{{ message }}
Repository navigation
Expand file tree
/
Copy pathattention.py
More file actions
366 lines (305 loc) · 13.8 KB
/
Copy pathattention.py
File metadata and controls
366 lines (305 loc) · 13.8 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
# 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.
from typing import Any, Dict, Optional
import torch
import torch.nn.functional as F
from torch import nn
from ..utils import maybe_allow_in_graph
from .activations import get_activation
from .attention_processor import Attention
from .embeddings import CombinedTimestepLabelEmbeddings
@maybe_allow_in_graph
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.
only_cross_attention (`bool`, *optional*):
Whether to use only cross-attention layers. In this case two cross attention layers are used.
double_self_attention (`bool`, *optional*):
Whether to use two self-attention layers. In this case no cross attention layers are used.
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,
double_self_attention: bool = False,
upcast_attention: bool = False,
norm_elementwise_affine: bool = True,
norm_type: str = "layer_norm",
final_dropout: bool = False,
):
super().__init__()
self.only_cross_attention = only_cross_attention
self.use_ada_layer_norm_zero = (num_embeds_ada_norm is not None) and norm_type == "ada_norm_zero"
self.use_ada_layer_norm = (num_embeds_ada_norm is not None) and norm_type == "ada_norm"
if norm_type in ("ada_norm", "ada_norm_zero") and num_embeds_ada_norm is None:
raise ValueError(
f"`norm_type` is set to {norm_type}, but `num_embeds_ada_norm` is not defined. Please make sure to"
f" define `num_embeds_ada_norm` if setting `norm_type` to {norm_type}."
)
# Define 3 blocks. Each block has its own normalization layer.
# 1. Self-Attn
if self.use_ada_layer_norm:
self.norm1 = AdaLayerNorm(dim, num_embeds_ada_norm)
elif self.use_ada_layer_norm_zero:
self.norm1 = AdaLayerNormZero(dim, num_embeds_ada_norm)
else:
self.norm1 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine)
self.attn1 = Attention(
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,
)
# 2. Cross-Attn
if cross_attention_dim is not None or double_self_attention:
# We currently only use AdaLayerNormZero for self attention where there will only be one attention block.
# I.e. the number of returned modulation chunks from AdaLayerZero would not make sense if returned during
# the second cross attention block.
self.norm2 = (
AdaLayerNorm(dim, num_embeds_ada_norm)
if self.use_ada_layer_norm
else nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine)
)
self.attn2 = Attention(
query_dim=dim,
cross_attention_dim=cross_attention_dim if not double_self_attention else None,
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.norm2 = None
self.attn2 = None
# 3. Feed-forward
self.norm3 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine)
self.ff = FeedForward(dim, dropout=dropout, activation_fn=activation_fn, final_dropout=final_dropout)
def forward(
self,
hidden_states: torch.FloatTensor,
attention_mask: Optional[torch.FloatTensor] = None,
encoder_hidden_states: Optional[torch.FloatTensor] = None,
encoder_attention_mask: Optional[torch.FloatTensor] = None,
timestep: Optional[torch.LongTensor] = None,
cross_attention_kwargs: Dict[str, Any] = None,
class_labels: Optional[torch.LongTensor] = None,
):
# Notice that normalization is always applied before the real computation in the following blocks.
# 1. Self-Attention
if self.use_ada_layer_norm:
norm_hidden_states = self.norm1(hidden_states, timestep)
elif self.use_ada_layer_norm_zero:
norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1(
hidden_states, timestep, class_labels, hidden_dtype=hidden_states.dtype
)
else:
norm_hidden_states = 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,
)
if self.use_ada_layer_norm_zero:
attn_output = gate_msa.unsqueeze(1) * attn_output
hidden_states = attn_output + hidden_states
# 2. Cross-Attention
if self.attn2 is not None:
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=encoder_attention_mask,
**cross_attention_kwargs,
)
hidden_states = attn_output + hidden_states
# 3. Feed-forward
norm_hidden_states = self.norm3(hidden_states)
if self.use_ada_layer_norm_zero:
norm_hidden_states = norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None]
ff_output = self.ff(norm_hidden_states)
if self.use_ada_layer_norm_zero:
ff_output = gate_mlp.unsqueeze(1) * ff_output
hidden_states = ff_output + 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.
final_dropout (`bool` *optional*, defaults to False): Apply a final dropout.
"""
def __init__(
self,
dim: int,
dim_out: Optional[int] = None,
mult: int = 4,
dropout: float = 0.0,
activation_fn: str = "geglu",
final_dropout: bool = False,
):
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)
if activation_fn == "gelu-approximate":
act_fn = GELU(dim, inner_dim, approximate="tanh")
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))
# FF as used in Vision Transformer, MLP-Mixer, etc. have a final dropout
if final_dropout:
self.net.append(nn.Dropout(dropout))
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 with tanh approximation support with `approximate="tanh"`.
"""
def __init__(self, dim_in: int, dim_out: int, approximate: str = "none"):
super().__init__()
self.proj = nn.Linear(dim_in, dim_out)
self.approximate = approximate
def gelu(self, gate):
if gate.device.type != "mps":
return F.gelu(gate, approximate=self.approximate)
# mps: gelu is not implemented for float16
return F.gelu(gate.to(dtype=torch.float32), approximate=self.approximate).to(dtype=gate.dtype)
def forward(self, hidden_states):
hidden_states = self.proj(hidden_states)
hidden_states = self.gelu(hidden_states)
return hidden_states
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
class AdaLayerNormZero(nn.Module):
"""
Norm layer adaptive layer norm zero (adaLN-Zero).
"""
def __init__(self, embedding_dim, num_embeddings):
super().__init__()
self.emb = CombinedTimestepLabelEmbeddings(num_embeddings, embedding_dim)
self.silu = nn.SiLU()
self.linear = nn.Linear(embedding_dim, 6 * embedding_dim, bias=True)
self.norm = nn.LayerNorm(embedding_dim, elementwise_affine=False, eps=1e-6)
def forward(self, x, timestep, class_labels, hidden_dtype=None):
emb = self.linear(self.silu(self.emb(timestep, class_labels, hidden_dtype=hidden_dtype)))
shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = emb.chunk(6, dim=1)
x = self.norm(x) * (1 + scale_msa[:, None]) + shift_msa[:, None]
return x, gate_msa, shift_mlp, scale_mlp, gate_mlp
class AdaGroupNorm(nn.Module):
"""
GroupNorm layer modified to incorporate timestep embeddings.
"""
def __init__(
self, embedding_dim: int, out_dim: int, num_groups: int, act_fn: Optional[str] = None, eps: float = 1e-5
):
super().__init__()
self.num_groups = num_groups
self.eps = eps
if act_fn is None:
self.act = None
else:
self.act = get_activation(act_fn)
self.linear = nn.Linear(embedding_dim, out_dim * 2)
def forward(self, x, emb):
if self.act:
emb = self.act(emb)
emb = self.linear(emb)
emb = emb[:, :, None, None]
scale, shift = emb.chunk(2, dim=1)
x = F.group_norm(x, self.num_groups, eps=self.eps)
x = x * (1 + scale) + shift
return x
You can’t perform that action at this time.