File size: 17,513 Bytes
d12790b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
import torch
import torch.nn as nn
import torch.nn.functional as F
from einops import rearrange
import pdb



def conv3x3(in_planes, out_planes, stride=1):
    return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride, padding=1, bias=False)
def conv1x1(in_planes, out_planes, stride=1):
    return nn.Conv2d(in_planes, out_planes, kernel_size=1, stride=stride, padding=0, bias=False)

class depthwise_separable_conv(nn.Module):
    def __init__(self, in_ch, out_ch, stride=1, kernel_size=3, padding=1, bias=False):
        super().__init__()
        self.depthwise = nn.Conv2d(in_ch, in_ch, kernel_size=kernel_size, padding=padding, groups=in_ch, bias=bias, stride=stride)
        self.pointwise = nn.Conv2d(in_ch, out_ch, kernel_size=1, bias=bias)

    def forward(self, x): 
        out = self.depthwise(x)
        out = self.pointwise(out)

        return out 

class Mlp(nn.Module):
    def __init__(self, in_ch, hid_ch=None, out_ch=None, act_layer=nn.GELU, drop=0.):
        super().__init__()
        out_ch = out_ch or in_ch
        hid_ch = hid_ch or in_ch

        self.fc1 = nn.Conv2d(in_ch, hid_ch, kernel_size=1)
        self.act = act_layer()
        self.fc2 = nn.Conv2d(hid_ch, out_ch, kernel_size=1)
        self.drop = nn.Dropout(drop)

    def forward(self, x):
        x = self.fc1(x)
        x = self.act(x)
        x = self.drop(x)
        x = self.fc2(x)
        x = self.drop(x)

        return x

class BasicBlock(nn.Module):

    def __init__(self, inplanes, planes, stride=1):
        super().__init__()
        self.conv1 = conv3x3(inplanes, planes, stride)
        self.bn1 = nn.BatchNorm2d(inplanes)
        self.relu = nn.ReLU(inplace=True)

        self.conv2 = conv3x3(planes, planes)
        self.bn2 = nn.BatchNorm2d(planes)

        self.shortcut = nn.Sequential()
        if stride != 1 or inplanes != planes:
            self.shortcut = nn.Sequential(
                    nn.BatchNorm2d(inplanes),
                    self.relu, 
                    nn.Conv2d(inplanes, planes, kernel_size=1, stride=stride, bias=False)
                    )

    def forward(self, x): 
        residue = x

        out = self.bn1(x)
        out = self.relu(out)
        out = self.conv1(out)

        out = self.bn2(out)
        out = self.relu(out)
        out = self.conv2(out)

        out += self.shortcut(residue)

        return out

class BasicTransBlock(nn.Module):

    def __init__(self, in_ch, heads, dim_head, attn_drop=0., proj_drop=0., reduce_size=16, projection='interp', rel_pos=True):
        super().__init__()
        self.bn1 = nn.BatchNorm2d(in_ch)

        self.attn = LinearAttention(in_ch, heads=heads, dim_head=in_ch//heads, attn_drop=attn_drop, proj_drop=proj_drop, reduce_size=reduce_size, projection=projection, rel_pos=rel_pos)
        
        self.bn2 = nn.BatchNorm2d(in_ch)
        self.relu = nn.ReLU(inplace=True)
        self.mlp = nn.Conv2d(in_ch, in_ch, kernel_size=1, bias=False)
        # conv1x1 has not difference with mlp in performance

    def forward(self, x):

        out = self.bn1(x)
        out, q_k_attn = self.attn(out)
        
        out = out + x
        residue = out

        out = self.bn2(out)
        out = self.relu(out)
        out = self.mlp(out)

        out += residue

        return out

class BasicTransDecoderBlock(nn.Module):

    def __init__(self, in_ch, out_ch, heads, dim_head, attn_drop=0., proj_drop=0., reduce_size=16, projection='interp', rel_pos=True):
        super().__init__()

        self.bn_l = nn.BatchNorm2d(in_ch)
        self.bn_h = nn.BatchNorm2d(out_ch)

        self.conv_ch = nn.Conv2d(in_ch, out_ch, kernel_size=1)
        self.attn = LinearAttentionDecoder(in_ch, out_ch, heads=heads, dim_head=out_ch//heads, attn_drop=attn_drop, proj_drop=proj_drop, reduce_size=reduce_size, projection=projection, rel_pos=rel_pos)
        
        self.bn2 = nn.BatchNorm2d(out_ch)
        self.relu = nn.ReLU(inplace=True)
        self.mlp = nn.Conv2d(out_ch, out_ch, kernel_size=1, bias=False)

    def forward(self, x1, x2):

        residue = F.interpolate(self.conv_ch(x1), size=x2.shape[-2:], mode='bilinear', align_corners=True)
        #x1: low-res, x2: high-res
        x1 = self.bn_l(x1)
        x2 = self.bn_h(x2)

        out, q_k_attn = self.attn(x2, x1)
        
        out = out + residue
        residue = out

        out = self.bn2(out)
        out = self.relu(out)
        out = self.mlp(out)

        out += residue

        return out




########################################################################
# Transformer components

class LinearAttention(nn.Module):
    
    def __init__(self, dim, heads=4, dim_head=64, attn_drop=0., proj_drop=0., reduce_size=16, projection='interp', rel_pos=True):
        super().__init__()

        self.inner_dim = dim_head * heads
        self.heads = heads
        self.scale = dim_head ** (-0.5)
        self.dim_head = dim_head
        self.reduce_size = reduce_size
        self.projection = projection
        self.rel_pos = rel_pos
        
        # depthwise conv is slightly better than conv1x1
        #self.to_qkv = nn.Conv2d(dim, self.inner_dim*3, kernel_size=1, stride=1, padding=0, bias=True)
        #self.to_out = nn.Conv2d(self.inner_dim, dim, kernel_size=1, stride=1, padding=0, bias=True)
       
        self.to_qkv = depthwise_separable_conv(dim, self.inner_dim*3)
        self.to_out = depthwise_separable_conv(self.inner_dim, dim)


        self.attn_drop = nn.Dropout(attn_drop)
        self.proj_drop = nn.Dropout(proj_drop)

        if self.rel_pos:
            # 2D input-independent relative position encoding is a little bit better than
            # 1D input-denpendent counterpart
            self.relative_position_encoding = RelativePositionBias(heads, reduce_size, reduce_size)
            #self.relative_position_encoding = RelativePositionEmbedding(dim_head, reduce_size)

    def forward(self, x):

        B, C, H, W = x.shape

        #B, inner_dim, H, W
        qkv = self.to_qkv(x)
        q, k, v = qkv.chunk(3, dim=1)

        if self.projection == 'interp' and H != self.reduce_size:
            k, v = map(lambda t: F.interpolate(t, size=self.reduce_size, mode='bilinear', align_corners=True), (k, v))

        elif self.projection == 'maxpool' and H != self.reduce_size:
            k, v = map(lambda t: F.adaptive_max_pool2d(t, output_size=self.reduce_size), (k, v))
        
        q = rearrange(q, 'b (dim_head heads) h w -> b heads (h w) dim_head', dim_head=self.dim_head, heads=self.heads, h=H, w=W)
        k, v = map(lambda t: rearrange(t, 'b (dim_head heads) h w -> b heads (h w) dim_head', dim_head=self.dim_head, heads=self.heads, h=self.reduce_size, w=self.reduce_size), (k, v))

        q_k_attn = torch.einsum('bhid,bhjd->bhij', q, k)
        
        if self.rel_pos:
            relative_position_bias = self.relative_position_encoding(H, W)
            q_k_attn += relative_position_bias
            #rel_attn_h, rel_attn_w = self.relative_position_encoding(q, self.heads, H, W, self.dim_head)
            #q_k_attn = q_k_attn + rel_attn_h + rel_attn_w

        q_k_attn *= self.scale
        q_k_attn = F.softmax(q_k_attn, dim=-1)
        q_k_attn = self.attn_drop(q_k_attn)

        out = torch.einsum('bhij,bhjd->bhid', q_k_attn, v)
        out = rearrange(out, 'b heads (h w) dim_head -> b (dim_head heads) h w', h=H, w=W, dim_head=self.dim_head, heads=self.heads)

        out = self.to_out(out)
        out = self.proj_drop(out)

        return out, q_k_attn

class LinearAttentionDecoder(nn.Module):
    
    def __init__(self, in_dim, out_dim, heads=4, dim_head=64, attn_drop=0., proj_drop=0., reduce_size=16, projection='interp', rel_pos=True):
        super().__init__()

        self.inner_dim = dim_head * heads
        self.heads = heads
        self.scale = dim_head ** (-0.5)
        self.dim_head = dim_head
        self.reduce_size = reduce_size
        self.projection = projection
        self.rel_pos = rel_pos
        
        # depthwise conv is slightly better than conv1x1
        #self.to_kv = nn.Conv2d(dim, self.inner_dim*2, kernel_size=1, stride=1, padding=0, bias=True)
        #self.to_q = nn.Conv2d(dim, self.inner_dim, kernel_size=1, stride=1, padding=0, bias=True)
        #self.to_out = nn.Conv2d(self.inner_dim, dim, kernel_size=1, stride=1, padding=0, bias=True)
       
        self.to_kv = depthwise_separable_conv(in_dim, self.inner_dim*2)
        self.to_q = depthwise_separable_conv(out_dim, self.inner_dim)
        self.to_out = depthwise_separable_conv(self.inner_dim, out_dim)


        self.attn_drop = nn.Dropout(attn_drop)
        self.proj_drop = nn.Dropout(proj_drop)

        if self.rel_pos:
            self.relative_position_encoding = RelativePositionBias(heads, reduce_size, reduce_size)
            #self.relative_position_encoding = RelativePositionEmbedding(dim_head, reduce_size)

    def forward(self, q, x):

        B, C, H, W = x.shape # low-res feature shape
        BH, CH, HH, WH = q.shape # high-res feature shape

        k, v = self.to_kv(x).chunk(2, dim=1) #B, inner_dim, H, W
        q = self.to_q(q) #BH, inner_dim, HH, WH

        if self.projection == 'interp' and H != self.reduce_size:
            k, v = map(lambda t: F.interpolate(t, size=self.reduce_size, mode='bilinear', align_corners=True), (k, v))

        elif self.projection == 'maxpool' and H != self.reduce_size:
            k, v = map(lambda t: F.adaptive_max_pool2d(t, output_size=self.reduce_size), (k, v))
        
        q = rearrange(q, 'b (dim_head heads) h w -> b heads (h w) dim_head', dim_head=self.dim_head, heads=self.heads, h=HH, w=WH)
        k, v = map(lambda t: rearrange(t, 'b (dim_head heads) h w -> b heads (h w) dim_head', dim_head=self.dim_head, heads=self.heads, h=self.reduce_size, w=self.reduce_size), (k, v))

        q_k_attn = torch.einsum('bhid,bhjd->bhij', q, k)
        
        if self.rel_pos:
            relative_position_bias = self.relative_position_encoding(HH, WH)
            q_k_attn += relative_position_bias
            #rel_attn_h, rel_attn_w = self.relative_position_encoding(q, self.heads, HH, WH, self.dim_head)
            #q_k_attn = q_k_attn + rel_attn_h + rel_attn_w
       
        q_k_attn *= self.scale
        q_k_attn = F.softmax(q_k_attn, dim=-1)
        q_k_attn = self.attn_drop(q_k_attn)

        out = torch.einsum('bhij,bhjd->bhid', q_k_attn, v)
        out = rearrange(out, 'b heads (h w) dim_head -> b (dim_head heads) h w', h=HH, w=WH, dim_head=self.dim_head, heads=self.heads)

        out = self.to_out(out)
        out = self.proj_drop(out)

        return out, q_k_attn

class RelativePositionEmbedding(nn.Module):
    # input-dependent relative position
    def __init__(self, dim, shape):
        super().__init__()

        self.dim = dim
        self.shape = shape

        self.key_rel_w = nn.Parameter(torch.randn((2*self.shape-1, dim))*0.02)
        self.key_rel_h = nn.Parameter(torch.randn((2*self.shape-1, dim))*0.02)

        coords = torch.arange(self.shape)
        relative_coords = coords[None, :] - coords[:, None] # h, h
        relative_coords += self.shape - 1 # shift to start from 0
        
        self.register_buffer('relative_position_index', relative_coords)



    def forward(self, q, Nh, H, W, dim_head):
        # q: B, Nh, HW, dim
        B, _, _, dim = q.shape

        # q: B, Nh, H, W, dim_head
        q = rearrange(q, 'b heads (h w) dim_head -> b heads h w dim_head', b=B, dim_head=dim_head, heads=Nh, h=H, w=W)

        rel_logits_w = self.relative_logits_1d(q, self.key_rel_w, 'w')

        rel_logits_h = self.relative_logits_1d(q.permute(0, 1, 3, 2, 4), self.key_rel_h, 'h')

        return rel_logits_w, rel_logits_h

    def relative_logits_1d(self, q, rel_k, case):
        
        B, Nh, H, W, dim = q.shape

        rel_logits = torch.einsum('bhxyd,md->bhxym', q, rel_k) # B, Nh, H, W, 2*shape-1

        if W != self.shape:
            # self_relative_position_index origin shape: w, w
            # after repeat: W, w
            relative_index= torch.repeat_interleave(self.relative_position_index, W//self.shape, dim=0) # W, shape
        relative_index = relative_index.view(1, 1, 1, W, self.shape)
        relative_index = relative_index.repeat(B, Nh, H, 1, 1)

        rel_logits = torch.gather(rel_logits, 4, relative_index) # B, Nh, H, W, shape
        rel_logits = rel_logits.unsqueeze(3)
        rel_logits = rel_logits.repeat(1, 1, 1, self.shape, 1, 1)

        if case == 'w':
            rel_logits = rearrange(rel_logits, 'b heads H h W w -> b heads (H W) (h w)')

        elif case == 'h':
            rel_logits = rearrange(rel_logits, 'b heads W w H h -> b heads (H W) (h w)')

        return rel_logits




class RelativePositionBias(nn.Module):
    # input-independent relative position attention
    # As the number of parameters is smaller, so use 2D here
    # Borrowed some code from SwinTransformer: https://github.com/microsoft/Swin-Transformer/blob/main/models/swin_transformer.py
    def __init__(self, num_heads, h, w):
        super().__init__()
        self.num_heads = num_heads
        self.h = h
        self.w = w

        self.relative_position_bias_table = nn.Parameter(
                torch.randn((2*h-1) * (2*w-1), num_heads)*0.02)

        coords_h = torch.arange(self.h)
        coords_w = torch.arange(self.w)
        coords = torch.stack(torch.meshgrid([coords_h, coords_w])) # 2, h, w
        coords_flatten = torch.flatten(coords, 1) # 2, hw

        relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :]
        relative_coords = relative_coords.permute(1, 2, 0).contiguous()
        relative_coords[:, :, 0] += self.h - 1
        relative_coords[:, :, 1] += self.w - 1
        relative_coords[:, :, 0] *= 2 * self.h - 1
        relative_position_index = relative_coords.sum(-1) # hw, hw
    
        self.register_buffer("relative_position_index", relative_position_index)

    def forward(self, H, W):
        
        relative_position_bias = self.relative_position_bias_table[self.relative_position_index.view(-1)].view(self.h, self.w, self.h*self.w, -1) #h, w, hw, nH
        relative_position_bias_expand_h = torch.repeat_interleave(relative_position_bias, H//self.h, dim=0)
        relative_position_bias_expanded = torch.repeat_interleave(relative_position_bias_expand_h, W//self.w, dim=1) #HW, hw, nH
        
        relative_position_bias_expanded = relative_position_bias_expanded.view(H*W, self.h*self.w, self.num_heads).permute(2, 0, 1).contiguous().unsqueeze(0)

        return relative_position_bias_expanded


###########################################################################
# Unet Transformer building block

class down_block_trans(nn.Module):
    def __init__(self, in_ch, out_ch, num_block, bottleneck=False, maxpool=True, heads=4, dim_head=64, attn_drop=0., proj_drop=0., reduce_size=16, projection='interp', rel_pos=True):

        super().__init__()

        block_list = []

        if bottleneck:
            block = BottleneckBlock
        else:
            block = BasicBlock

        attn_block = BasicTransBlock

        if maxpool:
            block_list.append(nn.MaxPool2d(2))
            block_list.append(block(in_ch, out_ch, stride=1))
        else:
            block_list.append(block(in_ch, out_ch, stride=2))
        
        assert num_block > 0
        for i in range(num_block):
            block_list.append(attn_block(out_ch, heads, dim_head, attn_drop=attn_drop, proj_drop=proj_drop, reduce_size=reduce_size, projection=projection, rel_pos=rel_pos))
        self.blocks = nn.Sequential(*block_list)

        
    def forward(self, x):
        
        out = self.blocks(x)


        return out

class up_block_trans(nn.Module):
    def __init__(self, in_ch, out_ch, num_block, bottleneck=False, heads=4, dim_head=64, attn_drop=0., proj_drop=0., reduce_size=16, projection='interp', rel_pos=True):
        super().__init__()
 
        self.attn_decoder = BasicTransDecoderBlock(in_ch, out_ch, heads=heads, dim_head=dim_head, attn_drop=attn_drop, proj_drop=proj_drop, reduce_size=reduce_size, projection=projection, rel_pos=rel_pos)

        if bottleneck:
            block = BottleneckBlock
        else:
            block = BasicBlock
        attn_block = BasicTransBlock
        
        block_list = []

        for i in range(num_block):
            block_list.append(attn_block(out_ch, heads, dim_head, attn_drop=attn_drop, proj_drop=proj_drop, reduce_size=reduce_size, projection=projection, rel_pos=rel_pos))

        block_list.append(block(2*out_ch, out_ch, stride=1))

        self.blocks = nn.Sequential(*block_list)

    def forward(self, x1, x2):
        # x1: low-res feature, x2: high-res feature
        out = self.attn_decoder(x1, x2)
        out = torch.cat([out, x2], dim=1)
        out = self.blocks(out)

        return out

class block_trans(nn.Module):
    def __init__(self, in_ch, num_block, heads=4, dim_head=64, attn_drop=0., proj_drop=0., reduce_size=16, projection='interp', rel_pos=True):

        super().__init__()

        block_list = []

        attn_block = BasicTransBlock

        assert num_block > 0
        for i in range(num_block):
            block_list.append(attn_block(in_ch, heads, dim_head, attn_drop=attn_drop, proj_drop=proj_drop, reduce_size=reduce_size, projection=projection, rel_pos=rel_pos))
        self.blocks = nn.Sequential(*block_list)

        
    def forward(self, x):
        
        out = self.blocks(x)


        return out