-
Notifications
You must be signed in to change notification settings - Fork 75
/
attention.py
294 lines (225 loc) · 11.2 KB
/
attention.py
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
import torch
import torch.nn as nn
import torch.nn.functional as F
from diffusers.models.lora import LoRALinearLayer
from functions import AttentionMLP
from diffusers.utils.import_utils import is_xformers_available
if is_xformers_available():
import xformers
class FuseModule(nn.Module):
def __init__(self, embed_dim):
super().__init__()
self.mlp1 = MLP(embed_dim * 2, embed_dim, embed_dim, use_residual=False)
self.mlp2 = MLP(embed_dim, embed_dim, embed_dim, use_residual=True)
self.layer_norm = nn.LayerNorm(embed_dim)
def fuse_fn(self, prompt_embeds, id_embeds):
stacked_id_embeds = torch.cat([prompt_embeds, id_embeds], dim=-1)
stacked_id_embeds = self.mlp1(stacked_id_embeds) + prompt_embeds
stacked_id_embeds = self.mlp2(stacked_id_embeds)
stacked_id_embeds = self.layer_norm(stacked_id_embeds)
return stacked_id_embeds
def forward(
self,
prompt_embeds,
id_embeds,
class_tokens_mask,
valid_id_mask,
) -> torch.Tensor:
id_embeds = id_embeds.to(prompt_embeds.dtype)
batch_size, max_num_inputs = id_embeds.shape[:2]
seq_length = prompt_embeds.shape[1]
flat_id_embeds = id_embeds.view(-1, id_embeds.shape[-2], id_embeds.shape[-1])
valid_id_embeds = flat_id_embeds[valid_id_mask.flatten()]
prompt_embeds = prompt_embeds.view(-1, prompt_embeds.shape[-1])
class_tokens_mask = class_tokens_mask.view(-1)
valid_id_embeds = valid_id_embeds.view(-1, valid_id_embeds.shape[-1])
image_token_embeds = prompt_embeds[class_tokens_mask]
stacked_id_embeds = self.fuse_fn(image_token_embeds, valid_id_embeds)
assert class_tokens_mask.sum() == stacked_id_embeds.shape[0], f"{class_tokens_mask.sum()} != {stacked_id_embeds.shape[0]}"
prompt_embeds.masked_scatter_(class_tokens_mask[:, None], stacked_id_embeds.to(prompt_embeds.dtype))
updated_prompt_embeds = prompt_embeds.view(batch_size, seq_length, -1)
return updated_prompt_embeds
class MLP(nn.Module):
def __init__(self, in_dim, out_dim, hidden_dim, use_residual=True):
super().__init__()
if use_residual:
assert in_dim == out_dim
self.layernorm = nn.LayerNorm(in_dim)
self.fc1 = nn.Linear(in_dim, hidden_dim)
self.fc2 = nn.Linear(hidden_dim, out_dim)
self.use_residual = use_residual
self.act_fn = nn.GELU()
def forward(self, x):
residual = x
x = self.layernorm(x)
x = self.fc1(x)
x = self.act_fn(x)
x = self.fc2(x)
if self.use_residual:
x = x + residual
return x
class FacialEncoder(nn.Module):
def __init__(self,image_CLIPModel_encoder=None):
super().__init__()
self.visual_projection = AttentionMLP()
self.fuse_module = FuseModule(768)
def forward(self, prompt_embeds, multi_image_embeds, class_tokens_mask, valid_id_mask):
bs, num_inputs, token_length, image_dim = multi_image_embeds.shape
multi_image_embeds_view = multi_image_embeds.view(bs * num_inputs, token_length, image_dim)
id_embeds = self.visual_projection(multi_image_embeds_view)
id_embeds = id_embeds.view(bs, num_inputs, 1, -1)
updated_prompt_embeds = self.fuse_module(prompt_embeds, id_embeds, class_tokens_mask, valid_id_mask)
return updated_prompt_embeds
class Consistent_AttProcessor(nn.Module):
def __init__(
self,
hidden_size=None,
cross_attention_dim=None,
rank=4,
network_alpha=None,
lora_scale=1.0,
):
super().__init__()
self.rank = rank
self.lora_scale = lora_scale
self.to_q_lora = LoRALinearLayer(hidden_size, hidden_size, rank, network_alpha)
self.to_k_lora = LoRALinearLayer(cross_attention_dim or hidden_size, hidden_size, rank, network_alpha)
self.to_v_lora = LoRALinearLayer(cross_attention_dim or hidden_size, hidden_size, rank, network_alpha)
self.to_out_lora = LoRALinearLayer(hidden_size, hidden_size, rank, network_alpha)
def __call__(
self,
attn,
hidden_states,
encoder_hidden_states=None,
attention_mask=None,
temb=None,
):
residual = hidden_states
if attn.spatial_norm is not None:
hidden_states = attn.spatial_norm(hidden_states, temb)
input_ndim = hidden_states.ndim
if input_ndim == 4:
batch_size, channel, height, width = hidden_states.shape
hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
batch_size, sequence_length, _ = (
hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
)
attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
if attn.group_norm is not None:
hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
query = attn.to_q(hidden_states) + self.lora_scale * self.to_q_lora(hidden_states)
if encoder_hidden_states is None:
encoder_hidden_states = hidden_states
elif attn.norm_cross:
encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
key = attn.to_k(encoder_hidden_states) + self.lora_scale * self.to_k_lora(encoder_hidden_states)
value = attn.to_v(encoder_hidden_states) + self.lora_scale * self.to_v_lora(encoder_hidden_states)
query = attn.head_to_batch_dim(query)
key = attn.head_to_batch_dim(key)
value = attn.head_to_batch_dim(value)
if is_xformers_available():
### xformers
hidden_states = xformers.ops.memory_efficient_attention(query, key, value, attn_bias=attention_mask)
hidden_states = hidden_states.to(query.dtype)
else:
attention_probs = attn.get_attention_scores(query, key, attention_mask)
hidden_states = torch.bmm(attention_probs, value)
hidden_states = attn.batch_to_head_dim(hidden_states)
# linear proj
hidden_states = attn.to_out[0](hidden_states) + self.lora_scale * self.to_out_lora(hidden_states)
# dropout
hidden_states = attn.to_out[1](hidden_states)
if input_ndim == 4:
hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
if attn.residual_connection:
hidden_states = hidden_states + residual
hidden_states = hidden_states / attn.rescale_output_factor
return hidden_states
class Consistent_IPAttProcessor(nn.Module):
def __init__(
self,
hidden_size,
cross_attention_dim=None,
rank=4,
network_alpha=None,
lora_scale=1.0,
scale=1.0,
num_tokens=4):
super().__init__()
self.rank = rank
self.lora_scale = lora_scale
self.num_tokens = num_tokens
self.to_q_lora = LoRALinearLayer(hidden_size, hidden_size, rank, network_alpha)
self.to_k_lora = LoRALinearLayer(cross_attention_dim or hidden_size, hidden_size, rank, network_alpha)
self.to_v_lora = LoRALinearLayer(cross_attention_dim or hidden_size, hidden_size, rank, network_alpha)
self.to_out_lora = LoRALinearLayer(hidden_size, hidden_size, rank, network_alpha)
self.hidden_size = hidden_size
self.cross_attention_dim = cross_attention_dim
self.scale = scale
self.to_k_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
self.to_v_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
def __call__(
self,
attn,
hidden_states,
encoder_hidden_states=None,
attention_mask=None,
scale=1.0,
temb=None,
):
residual = hidden_states
if attn.spatial_norm is not None:
hidden_states = attn.spatial_norm(hidden_states, temb)
input_ndim = hidden_states.ndim
if input_ndim == 4:
batch_size, channel, height, width = hidden_states.shape
hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
batch_size, sequence_length, _ = (
hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
)
attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
if attn.group_norm is not None:
hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
query = attn.to_q(hidden_states) + self.lora_scale * self.to_q_lora(hidden_states)
if encoder_hidden_states is None:
encoder_hidden_states = hidden_states
else:
end_pos = encoder_hidden_states.shape[1] - self.num_tokens
encoder_hidden_states, ip_hidden_states = (
encoder_hidden_states[:, :end_pos, :],
encoder_hidden_states[:, end_pos:, :],
)
if attn.norm_cross:
encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
key = attn.to_k(encoder_hidden_states) + self.lora_scale * self.to_k_lora(encoder_hidden_states)
value = attn.to_v(encoder_hidden_states) + self.lora_scale * self.to_v_lora(encoder_hidden_states)
inner_dim = key.shape[-1]
head_dim = inner_dim // attn.heads
query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
hidden_states = F.scaled_dot_product_attention(
query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
)
hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
hidden_states = hidden_states.to(query.dtype)
ip_key = self.to_k_ip(ip_hidden_states)
ip_value = self.to_v_ip(ip_hidden_states)
ip_key = ip_key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
ip_value = ip_value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
ip_hidden_states = F.scaled_dot_product_attention(
query, ip_key, ip_value, attn_mask=None, dropout_p=0.0, is_causal=False
)
ip_hidden_states = ip_hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
ip_hidden_states = ip_hidden_states.to(query.dtype)
hidden_states = hidden_states + self.scale * ip_hidden_states
# linear proj
hidden_states = attn.to_out[0](hidden_states) + self.lora_scale * self.to_out_lora(hidden_states)
# dropout
hidden_states = attn.to_out[1](hidden_states)
if input_ndim == 4:
hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
if attn.residual_connection:
hidden_states = hidden_states + residual
hidden_states = hidden_states / attn.rescale_output_factor
return hidden_states