File size: 7,290 Bytes
684943d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import numpy as np
import torch
import cv2
from torchvision.transforms import Resize, InterpolationMode, ToTensor, Compose, CenterCrop
from einops import rearrange
import glob
from diffusers.utils import USE_PEFT_BACKEND
from diffusers.utils import load_image
from natsort import natsorted


def read_mask(mask_dir):
    transform = Compose([
        Resize((512, 512), interpolation=InterpolationMode.BILINEAR, antialias=True),
        # CenterCrop((512, 512)),
        ToTensor()])
    mask_paths = glob.glob(mask_dir + '/*.png')
    mask_paths = natsorted(mask_paths)
    mask_list = []
    for mask_path in mask_paths:
        mask = load_image(mask_path)
        mask_torch = transform(mask).bool().unsqueeze(0)  # torch.Size([1, 3, 512, 512]) -1~1
        mask_list.append(mask_torch)
    return mask_list


def read_rgb(rgb_dir):
    transform = Compose([
        Resize((512, 512), interpolation=InterpolationMode.BILINEAR, antialias=True),
        # CenterCrop((512, 512)),
        ToTensor()])
    rgb_paths = sorted(glob.glob(rgb_dir + '/*.jpg'))
    rgb_list = []
    rgb_frame = []
    for rgb_path in rgb_paths:
        rgb = load_image(rgb_path);
        width, height = rgb.size
        file_name = rgb_path.split('/')[-1]
        frame_number = int(file_name.split('_')[1].split('.')[0].lstrip('0') or '0')
        rgb_frame.append(frame_number)
        rgb_torch = transform(rgb).unsqueeze(0)  # torch.Size([1, 3, 512, 512])
        rgb_list.append(rgb_torch)
    return rgb_list, (width, height), rgb_frame


def read_depth2disparity(depth_dir):
    depth_paths = sorted(glob.glob(depth_dir + '/*.npy'))
    disparity_list = []
    for depth_path in depth_paths:
        depth = np.load(depth_path)
        depth = cv2.resize(depth, (512, 512)).reshape((512, 512, 1))  # [512,512,1]
        # depth = CenterCrop((512, 512))(torch.from_numpy(depth))[..., None].numpy() # [512,512,1]

        disparity = 1 / (depth + 1e-5)
        disparity_map = disparity / np.max(disparity)  # 0.00233~1
        # disparity_map = disparity_map.astype(np.uint8)[:,:,0]
        disparity_map = np.concatenate([disparity_map, disparity_map, disparity_map], axis=2)
        disparity_list.append(torch.from_numpy(disparity_map[None]).permute(0, 3, 1, 2).float())  # [1,512,512,3]
    return disparity_list


def compute_attn(attn, query, key, value, video_length, ref_frame_index, attention_mask):
    key_ref_cross = rearrange(key, "(b f) d c -> b f d c", f=video_length)
    key_ref_cross = key_ref_cross[:, ref_frame_index]
    key_ref_cross = rearrange(key_ref_cross, "b f d c -> (b f) d c")
    value_ref_cross = rearrange(value, "(b f) d c -> b f d c", f=video_length)
    value_ref_cross = value_ref_cross[:, ref_frame_index]
    value_ref_cross = rearrange(value_ref_cross, "b f d c -> (b f) d c")

    key_ref_cross = attn.head_to_batch_dim(key_ref_cross)
    value_ref_cross = attn.head_to_batch_dim(value_ref_cross)
    attention_probs = attn.get_attention_scores(query, key_ref_cross, attention_mask)
    hidden_states_ref_cross = torch.bmm(attention_probs, value_ref_cross)
    return hidden_states_ref_cross


class CrossViewAttnProcessor:
    def __init__(self, self_attn_coeff, unet_chunk_size=2):
        self.unet_chunk_size = unet_chunk_size
        self.self_attn_coeff = self_attn_coeff

    def __call__(
            self,
            attn,
            hidden_states,
            encoder_hidden_states=None,
            attention_mask=None,
            temb=None,
            scale=1.0, ):

        residual = hidden_states

        args = () if USE_PEFT_BACKEND else (scale,)

        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, *args)

        is_cross_attention = encoder_hidden_states is not None
        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, *args)
        value = attn.to_v(encoder_hidden_states, *args)
        query = attn.head_to_batch_dim(query)
        # Sparse Attention
        if not is_cross_attention:
            ################## Perform self attention
            key_self = attn.head_to_batch_dim(key)
            value_self = attn.head_to_batch_dim(value)
            attention_probs = attn.get_attention_scores(query, key_self, attention_mask)
            hidden_states_self = torch.bmm(attention_probs, value_self)
            #######################################

            video_length = key.size()[0] // self.unet_chunk_size
            ref0_frame_index = [0] * video_length
            ref1_frame_index = [1] * video_length
            ref2_frame_index = [2] * video_length
            ref3_frame_index = [3] * video_length

            hidden_states_ref0 = compute_attn(attn, query, key, value, video_length, ref0_frame_index, attention_mask)
            hidden_states_ref1 = compute_attn(attn, query, key, value, video_length, ref1_frame_index, attention_mask)
            hidden_states_ref2 = compute_attn(attn, query, key, value, video_length, ref2_frame_index, attention_mask)

            key = rearrange(key, "(b f) d c -> b f d c", f=video_length)
            key = key[:, ref3_frame_index]
            key = rearrange(key, "b f d c -> (b f) d c")
            value = rearrange(value, "(b f) d c -> b f d c", f=video_length)
            value = value[:, ref3_frame_index]
            value = rearrange(value, "b f d c -> (b f) d c")

        key = attn.head_to_batch_dim(key)
        value = attn.head_to_batch_dim(value)

        attention_probs = attn.get_attention_scores(query, key, attention_mask)
        hidden_states_ref3 = torch.bmm(attention_probs, value)

        hidden_states = self.self_attn_coeff * hidden_states_self + (1 - self.self_attn_coeff) * torch.mean(
            torch.stack([hidden_states_ref0, hidden_states_ref1, hidden_states_ref2, hidden_states_ref3]),
            dim=0) if not is_cross_attention else hidden_states_ref3
        hidden_states = attn.batch_to_head_dim(hidden_states)

        # linear proj
        hidden_states = attn.to_out[0](hidden_states, *args)
        # 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