File size: 8,692 Bytes
8d015d4
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import loguru
import torch
import math
import torch.nn as nn
import torch.nn.functional as F
from torch import einsum

from einops.layers.torch import Rearrange
from einops import rearrange

from .utils import coords_grid, bilinear_sampler, upflow8
from .attention import MultiHeadAttention, LinearPositionEmbeddingSine, ExpPositionEmbeddingSine
from typing import Optional, Tuple

from timm.models.layers import DropPath, to_2tuple, trunc_normal_

from .gru import BasicUpdateBlock, GMAUpdateBlock
from .gma import Attention

def initialize_flow(img):
    """ Flow is represented as difference between two means flow = mean1 - mean0"""
    N, C, H, W = img.shape
    mean = coords_grid(N, H, W).to(img.device)
    mean_init = coords_grid(N, H, W).to(img.device)

    # optical flow computed as difference: flow = mean1 - mean0
    return mean, mean_init

class CrossAttentionLayer(nn.Module):
    # def __init__(self, dim, cfg, num_heads=8, attn_drop=0., proj_drop=0., drop_path=0., dropout=0.):
    def __init__(self, qk_dim, v_dim, query_token_dim, tgt_token_dim, add_flow_token=True, num_heads=8, attn_drop=0., proj_drop=0., drop_path=0., dropout=0., pe='linear'):
        super(CrossAttentionLayer, self).__init__()

        head_dim = qk_dim // num_heads
        self.scale = head_dim ** -0.5
        self.query_token_dim = query_token_dim
        self.pe = pe

        self.norm1 = nn.LayerNorm(query_token_dim)
        self.norm2 = nn.LayerNorm(query_token_dim)
        self.multi_head_attn = MultiHeadAttention(qk_dim, num_heads)
        self.q, self.k, self.v = nn.Linear(query_token_dim, qk_dim, bias=True), nn.Linear(tgt_token_dim, qk_dim, bias=True), nn.Linear(tgt_token_dim, v_dim, bias=True)

        self.proj = nn.Linear(v_dim*2, query_token_dim)
        self.proj_drop = nn.Dropout(proj_drop)
        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()

        self.ffn = nn.Sequential(
            nn.Linear(query_token_dim, query_token_dim),
            nn.GELU(),
            nn.Dropout(dropout),
            nn.Linear(query_token_dim, query_token_dim),
            nn.Dropout(dropout)
        )
        self.add_flow_token = add_flow_token
        self.dim = qk_dim
    def forward(self, query, key, value, memory, query_coord, patch_size, size_h3w3):
        """
            query_coord [B, 2, H1, W1]
        """
        B, _, H1, W1 = query_coord.shape

        if key is None and value is None:
            key = self.k(memory)
            value = self.v(memory)

        # [B, 2, H1, W1] -> [BH1W1, 1, 2]
        query_coord = query_coord.contiguous()
        query_coord = query_coord.view(B, 2, -1).permute(0, 2, 1)[:,:,None,:].contiguous().view(B*H1*W1, 1, 2)
        if self.pe == 'linear':
            query_coord_enc = LinearPositionEmbeddingSine(query_coord, dim=self.dim)
        elif self.pe == 'exp':
            query_coord_enc = ExpPositionEmbeddingSine(query_coord, dim=self.dim)

        short_cut = query
        query = self.norm1(query)

        if self.add_flow_token:
            q = self.q(query+query_coord_enc)
        else:
            q = self.q(query_coord_enc)
        k, v = key, value

        x = self.multi_head_attn(q, k, v)

        x = self.proj(torch.cat([x, short_cut],dim=2))
        x = short_cut + self.proj_drop(x)

        x = x + self.drop_path(self.ffn(self.norm2(x)))

        return x, k, v

class MemoryDecoderLayer(nn.Module):
    def __init__(self, patch_size, query_latent_dim=64, cost_latent_dim=128):
        super(MemoryDecoderLayer, self).__init__()
        self.patch_size = patch_size # for converting coords into H2', W2' space
        self.query_latent_dim = query_latent_dim
        query_token_dim, tgt_token_dim = query_latent_dim, cost_latent_dim
        qk_dim, v_dim = query_token_dim, query_token_dim
        self.cross_attend = CrossAttentionLayer(qk_dim, v_dim, query_token_dim, tgt_token_dim, add_flow_token=True)

    def forward(self, query, key, value, memory, coords1, size, size_h3w3):
        """
            x:      [B*H1*W1, 1, C]
            memory: [B*H1*W1, H2'*W2', C]
            coords1 [B, 2, H2, W2]
            size: B, C, H1, W1
            1. Note that here coords0 and coords1 are in H2, W2 space.
               Should first convert it into H2', W2' space.
            2. We assume the upper-left point to be [0, 0], instead of letting center of upper-left patch to be [0, 0]
        """
        x_global, k, v = self.cross_attend(query, key, value, memory, coords1, self.patch_size, size_h3w3)
        B, C, H1, W1 = size
        C = self.query_latent_dim
        x_global = x_global.view(B, H1, W1, C).permute(0, 3, 1, 2)
        return x_global, k, v



class MemoryDecoder(nn.Module):
    def __init__(self, cfg):
        super(MemoryDecoder, self).__init__()
        dim = self.dim = cfg.query_latent_dim
        self.cfg = cfg

        self.flow_token_encoder = nn.Sequential(
            nn.Conv2d(81*cfg.cost_heads_num, dim, 1, 1),
            nn.GELU(),
            nn.Conv2d(dim, dim, 1, 1)
        )
        self.proj = nn.Conv2d(256, 256, 1)
        self.depth = cfg.decoder_depth
        self.decoder_layer = MemoryDecoderLayer(dim, cfg)
        
        if self.cfg.gma:
            self.update_block = GMAUpdateBlock(self.cfg, hidden_dim=128)
            self.att = Attention(args=self.cfg, dim=128, heads=1, max_pos_size=160, dim_head=128)
        else:
            self.update_block = BasicUpdateBlock(self.cfg, hidden_dim=128)
        
    def upsample_flow(self, flow, mask):
        """ Upsample flow field [H/8, W/8, 2] -> [H, W, 2] using convex combination """
        N, _, H, W = flow.shape
        mask = mask.view(N, 1, 9, 8, 8, H, W)
        mask = torch.softmax(mask, dim=2)

        up_flow = F.unfold(8 * flow, [3,3], padding=1)
        up_flow = up_flow.view(N, 2, 9, 1, 1, H, W)

        up_flow = torch.sum(mask * up_flow, dim=2)
        up_flow = up_flow.permute(0, 1, 4, 2, 5, 3)
        return up_flow.reshape(N, 2, 8*H, 8*W)

    def encode_flow_token(self, cost_maps, coords):
        """
            cost_maps   -   B*H1*W1, cost_heads_num, H2, W2
            coords      -   B, 2, H1, W1
        """
        coords = coords.permute(0, 2, 3, 1)
        batch, h1, w1, _ = coords.shape

        r = 4
        dx = torch.linspace(-r, r, 2*r+1)
        dy = torch.linspace(-r, r, 2*r+1)
        delta = torch.stack(torch.meshgrid(dy, dx), axis=-1).to(coords.device)

        centroid = coords.reshape(batch*h1*w1, 1, 1, 2)
        delta = delta.view(1, 2*r+1, 2*r+1, 2)
        coords = centroid + delta
        corr = bilinear_sampler(cost_maps, coords)
        corr = corr.view(batch, h1, w1, -1).permute(0, 3, 1, 2)
        return corr

    def forward(self, cost_memory, context, data={}, flow_init=None):
        """
            memory: [B*H1*W1, H2'*W2', C]
            context: [B, D, H1, W1]
        """
        cost_maps = data['cost_maps']
        coords0, coords1 = initialize_flow(context)

        if flow_init is not None:
            #print("[Using warm start]")
            coords1 = coords1 + flow_init

        #flow = coords1

        flow_predictions = []

        context = self.proj(context)
        net, inp = torch.split(context, [128, 128], dim=1)
        net = torch.tanh(net)
        inp = torch.relu(inp)
        if self.cfg.gma:
            attention = self.att(inp)

        size = net.shape
        key, value = None, None

        for idx in range(self.depth):
            coords1 = coords1.detach()

            cost_forward = self.encode_flow_token(cost_maps, coords1)
            #cost_backward = self.reverse_cost_extractor(cost_maps, coords0, coords1)

            query = self.flow_token_encoder(cost_forward)
            query = query.permute(0, 2, 3, 1).contiguous().view(size[0]*size[2]*size[3], 1, self.dim)
            cost_global, key, value = self.decoder_layer(query, key, value, cost_memory, coords1, size, data['H3W3'])
            if self.cfg.only_global:
                corr = cost_global
            else:
                corr = torch.cat([cost_global, cost_forward], dim=1)

            flow = coords1 - coords0
             
            if self.cfg.gma:
                net, up_mask, delta_flow = self.update_block(net, inp, corr, flow, attention)
            else:
                net, up_mask, delta_flow = self.update_block(net, inp, corr, flow)

            # flow = delta_flow
            coords1 = coords1 + delta_flow
            flow_up = self.upsample_flow(coords1 - coords0, up_mask)
            flow_predictions.append(flow_up)
        
        if self.training:
            return flow_predictions
        else:
            return flow_predictions[-1:]