File size: 17,753 Bytes
0f90f73
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# Fast Fourier Convolution NeurIPS 2020
# original implementation https://github.com/pkumivision/FFC/blob/main/model_zoo/ffc.py
# paper https://proceedings.neurips.cc/paper/2020/file/2fd5d41ec6cfab47e32164d5624269b1-Paper.pdf

import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F

from saicinpainting.training.modules.base import get_activation, BaseDiscriminator
from saicinpainting.training.modules.spatial_transform import LearnableSpatialTransformWrapper
from saicinpainting.training.modules.squeeze_excitation import SELayer
from saicinpainting.utils import get_shape


class FFCSE_block(nn.Module):

    def __init__(self, channels, ratio_g):
        super(FFCSE_block, self).__init__()
        in_cg = int(channels * ratio_g)
        in_cl = channels - in_cg
        r = 16

        self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
        self.conv1 = nn.Conv2d(channels, channels // r,
                               kernel_size=1, bias=True)
        self.relu1 = nn.ReLU(inplace=True)
        self.conv_a2l = None if in_cl == 0 else nn.Conv2d(
            channels // r, in_cl, kernel_size=1, bias=True)
        self.conv_a2g = None if in_cg == 0 else nn.Conv2d(
            channels // r, in_cg, kernel_size=1, bias=True)
        self.sigmoid = nn.Sigmoid()

    def forward(self, x):
        x = x if type(x) is tuple else (x, 0)
        id_l, id_g = x

        x = id_l if type(id_g) is int else torch.cat([id_l, id_g], dim=1)
        x = self.avgpool(x)
        x = self.relu1(self.conv1(x))

        x_l = 0 if self.conv_a2l is None else id_l * \
            self.sigmoid(self.conv_a2l(x))
        x_g = 0 if self.conv_a2g is None else id_g * \
            self.sigmoid(self.conv_a2g(x))
        return x_l, x_g


class FourierUnit(nn.Module):

    def __init__(self, in_channels, out_channels, groups=1, spatial_scale_factor=None, spatial_scale_mode='bilinear',
                 spectral_pos_encoding=False, use_se=False, se_kwargs=None, ffc3d=False, fft_norm='ortho'):
        # bn_layer not used
        super(FourierUnit, self).__init__()
        self.groups = groups

        self.conv_layer = torch.nn.Conv2d(in_channels=in_channels * 2 + (2 if spectral_pos_encoding else 0),
                                          out_channels=out_channels * 2,
                                          kernel_size=1, stride=1, padding=0, groups=self.groups, bias=False)
        self.bn = torch.nn.BatchNorm2d(out_channels * 2)
        self.relu = torch.nn.ReLU(inplace=True)

        # squeeze and excitation block
        self.use_se = use_se
        if use_se:
            if se_kwargs is None:
                se_kwargs = {}
            self.se = SELayer(self.conv_layer.in_channels, **se_kwargs)

        self.spatial_scale_factor = spatial_scale_factor
        self.spatial_scale_mode = spatial_scale_mode
        self.spectral_pos_encoding = spectral_pos_encoding
        self.ffc3d = ffc3d
        self.fft_norm = fft_norm

    def forward(self, x):
        batch = x.shape[0]

        if self.spatial_scale_factor is not None:
            orig_size = x.shape[-2:]
            x = F.interpolate(x, scale_factor=self.spatial_scale_factor, mode=self.spatial_scale_mode, align_corners=False)

        r_size = x.size()
        # (batch, c, h, w/2+1, 2)
        fft_dim = (-3, -2, -1) if self.ffc3d else (-2, -1)
        ffted = torch.fft.rfftn(x, dim=fft_dim, norm=self.fft_norm)
        ffted = torch.stack((ffted.real, ffted.imag), dim=-1)
        ffted = ffted.permute(0, 1, 4, 2, 3).contiguous()  # (batch, c, 2, h, w/2+1)
        ffted = ffted.view((batch, -1,) + ffted.size()[3:])

        if self.spectral_pos_encoding:
            height, width = ffted.shape[-2:]
            coords_vert = torch.linspace(0, 1, height)[None, None, :, None].expand(batch, 1, height, width).to(ffted)
            coords_hor = torch.linspace(0, 1, width)[None, None, None, :].expand(batch, 1, height, width).to(ffted)
            ffted = torch.cat((coords_vert, coords_hor, ffted), dim=1)

        if self.use_se:
            ffted = self.se(ffted)

        ffted = self.conv_layer(ffted)  # (batch, c*2, h, w/2+1)
        ffted = self.relu(self.bn(ffted))

        ffted = ffted.view((batch, -1, 2,) + ffted.size()[2:]).permute(
            0, 1, 3, 4, 2).contiguous()  # (batch,c, t, h, w/2+1, 2)
        ffted = torch.complex(ffted[..., 0], ffted[..., 1])

        ifft_shape_slice = x.shape[-3:] if self.ffc3d else x.shape[-2:]
        output = torch.fft.irfftn(ffted, s=ifft_shape_slice, dim=fft_dim, norm=self.fft_norm)

        if self.spatial_scale_factor is not None:
            output = F.interpolate(output, size=orig_size, mode=self.spatial_scale_mode, align_corners=False)

        return output


class SpectralTransform(nn.Module):

    def __init__(self, in_channels, out_channels, stride=1, groups=1, enable_lfu=True, **fu_kwargs):
        # bn_layer not used
        super(SpectralTransform, self).__init__()
        self.enable_lfu = enable_lfu
        if stride == 2:
            self.downsample = nn.AvgPool2d(kernel_size=(2, 2), stride=2)
        else:
            self.downsample = nn.Identity()

        self.stride = stride
        self.conv1 = nn.Sequential(
            nn.Conv2d(in_channels, out_channels //
                      2, kernel_size=1, groups=groups, bias=False),
            nn.BatchNorm2d(out_channels // 2),
            nn.ReLU(inplace=True)
        )
        self.fu = FourierUnit(
            out_channels // 2, out_channels // 2, groups, **fu_kwargs)
        if self.enable_lfu:
            self.lfu = FourierUnit(
                out_channels // 2, out_channels // 2, groups)
        self.conv2 = torch.nn.Conv2d(
            out_channels // 2, out_channels, kernel_size=1, groups=groups, bias=False)

    def forward(self, x):

        x = self.downsample(x)
        x = self.conv1(x)
        output = self.fu(x)

        if self.enable_lfu:
            n, c, h, w = x.shape
            split_no = 2
            split_s = h // split_no
            xs = torch.cat(torch.split(
                x[:, :c // 4], split_s, dim=-2), dim=1).contiguous()
            xs = torch.cat(torch.split(xs, split_s, dim=-1),
                           dim=1).contiguous()
            xs = self.lfu(xs)
            xs = xs.repeat(1, 1, split_no, split_no).contiguous()
        else:
            xs = 0

        output = self.conv2(x + output + xs)

        return output


class FFC(nn.Module):

    def __init__(self, in_channels, out_channels, kernel_size,
                 ratio_gin, ratio_gout, stride=1, padding=0,
                 dilation=1, groups=1, bias=False, enable_lfu=True,
                 padding_type='reflect', gated=False, **spectral_kwargs):
        super(FFC, self).__init__()

        assert stride == 1 or stride == 2, "Stride should be 1 or 2."
        self.stride = stride

        in_cg = int(in_channels * ratio_gin)
        in_cl = in_channels - in_cg
        out_cg = int(out_channels * ratio_gout)
        out_cl = out_channels - out_cg
        #groups_g = 1 if groups == 1 else int(groups * ratio_gout)
        #groups_l = 1 if groups == 1 else groups - groups_g

        self.ratio_gin = ratio_gin
        self.ratio_gout = ratio_gout
        self.global_in_num = in_cg

        module = nn.Identity if in_cl == 0 or out_cl == 0 else nn.Conv2d
        self.convl2l = module(in_cl, out_cl, kernel_size,
                              stride, padding, dilation, groups, bias, padding_mode=padding_type)
        module = nn.Identity if in_cl == 0 or out_cg == 0 else nn.Conv2d
        self.convl2g = module(in_cl, out_cg, kernel_size,
                              stride, padding, dilation, groups, bias, padding_mode=padding_type)
        module = nn.Identity if in_cg == 0 or out_cl == 0 else nn.Conv2d
        self.convg2l = module(in_cg, out_cl, kernel_size,
                              stride, padding, dilation, groups, bias, padding_mode=padding_type)
        module = nn.Identity if in_cg == 0 or out_cg == 0 else SpectralTransform
        self.convg2g = module(
            in_cg, out_cg, stride, 1 if groups == 1 else groups // 2, enable_lfu, **spectral_kwargs)

        self.gated = gated
        module = nn.Identity if in_cg == 0 or out_cl == 0 or not self.gated else nn.Conv2d
        self.gate = module(in_channels, 2, 1)

    def forward(self, x):
        x_l, x_g = x if type(x) is tuple else (x, 0)
        out_xl, out_xg = 0, 0

        if self.gated:
            total_input_parts = [x_l]
            if torch.is_tensor(x_g):
                total_input_parts.append(x_g)
            total_input = torch.cat(total_input_parts, dim=1)

            gates = torch.sigmoid(self.gate(total_input))
            g2l_gate, l2g_gate = gates.chunk(2, dim=1)
        else:
            g2l_gate, l2g_gate = 1, 1

        if self.ratio_gout != 1:
            out_xl = self.convl2l(x_l) + self.convg2l(x_g) * g2l_gate
        if self.ratio_gout != 0:
            out_xg = self.convl2g(x_l) * l2g_gate + self.convg2g(x_g)

        return out_xl, out_xg


class FFC_BN_ACT(nn.Module):

    def __init__(self, in_channels, out_channels,
                 kernel_size, ratio_gin, ratio_gout,
                 stride=1, padding=0, dilation=1, groups=1, bias=False,
                 norm_layer=nn.BatchNorm2d, activation_layer=nn.Identity,
                 padding_type='reflect',
                 enable_lfu=True, **kwargs):
        super(FFC_BN_ACT, self).__init__()
        self.ffc = FFC(in_channels, out_channels, kernel_size,
                       ratio_gin, ratio_gout, stride, padding, dilation,
                       groups, bias, enable_lfu, padding_type=padding_type, **kwargs)
        lnorm = nn.Identity if ratio_gout == 1 else norm_layer
        gnorm = nn.Identity if ratio_gout == 0 else norm_layer
        global_channels = int(out_channels * ratio_gout)
        self.bn_l = lnorm(out_channels - global_channels)
        self.bn_g = gnorm(global_channels)

        lact = nn.Identity if ratio_gout == 1 else activation_layer
        gact = nn.Identity if ratio_gout == 0 else activation_layer
        self.act_l = lact(inplace=True)
        self.act_g = gact(inplace=True)

    def forward(self, x):
        x_l, x_g = self.ffc(x)
        x_l = self.act_l(self.bn_l(x_l))
        x_g = self.act_g(self.bn_g(x_g))
        return x_l, x_g


class FFCResnetBlock(nn.Module):
    def __init__(self, dim, padding_type, norm_layer, activation_layer=nn.ReLU, dilation=1,
                 spatial_transform_kwargs=None, inline=False, **conv_kwargs):
        super().__init__()
        self.conv1 = FFC_BN_ACT(dim, dim, kernel_size=3, padding=dilation, dilation=dilation,
                                norm_layer=norm_layer,
                                activation_layer=activation_layer,
                                padding_type=padding_type,
                                **conv_kwargs)
        self.conv2 = FFC_BN_ACT(dim, dim, kernel_size=3, padding=dilation, dilation=dilation,
                                norm_layer=norm_layer,
                                activation_layer=activation_layer,
                                padding_type=padding_type,
                                **conv_kwargs)
        if spatial_transform_kwargs is not None:
            self.conv1 = LearnableSpatialTransformWrapper(self.conv1, **spatial_transform_kwargs)
            self.conv2 = LearnableSpatialTransformWrapper(self.conv2, **spatial_transform_kwargs)
        self.inline = inline

    def forward(self, x):
        if self.inline:
            x_l, x_g = x[:, :-self.conv1.ffc.global_in_num], x[:, -self.conv1.ffc.global_in_num:]
        else:
            x_l, x_g = x if type(x) is tuple else (x, 0)

        id_l, id_g = x_l, x_g

        x_l, x_g = self.conv1((x_l, x_g))
        x_l, x_g = self.conv2((x_l, x_g))

        x_l, x_g = id_l + x_l, id_g + x_g
        out = x_l, x_g
        if self.inline:
            out = torch.cat(out, dim=1)
        return out


class ConcatTupleLayer(nn.Module):
    def forward(self, x):
        assert isinstance(x, tuple)
        x_l, x_g = x
        assert torch.is_tensor(x_l) or torch.is_tensor(x_g)
        if not torch.is_tensor(x_g):
            return x_l
        return torch.cat(x, dim=1)


class FFCResNetGenerator(nn.Module):
    def __init__(self, input_nc, output_nc, ngf=64, n_downsampling=3, n_blocks=9, norm_layer=nn.BatchNorm2d,
                 padding_type='reflect', activation_layer=nn.ReLU,
                 up_norm_layer=nn.BatchNorm2d, up_activation=nn.ReLU(True),
                 init_conv_kwargs={}, downsample_conv_kwargs={}, resnet_conv_kwargs={},
                 spatial_transform_layers=None, spatial_transform_kwargs={},
                 add_out_act=True, max_features=1024, out_ffc=False, out_ffc_kwargs={}):
        assert (n_blocks >= 0)
        super().__init__()

        model = [nn.ReflectionPad2d(3),
                 FFC_BN_ACT(input_nc, ngf, kernel_size=7, padding=0, norm_layer=norm_layer,
                            activation_layer=activation_layer, **init_conv_kwargs)]

        ### downsample
        for i in range(n_downsampling):
            mult = 2 ** i
            if i == n_downsampling - 1:
                cur_conv_kwargs = dict(downsample_conv_kwargs)
                cur_conv_kwargs['ratio_gout'] = resnet_conv_kwargs.get('ratio_gin', 0)
            else:
                cur_conv_kwargs = downsample_conv_kwargs
            model += [FFC_BN_ACT(min(max_features, ngf * mult),
                                 min(max_features, ngf * mult * 2),
                                 kernel_size=3, stride=2, padding=1,
                                 norm_layer=norm_layer,
                                 activation_layer=activation_layer,
                                 **cur_conv_kwargs)]

        mult = 2 ** n_downsampling
        feats_num_bottleneck = min(max_features, ngf * mult)

        ### resnet blocks
        for i in range(n_blocks):
            cur_resblock = FFCResnetBlock(feats_num_bottleneck, padding_type=padding_type, activation_layer=activation_layer,
                                          norm_layer=norm_layer, **resnet_conv_kwargs)
            if spatial_transform_layers is not None and i in spatial_transform_layers:
                cur_resblock = LearnableSpatialTransformWrapper(cur_resblock, **spatial_transform_kwargs)
            model += [cur_resblock]

        model += [ConcatTupleLayer()]

        ### upsample
        for i in range(n_downsampling):
            mult = 2 ** (n_downsampling - i)
            model += [nn.ConvTranspose2d(min(max_features, ngf * mult),
                                         min(max_features, int(ngf * mult / 2)),
                                         kernel_size=3, stride=2, padding=1, output_padding=1),
                      up_norm_layer(min(max_features, int(ngf * mult / 2))),
                      up_activation]

        if out_ffc:
            model += [FFCResnetBlock(ngf, padding_type=padding_type, activation_layer=activation_layer,
                                     norm_layer=norm_layer, inline=True, **out_ffc_kwargs)]

        model += [nn.ReflectionPad2d(3),
                  nn.Conv2d(ngf, output_nc, kernel_size=7, padding=0)]
        if add_out_act:
            model.append(get_activation('tanh' if add_out_act is True else add_out_act))
        self.model = nn.Sequential(*model)

    def forward(self, input):
        return self.model(input)


class FFCNLayerDiscriminator(BaseDiscriminator):
    def __init__(self, input_nc, ndf=64, n_layers=3, norm_layer=nn.BatchNorm2d, max_features=512,
                 init_conv_kwargs={}, conv_kwargs={}):
        super().__init__()
        self.n_layers = n_layers

        def _act_ctor(inplace=True):
            return nn.LeakyReLU(negative_slope=0.2, inplace=inplace)

        kw = 3
        padw = int(np.ceil((kw-1.0)/2))
        sequence = [[FFC_BN_ACT(input_nc, ndf, kernel_size=kw, padding=padw, norm_layer=norm_layer,
                                activation_layer=_act_ctor, **init_conv_kwargs)]]

        nf = ndf
        for n in range(1, n_layers):
            nf_prev = nf
            nf = min(nf * 2, max_features)

            cur_model = [
                FFC_BN_ACT(nf_prev, nf,
                           kernel_size=kw, stride=2, padding=padw,
                           norm_layer=norm_layer,
                           activation_layer=_act_ctor,
                           **conv_kwargs)
            ]
            sequence.append(cur_model)

        nf_prev = nf
        nf = min(nf * 2, 512)

        cur_model = [
            FFC_BN_ACT(nf_prev, nf,
                       kernel_size=kw, stride=1, padding=padw,
                       norm_layer=norm_layer,
                       activation_layer=lambda *args, **kwargs: nn.LeakyReLU(*args, negative_slope=0.2, **kwargs),
                       **conv_kwargs),
            ConcatTupleLayer()
        ]
        sequence.append(cur_model)

        sequence += [[nn.Conv2d(nf, 1, kernel_size=kw, stride=1, padding=padw)]]

        for n in range(len(sequence)):
            setattr(self, 'model'+str(n), nn.Sequential(*sequence[n]))

    def get_all_activations(self, x):
        res = [x]
        for n in range(self.n_layers + 2):
            model = getattr(self, 'model' + str(n))
            res.append(model(res[-1]))
        return res[1:]

    def forward(self, x):
        act = self.get_all_activations(x)
        feats = []
        for out in act[:-1]:
            if isinstance(out, tuple):
                if torch.is_tensor(out[1]):
                    out = torch.cat(out, dim=1)
                else:
                    out = out[0]
            feats.append(out)
        return act[-1], feats