Spaces:
Runtime error
Runtime error
Update ip_adapter/attention_processor.py
Browse files- ip_adapter/attention_processor.py +354 -57
ip_adapter/attention_processor.py
CHANGED
@@ -1,3 +1,4 @@
|
|
|
|
1 |
import torch
|
2 |
import torch.nn as nn
|
3 |
import torch.nn.functional as F
|
@@ -6,31 +7,33 @@ try:
|
|
6 |
import xformers
|
7 |
import xformers.ops
|
8 |
xformers_available = True
|
9 |
-
except Exception:
|
10 |
xformers_available = False
|
11 |
|
12 |
-
|
13 |
-
# Region Controller (unchanged)
|
14 |
-
class RegionControler:
|
15 |
def __init__(self) -> None:
|
16 |
self.prompt_image_conditioning = []
|
17 |
-
|
18 |
region_control = RegionControler()
|
19 |
|
20 |
-
|
21 |
-
|
22 |
-
|
23 |
-
|
24 |
-
|
25 |
-
|
26 |
-
|
27 |
-
|
28 |
-
|
29 |
-
def __init__(self):
|
30 |
super().__init__()
|
31 |
|
32 |
-
def
|
33 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
34 |
residual = hidden_states
|
35 |
|
36 |
if attn.spatial_norm is not None:
|
@@ -41,60 +44,286 @@ class BaseAttnProcessor(nn.Module):
|
|
41 |
if input_ndim == 4:
|
42 |
batch_size, channel, height, width = hidden_states.shape
|
43 |
hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
|
44 |
-
|
45 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
46 |
|
47 |
if encoder_hidden_states is None:
|
48 |
encoder_hidden_states = hidden_states
|
49 |
elif attn.norm_cross:
|
50 |
encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
|
51 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
52 |
attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
|
53 |
|
54 |
if attn.group_norm is not None:
|
55 |
hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
|
56 |
|
57 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
58 |
|
59 |
-
def _apply_attention(self, attn, query, key, value, attention_mask):
|
60 |
-
"""Handles the actual attention operation using either xformers or standard PyTorch"""
|
61 |
if xformers_available:
|
62 |
-
|
63 |
else:
|
64 |
attention_probs = attn.get_attention_scores(query, key, attention_mask)
|
65 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
66 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
67 |
|
68 |
-
|
69 |
-
|
70 |
-
|
71 |
-
|
72 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
73 |
|
74 |
query = attn.to_q(hidden_states)
|
|
|
|
|
|
|
|
|
|
|
|
|
75 |
key = attn.to_k(encoder_hidden_states)
|
76 |
value = attn.to_v(encoder_hidden_states)
|
77 |
|
78 |
-
|
79 |
-
|
80 |
-
hidden_states = attn.batch_to_head_dim(hidden_states)
|
81 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
82 |
hidden_states = attn.to_out[0](hidden_states)
|
|
|
83 |
hidden_states = attn.to_out[1](hidden_states)
|
84 |
|
85 |
if input_ndim == 4:
|
86 |
-
hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size,
|
87 |
|
88 |
if attn.residual_connection:
|
89 |
hidden_states = hidden_states + residual
|
90 |
|
91 |
-
|
|
|
|
|
92 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
93 |
|
94 |
-
# Optimized IPAttnProcessor
|
95 |
-
class IPAttnProcessor(BaseAttnProcessor):
|
96 |
def __init__(self, hidden_size, cross_attention_dim=None, scale=1.0, num_tokens=4):
|
97 |
super().__init__()
|
|
|
|
|
|
|
|
|
98 |
self.hidden_size = hidden_size
|
99 |
self.cross_attention_dim = cross_attention_dim
|
100 |
self.scale = scale
|
@@ -102,49 +331,117 @@ class IPAttnProcessor(BaseAttnProcessor):
|
|
102 |
|
103 |
self.to_k_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
|
104 |
self.to_v_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
|
105 |
-
self.apply(init_weights)
|
106 |
|
107 |
-
def forward(
|
108 |
-
|
109 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
110 |
|
111 |
-
|
112 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
113 |
|
114 |
query = attn.to_q(hidden_states)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
115 |
key = attn.to_k(encoder_hidden_states)
|
116 |
value = attn.to_v(encoder_hidden_states)
|
117 |
|
118 |
-
|
119 |
-
|
120 |
-
hidden_states = attn.batch_to_head_dim(hidden_states)
|
121 |
|
122 |
-
|
123 |
-
ip_key = attn.head_to_batch_dim(self.to_k_ip(ip_hidden_states))
|
124 |
-
ip_value = attn.head_to_batch_dim(self.to_v_ip(ip_hidden_states))
|
125 |
|
126 |
-
|
127 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
128 |
|
129 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
130 |
if len(region_control.prompt_image_conditioning) == 1:
|
131 |
-
region_mask = region_control.prompt_image_conditioning[0].get(
|
132 |
if region_mask is not None:
|
133 |
-
|
|
|
|
|
|
|
134 |
else:
|
135 |
mask = torch.ones_like(ip_hidden_states)
|
136 |
-
ip_hidden_states
|
137 |
|
138 |
hidden_states = hidden_states + self.scale * ip_hidden_states
|
139 |
|
140 |
-
#
|
141 |
hidden_states = attn.to_out[0](hidden_states)
|
|
|
142 |
hidden_states = attn.to_out[1](hidden_states)
|
143 |
|
144 |
if input_ndim == 4:
|
145 |
-
hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size,
|
146 |
|
147 |
if attn.residual_connection:
|
148 |
hidden_states = hidden_states + residual
|
149 |
|
150 |
-
|
|
|
|
|
|
1 |
+
# modified from https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py
|
2 |
import torch
|
3 |
import torch.nn as nn
|
4 |
import torch.nn.functional as F
|
|
|
7 |
import xformers
|
8 |
import xformers.ops
|
9 |
xformers_available = True
|
10 |
+
except Exception as e:
|
11 |
xformers_available = False
|
12 |
|
13 |
+
class RegionControler(object):
|
|
|
|
|
14 |
def __init__(self) -> None:
|
15 |
self.prompt_image_conditioning = []
|
|
|
16 |
region_control = RegionControler()
|
17 |
|
18 |
+
class AttnProcessor(nn.Module):
|
19 |
+
r"""
|
20 |
+
Default processor for performing attention-related computations.
|
21 |
+
"""
|
22 |
+
def __init__(
|
23 |
+
self,
|
24 |
+
hidden_size=None,
|
25 |
+
cross_attention_dim=None,
|
26 |
+
):
|
|
|
27 |
super().__init__()
|
28 |
|
29 |
+
def forward(
|
30 |
+
self,
|
31 |
+
attn,
|
32 |
+
hidden_states,
|
33 |
+
encoder_hidden_states=None,
|
34 |
+
attention_mask=None,
|
35 |
+
temb=None,
|
36 |
+
):
|
37 |
residual = hidden_states
|
38 |
|
39 |
if attn.spatial_norm is not None:
|
|
|
44 |
if input_ndim == 4:
|
45 |
batch_size, channel, height, width = hidden_states.shape
|
46 |
hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
|
47 |
+
|
48 |
+
batch_size, sequence_length, _ = (
|
49 |
+
hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
|
50 |
+
)
|
51 |
+
attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
|
52 |
+
|
53 |
+
if attn.group_norm is not None:
|
54 |
+
hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
|
55 |
+
|
56 |
+
query = attn.to_q(hidden_states)
|
57 |
|
58 |
if encoder_hidden_states is None:
|
59 |
encoder_hidden_states = hidden_states
|
60 |
elif attn.norm_cross:
|
61 |
encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
|
62 |
|
63 |
+
key = attn.to_k(encoder_hidden_states)
|
64 |
+
value = attn.to_v(encoder_hidden_states)
|
65 |
+
|
66 |
+
query = attn.head_to_batch_dim(query)
|
67 |
+
key = attn.head_to_batch_dim(key)
|
68 |
+
value = attn.head_to_batch_dim(value)
|
69 |
+
|
70 |
+
attention_probs = attn.get_attention_scores(query, key, attention_mask)
|
71 |
+
hidden_states = torch.bmm(attention_probs, value)
|
72 |
+
hidden_states = attn.batch_to_head_dim(hidden_states)
|
73 |
+
|
74 |
+
# linear proj
|
75 |
+
hidden_states = attn.to_out[0](hidden_states)
|
76 |
+
# dropout
|
77 |
+
hidden_states = attn.to_out[1](hidden_states)
|
78 |
+
|
79 |
+
if input_ndim == 4:
|
80 |
+
hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
|
81 |
+
|
82 |
+
if attn.residual_connection:
|
83 |
+
hidden_states = hidden_states + residual
|
84 |
+
|
85 |
+
hidden_states = hidden_states / attn.rescale_output_factor
|
86 |
+
|
87 |
+
return hidden_states
|
88 |
+
|
89 |
+
|
90 |
+
class IPAttnProcessor(nn.Module):
|
91 |
+
r"""
|
92 |
+
Attention processor for IP-Adapater.
|
93 |
+
Args:
|
94 |
+
hidden_size (`int`):
|
95 |
+
The hidden size of the attention layer.
|
96 |
+
cross_attention_dim (`int`):
|
97 |
+
The number of channels in the `encoder_hidden_states`.
|
98 |
+
scale (`float`, defaults to 1.0):
|
99 |
+
the weight scale of image prompt.
|
100 |
+
num_tokens (`int`, defaults to 4 when do ip_adapter_plus it should be 16):
|
101 |
+
The context length of the image features.
|
102 |
+
"""
|
103 |
+
|
104 |
+
def __init__(self, hidden_size, cross_attention_dim=None, scale=1.0, num_tokens=4):
|
105 |
+
super().__init__()
|
106 |
+
|
107 |
+
self.hidden_size = hidden_size
|
108 |
+
self.cross_attention_dim = cross_attention_dim
|
109 |
+
self.scale = scale
|
110 |
+
self.num_tokens = num_tokens
|
111 |
+
|
112 |
+
self.to_k_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
|
113 |
+
self.to_v_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
|
114 |
+
|
115 |
+
def forward(
|
116 |
+
self,
|
117 |
+
attn,
|
118 |
+
hidden_states,
|
119 |
+
encoder_hidden_states=None,
|
120 |
+
attention_mask=None,
|
121 |
+
temb=None,
|
122 |
+
):
|
123 |
+
residual = hidden_states
|
124 |
+
|
125 |
+
if attn.spatial_norm is not None:
|
126 |
+
hidden_states = attn.spatial_norm(hidden_states, temb)
|
127 |
+
|
128 |
+
input_ndim = hidden_states.ndim
|
129 |
+
|
130 |
+
if input_ndim == 4:
|
131 |
+
batch_size, channel, height, width = hidden_states.shape
|
132 |
+
hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
|
133 |
+
|
134 |
+
batch_size, sequence_length, _ = (
|
135 |
+
hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
|
136 |
+
)
|
137 |
attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
|
138 |
|
139 |
if attn.group_norm is not None:
|
140 |
hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
|
141 |
|
142 |
+
query = attn.to_q(hidden_states)
|
143 |
+
|
144 |
+
if encoder_hidden_states is None:
|
145 |
+
encoder_hidden_states = hidden_states
|
146 |
+
else:
|
147 |
+
# get encoder_hidden_states, ip_hidden_states
|
148 |
+
end_pos = encoder_hidden_states.shape[1] - self.num_tokens
|
149 |
+
encoder_hidden_states, ip_hidden_states = encoder_hidden_states[:, :end_pos, :], encoder_hidden_states[:, end_pos:, :]
|
150 |
+
if attn.norm_cross:
|
151 |
+
encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
|
152 |
+
|
153 |
+
key = attn.to_k(encoder_hidden_states)
|
154 |
+
value = attn.to_v(encoder_hidden_states)
|
155 |
+
|
156 |
+
query = attn.head_to_batch_dim(query)
|
157 |
+
key = attn.head_to_batch_dim(key)
|
158 |
+
value = attn.head_to_batch_dim(value)
|
159 |
|
|
|
|
|
160 |
if xformers_available:
|
161 |
+
hidden_states = self._memory_efficient_attention_xformers(query, key, value, attention_mask)
|
162 |
else:
|
163 |
attention_probs = attn.get_attention_scores(query, key, attention_mask)
|
164 |
+
hidden_states = torch.bmm(attention_probs, value)
|
165 |
+
hidden_states = attn.batch_to_head_dim(hidden_states)
|
166 |
+
|
167 |
+
# for ip-adapter
|
168 |
+
ip_key = self.to_k_ip(ip_hidden_states)
|
169 |
+
ip_value = self.to_v_ip(ip_hidden_states)
|
170 |
+
|
171 |
+
ip_key = attn.head_to_batch_dim(ip_key)
|
172 |
+
ip_value = attn.head_to_batch_dim(ip_value)
|
173 |
+
|
174 |
+
if xformers_available:
|
175 |
+
ip_hidden_states = self._memory_efficient_attention_xformers(query, ip_key, ip_value, None)
|
176 |
+
else:
|
177 |
+
ip_attention_probs = attn.get_attention_scores(query, ip_key, None)
|
178 |
+
ip_hidden_states = torch.bmm(ip_attention_probs, ip_value)
|
179 |
+
ip_hidden_states = attn.batch_to_head_dim(ip_hidden_states)
|
180 |
+
|
181 |
+
# region control
|
182 |
+
if len(region_control.prompt_image_conditioning) == 1:
|
183 |
+
region_mask = region_control.prompt_image_conditioning[0].get('region_mask', None)
|
184 |
+
if region_mask is not None:
|
185 |
+
h, w = region_mask.shape[:2]
|
186 |
+
ratio = (h * w / query.shape[1]) ** 0.5
|
187 |
+
mask = F.interpolate(region_mask[None, None], scale_factor=1/ratio, mode='nearest').reshape([1, -1, 1])
|
188 |
+
else:
|
189 |
+
mask = torch.ones_like(ip_hidden_states)
|
190 |
+
ip_hidden_states = ip_hidden_states * mask
|
191 |
+
|
192 |
+
hidden_states = hidden_states + self.scale * ip_hidden_states
|
193 |
|
194 |
+
# linear proj
|
195 |
+
hidden_states = attn.to_out[0](hidden_states)
|
196 |
+
# dropout
|
197 |
+
hidden_states = attn.to_out[1](hidden_states)
|
198 |
+
|
199 |
+
if input_ndim == 4:
|
200 |
+
hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
|
201 |
+
|
202 |
+
if attn.residual_connection:
|
203 |
+
hidden_states = hidden_states + residual
|
204 |
+
|
205 |
+
hidden_states = hidden_states / attn.rescale_output_factor
|
206 |
+
|
207 |
+
return hidden_states
|
208 |
|
209 |
+
|
210 |
+
def _memory_efficient_attention_xformers(self, query, key, value, attention_mask):
|
211 |
+
# TODO attention_mask
|
212 |
+
query = query.contiguous()
|
213 |
+
key = key.contiguous()
|
214 |
+
value = value.contiguous()
|
215 |
+
hidden_states = xformers.ops.memory_efficient_attention(query, key, value, attn_bias=attention_mask)
|
216 |
+
# hidden_states = self.reshape_batch_dim_to_heads(hidden_states)
|
217 |
+
return hidden_states
|
218 |
+
|
219 |
+
|
220 |
+
class AttnProcessor2_0(torch.nn.Module):
|
221 |
+
r"""
|
222 |
+
Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0).
|
223 |
+
"""
|
224 |
+
def __init__(
|
225 |
+
self,
|
226 |
+
hidden_size=None,
|
227 |
+
cross_attention_dim=None,
|
228 |
+
):
|
229 |
+
super().__init__()
|
230 |
+
if not hasattr(F, "scaled_dot_product_attention"):
|
231 |
+
raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
|
232 |
+
|
233 |
+
def forward(
|
234 |
+
self,
|
235 |
+
attn,
|
236 |
+
hidden_states,
|
237 |
+
encoder_hidden_states=None,
|
238 |
+
attention_mask=None,
|
239 |
+
temb=None,
|
240 |
+
):
|
241 |
+
residual = hidden_states
|
242 |
+
|
243 |
+
if attn.spatial_norm is not None:
|
244 |
+
hidden_states = attn.spatial_norm(hidden_states, temb)
|
245 |
+
|
246 |
+
input_ndim = hidden_states.ndim
|
247 |
+
|
248 |
+
if input_ndim == 4:
|
249 |
+
batch_size, channel, height, width = hidden_states.shape
|
250 |
+
hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
|
251 |
+
|
252 |
+
batch_size, sequence_length, _ = (
|
253 |
+
hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
|
254 |
+
)
|
255 |
+
|
256 |
+
if attention_mask is not None:
|
257 |
+
attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
|
258 |
+
# scaled_dot_product_attention expects attention_mask shape to be
|
259 |
+
# (batch, heads, source_length, target_length)
|
260 |
+
attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
|
261 |
+
|
262 |
+
if attn.group_norm is not None:
|
263 |
+
hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
|
264 |
|
265 |
query = attn.to_q(hidden_states)
|
266 |
+
|
267 |
+
if encoder_hidden_states is None:
|
268 |
+
encoder_hidden_states = hidden_states
|
269 |
+
elif attn.norm_cross:
|
270 |
+
encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
|
271 |
+
|
272 |
key = attn.to_k(encoder_hidden_states)
|
273 |
value = attn.to_v(encoder_hidden_states)
|
274 |
|
275 |
+
inner_dim = key.shape[-1]
|
276 |
+
head_dim = inner_dim // attn.heads
|
|
|
277 |
|
278 |
+
query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
|
279 |
+
|
280 |
+
key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
|
281 |
+
value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
|
282 |
+
|
283 |
+
# the output of sdp = (batch, num_heads, seq_len, head_dim)
|
284 |
+
# TODO: add support for attn.scale when we move to Torch 2.1
|
285 |
+
hidden_states = F.scaled_dot_product_attention(
|
286 |
+
query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
|
287 |
+
)
|
288 |
+
|
289 |
+
hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
|
290 |
+
hidden_states = hidden_states.to(query.dtype)
|
291 |
+
|
292 |
+
# linear proj
|
293 |
hidden_states = attn.to_out[0](hidden_states)
|
294 |
+
# dropout
|
295 |
hidden_states = attn.to_out[1](hidden_states)
|
296 |
|
297 |
if input_ndim == 4:
|
298 |
+
hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
|
299 |
|
300 |
if attn.residual_connection:
|
301 |
hidden_states = hidden_states + residual
|
302 |
|
303 |
+
hidden_states = hidden_states / attn.rescale_output_factor
|
304 |
+
|
305 |
+
return hidden_states
|
306 |
|
307 |
+
class IPAttnProcessor2_0(torch.nn.Module):
|
308 |
+
r"""
|
309 |
+
Attention processor for IP-Adapater for PyTorch 2.0.
|
310 |
+
Args:
|
311 |
+
hidden_size (`int`):
|
312 |
+
The hidden size of the attention layer.
|
313 |
+
cross_attention_dim (`int`):
|
314 |
+
The number of channels in the `encoder_hidden_states`.
|
315 |
+
scale (`float`, defaults to 1.0):
|
316 |
+
the weight scale of image prompt.
|
317 |
+
num_tokens (`int`, defaults to 4 when do ip_adapter_plus it should be 16):
|
318 |
+
The context length of the image features.
|
319 |
+
"""
|
320 |
|
|
|
|
|
321 |
def __init__(self, hidden_size, cross_attention_dim=None, scale=1.0, num_tokens=4):
|
322 |
super().__init__()
|
323 |
+
|
324 |
+
if not hasattr(F, "scaled_dot_product_attention"):
|
325 |
+
raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
|
326 |
+
|
327 |
self.hidden_size = hidden_size
|
328 |
self.cross_attention_dim = cross_attention_dim
|
329 |
self.scale = scale
|
|
|
331 |
|
332 |
self.to_k_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
|
333 |
self.to_v_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
|
|
|
334 |
|
335 |
+
def forward(
|
336 |
+
self,
|
337 |
+
attn,
|
338 |
+
hidden_states,
|
339 |
+
encoder_hidden_states=None,
|
340 |
+
attention_mask=None,
|
341 |
+
temb=None,
|
342 |
+
):
|
343 |
+
residual = hidden_states
|
344 |
+
|
345 |
+
if attn.spatial_norm is not None:
|
346 |
+
hidden_states = attn.spatial_norm(hidden_states, temb)
|
347 |
+
|
348 |
+
input_ndim = hidden_states.ndim
|
349 |
+
|
350 |
+
if input_ndim == 4:
|
351 |
+
batch_size, channel, height, width = hidden_states.shape
|
352 |
+
hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
|
353 |
+
|
354 |
+
batch_size, sequence_length, _ = (
|
355 |
+
hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
|
356 |
+
)
|
357 |
|
358 |
+
if attention_mask is not None:
|
359 |
+
attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
|
360 |
+
# scaled_dot_product_attention expects attention_mask shape to be
|
361 |
+
# (batch, heads, source_length, target_length)
|
362 |
+
attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
|
363 |
+
|
364 |
+
if attn.group_norm is not None:
|
365 |
+
hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
|
366 |
|
367 |
query = attn.to_q(hidden_states)
|
368 |
+
|
369 |
+
if encoder_hidden_states is None:
|
370 |
+
encoder_hidden_states = hidden_states
|
371 |
+
else:
|
372 |
+
# get encoder_hidden_states, ip_hidden_states
|
373 |
+
end_pos = encoder_hidden_states.shape[1] - self.num_tokens
|
374 |
+
encoder_hidden_states, ip_hidden_states = (
|
375 |
+
encoder_hidden_states[:, :end_pos, :],
|
376 |
+
encoder_hidden_states[:, end_pos:, :],
|
377 |
+
)
|
378 |
+
if attn.norm_cross:
|
379 |
+
encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
|
380 |
+
|
381 |
key = attn.to_k(encoder_hidden_states)
|
382 |
value = attn.to_v(encoder_hidden_states)
|
383 |
|
384 |
+
inner_dim = key.shape[-1]
|
385 |
+
head_dim = inner_dim // attn.heads
|
|
|
386 |
|
387 |
+
query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
|
|
|
|
|
388 |
|
389 |
+
key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
|
390 |
+
value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
|
391 |
+
|
392 |
+
# the output of sdp = (batch, num_heads, seq_len, head_dim)
|
393 |
+
# TODO: add support for attn.scale when we move to Torch 2.1
|
394 |
+
hidden_states = F.scaled_dot_product_attention(
|
395 |
+
query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
|
396 |
+
)
|
397 |
+
|
398 |
+
hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
|
399 |
+
hidden_states = hidden_states.to(query.dtype)
|
400 |
+
|
401 |
+
# for ip-adapter
|
402 |
+
ip_key = self.to_k_ip(ip_hidden_states)
|
403 |
+
ip_value = self.to_v_ip(ip_hidden_states)
|
404 |
|
405 |
+
ip_key = ip_key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
|
406 |
+
ip_value = ip_value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
|
407 |
+
|
408 |
+
# the output of sdp = (batch, num_heads, seq_len, head_dim)
|
409 |
+
# TODO: add support for attn.scale when we move to Torch 2.1
|
410 |
+
ip_hidden_states = F.scaled_dot_product_attention(
|
411 |
+
query, ip_key, ip_value, attn_mask=None, dropout_p=0.0, is_causal=False
|
412 |
+
)
|
413 |
+
with torch.no_grad():
|
414 |
+
self.attn_map = query @ ip_key.transpose(-2, -1).softmax(dim=-1)
|
415 |
+
#print(self.attn_map.shape)
|
416 |
+
|
417 |
+
ip_hidden_states = ip_hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
|
418 |
+
ip_hidden_states = ip_hidden_states.to(query.dtype)
|
419 |
+
|
420 |
+
# region control
|
421 |
if len(region_control.prompt_image_conditioning) == 1:
|
422 |
+
region_mask = region_control.prompt_image_conditioning[0].get('region_mask', None)
|
423 |
if region_mask is not None:
|
424 |
+
query = query.reshape([-1, query.shape[-2], query.shape[-1]])
|
425 |
+
h, w = region_mask.shape[:2]
|
426 |
+
ratio = (h * w / query.shape[1]) ** 0.5
|
427 |
+
mask = F.interpolate(region_mask[None, None], scale_factor=1/ratio, mode='nearest').reshape([1, -1, 1])
|
428 |
else:
|
429 |
mask = torch.ones_like(ip_hidden_states)
|
430 |
+
ip_hidden_states = ip_hidden_states * mask
|
431 |
|
432 |
hidden_states = hidden_states + self.scale * ip_hidden_states
|
433 |
|
434 |
+
# linear proj
|
435 |
hidden_states = attn.to_out[0](hidden_states)
|
436 |
+
# dropout
|
437 |
hidden_states = attn.to_out[1](hidden_states)
|
438 |
|
439 |
if input_ndim == 4:
|
440 |
+
hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
|
441 |
|
442 |
if attn.residual_connection:
|
443 |
hidden_states = hidden_states + residual
|
444 |
|
445 |
+
hidden_states = hidden_states / attn.rescale_output_factor
|
446 |
+
|
447 |
+
return hidden_states
|