Spaces:
Build error
Build error
Create cross_frame_attention.py
Browse files- cross_frame_attention.py +121 -0
cross_frame_attention.py
ADDED
@@ -0,0 +1,121 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Adapted from https://github.com/Picsart-AI-Research/Text2Video-Zero
|
2 |
+
import torch
|
3 |
+
from einops import rearrange
|
4 |
+
|
5 |
+
class CrossFrameAttnProcessor:
|
6 |
+
def __init__(self, unet_chunk_size=2):
|
7 |
+
self.unet_chunk_size = unet_chunk_size
|
8 |
+
|
9 |
+
def __call__(
|
10 |
+
self,
|
11 |
+
attn,
|
12 |
+
hidden_states,
|
13 |
+
encoder_hidden_states=None,
|
14 |
+
attention_mask=None, **kwargs):
|
15 |
+
batch_size, sequence_length, _ = hidden_states.shape
|
16 |
+
attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
|
17 |
+
query = attn.to_q(hidden_states)
|
18 |
+
|
19 |
+
is_cross_attention = encoder_hidden_states is not None
|
20 |
+
if encoder_hidden_states is None:
|
21 |
+
encoder_hidden_states = hidden_states
|
22 |
+
elif attn.norm_cross:
|
23 |
+
encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
|
24 |
+
key = attn.to_k(encoder_hidden_states)
|
25 |
+
value = attn.to_v(encoder_hidden_states)
|
26 |
+
# Sparse Attention
|
27 |
+
if not is_cross_attention:
|
28 |
+
video_length = key.size()[0] // self.unet_chunk_size
|
29 |
+
# print("Video length is", video_length)
|
30 |
+
# former_frame_index = torch.arange(video_length) - 1
|
31 |
+
# former_frame_index[0] = 0
|
32 |
+
former_frame_index = [0] * video_length
|
33 |
+
key = rearrange(key, "(b f) d c -> b f d c", f=video_length)
|
34 |
+
key = key[:, former_frame_index]
|
35 |
+
key = rearrange(key, "b f d c -> (b f) d c")
|
36 |
+
value = rearrange(value, "(b f) d c -> b f d c", f=video_length)
|
37 |
+
value = value[:, former_frame_index]
|
38 |
+
value = rearrange(value, "b f d c -> (b f) d c")
|
39 |
+
|
40 |
+
query = attn.head_to_batch_dim(query)
|
41 |
+
key = attn.head_to_batch_dim(key)
|
42 |
+
value = attn.head_to_batch_dim(value)
|
43 |
+
|
44 |
+
attention_probs = attn.get_attention_scores(query, key, attention_mask)
|
45 |
+
hidden_states = torch.bmm(attention_probs, value)
|
46 |
+
hidden_states = attn.batch_to_head_dim(hidden_states)
|
47 |
+
|
48 |
+
# linear proj
|
49 |
+
hidden_states = attn.to_out[0](hidden_states)
|
50 |
+
# dropout
|
51 |
+
hidden_states = attn.to_out[1](hidden_states)
|
52 |
+
|
53 |
+
return hidden_states
|
54 |
+
|
55 |
+
|
56 |
+
|
57 |
+
class AttnProcessorX:
|
58 |
+
r"""
|
59 |
+
Default processor for performing attention-related computations.
|
60 |
+
"""
|
61 |
+
|
62 |
+
def __call__(
|
63 |
+
self,
|
64 |
+
attn,
|
65 |
+
hidden_states,
|
66 |
+
encoder_hidden_states=None,
|
67 |
+
attention_mask=None,
|
68 |
+
temb=None,
|
69 |
+
scale=1.0,
|
70 |
+
):
|
71 |
+
residual = hidden_states
|
72 |
+
|
73 |
+
if attn.spatial_norm is not None:
|
74 |
+
hidden_states = attn.spatial_norm(hidden_states, temb)
|
75 |
+
|
76 |
+
input_ndim = hidden_states.ndim
|
77 |
+
|
78 |
+
if input_ndim == 4:
|
79 |
+
batch_size, channel, height, width = hidden_states.shape
|
80 |
+
hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
|
81 |
+
|
82 |
+
batch_size, sequence_length, _ = (
|
83 |
+
hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
|
84 |
+
)
|
85 |
+
attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
|
86 |
+
|
87 |
+
if attn.group_norm is not None:
|
88 |
+
hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
|
89 |
+
|
90 |
+
query = attn.to_q(hidden_states, scale=scale)
|
91 |
+
|
92 |
+
if encoder_hidden_states is None:
|
93 |
+
encoder_hidden_states = hidden_states
|
94 |
+
elif attn.norm_cross:
|
95 |
+
encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
|
96 |
+
|
97 |
+
key = attn.to_k(encoder_hidden_states, scale=scale)
|
98 |
+
value = attn.to_v(encoder_hidden_states, scale=scale)
|
99 |
+
|
100 |
+
query = attn.head_to_batch_dim(query)
|
101 |
+
key = attn.head_to_batch_dim(key)
|
102 |
+
value = attn.head_to_batch_dim(value)
|
103 |
+
|
104 |
+
attention_probs = attn.get_attention_scores(query, key, attention_mask)
|
105 |
+
hidden_states = torch.bmm(attention_probs, value)
|
106 |
+
hidden_states = attn.batch_to_head_dim(hidden_states)
|
107 |
+
|
108 |
+
# linear proj
|
109 |
+
hidden_states = attn.to_out[0](hidden_states, scale=scale)
|
110 |
+
# dropout
|
111 |
+
hidden_states = attn.to_out[1](hidden_states)
|
112 |
+
|
113 |
+
if input_ndim == 4:
|
114 |
+
hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
|
115 |
+
|
116 |
+
if attn.residual_connection:
|
117 |
+
hidden_states = hidden_states + residual
|
118 |
+
|
119 |
+
hidden_states = hidden_states / attn.rescale_output_factor
|
120 |
+
|
121 |
+
return hidden_states
|