hidden_states
if self.attn2 is not None:
# 2. Cross-Attention
norm_hidden_states = (
self.norm2(hidden_states, timestep) if self.use_ada_layer_norm else self.norm2(hidden_states)
)
attn_output = self.attn2(
norm_hidden_states,
encoder_hidden_states=encoder_hidden_states,
attention_mask=attention_mask,
**cross_attention_kwargs,
)
hidden_states = attn_output + hidden_states
# 3. Feed-forward
hidden_states = self.ff(self.norm3(hidden_states)) + hidden_states
return hidden_states
class FeedForward(nn.Module):
r"""
A feed-forward layer.
Parameters:
dim (`int`): The number of channels in the input.
dim_out (`int`, *optional*): The number of channels in the output. If not given, defaults to `dim`.
mult (`int`, *optional*, defaults to 4): The multiplier to use for the hidden dimension.
dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
"""
def __init__(
self,
dim: int,
dim_out: Optional[int] = None,
mult: int = 4,
dropout: float = 0.0,
activation_fn: str = "geglu",
):
super().__init__()
inner_dim = int(dim * mult)
dim_out = dim_out if dim_out is not None else dim
if activation_fn == "gelu":
act_fn = GELU(dim, inner_dim)
elif activation_fn == "geglu":
act_fn = GEGLU(dim, inner_dim)
elif activation_fn == "geglu-approximate":
act_fn = ApproximateGELU(dim, inner_dim)
self.net = nn.ModuleList([])
# project in
self.net.append(act_fn)
# project dropout
self.net.append(nn.Dropout(dropout))
# project out
self.net.append(nn.Linear(inner_dim, dim_out))
def forward(self, hidden_states):
for module in self.net:
hidden_states = module(hidden_states)
return hidden_states
class GELU(nn.Module):
r"""
GELU activation function
"""
def __init__(self, dim_in: int, dim_out: int):
super().__init__()
self.proj = nn.Linear(dim_in, dim_out)
def gelu(self, gate):
if gate.device.type != "mps":
return F.gelu(gate)
# mps: gelu is not implemented for float16
return F.gelu(gate.to(dtype=torch.float32)).to(dtype=gate.dtype)
def forward(self, hidden_states):
hidden_states = self.proj(hidden_states)
hidden_states = self.gelu(hidden_states)
return hidden_states
# feedforward
class GEGLU(nn.Module):
r"""
A variant of the gated linear unit activation function from https://arxiv.org/abs/2002.05202.
Parameters:
dim_in (`int`): The number of channels in the input.
dim_out (`int`): The number of channels in the output.
"""
def __init__(self, dim_in: int, dim_out: int):
super().__init__()
self.proj = nn.Linear(dim_in, dim_out * 2)
def gelu(self, gate):
if gate.device.type != "mps":
return F.gelu(gate)
# mps: gelu is not implemented for float16
return F.gelu(gate.to(dtype=torch.float32)).to(dtype=gate.dtype)
def forward(self, hidden_states):
hidden_states, gate = self.proj(hidden_states).chunk(2, dim=-1)
return hidden_states * self.gelu(gate)
class ApproximateGELU(nn.Module):
"""
The approximate form of Gaussian Error Linear Unit (GELU)
For more details, see section 2: https://arxiv.org/abs/1606.08415
"""
def __init__(self, dim_in: int, dim_out: int):
super().__init__()
self.proj = nn.Linear(dim_in, dim_out)
def forward(self, x):
x = self.proj(x)
return x * torch.sigmoid(1.702 * x)
class AdaLayerNorm(nn.Module):
"""
Norm layer modified to incorporate timestep embeddings.
"""
def __init__(self, embedding_dim, num_embeddings):
super().__init__()
self.emb = nn.Embedding(num_embeddings, embedding_dim)
self.silu = nn.SiLU()
self.linear = nn.Linear(embedding_dim, embedding_dim * 2)
self.norm = nn.LayerNorm(embedding_dim, elementwise_affine=False)
def forward(self, x, timestep):
emb = self.linear(self.silu(self.emb(timestep)))
scale, shift = torch.chunk(emb, 2)
x = self.norm(x) * (1 + scale) + shift
return x