File size: 17,391 Bytes
5a5d1a8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
"""Modules for generator blocks."""

from typing import Tuple

import einops
import torch
import torch.nn.functional as F

from torch.distributions import normal
from torch.nn.modules.pixelshuffle import PixelUnshuffle
from torch.nn.utils.parametrizations import spectral_norm

from .layers import AttentionLayer
from .layers.utils import get_conv_layer


class GBlock(torch.nn.Module):
    """Residual generator block without upsampling."""

    def __init__(
        self,
        input_channels: int = 12,
        output_channels: int = 12,
        conv_type: str = "standard",
        spectral_normalized_eps=0.0001,
    ):
        """
        G Block from Skillful Nowcasting, see https://arxiv.org/pdf/2104.00954.pdf.

        Args:
            input_channels: Number of input channels
            output_channels: Number of output channels
            conv_type: Type of convolution desired, see satflow/models/utils.py for options
            spectral_normalized_eps: constrains the spectral norm of the weights.
        """
        super().__init__()
        self.output_channels = output_channels
        self.bn1 = torch.nn.BatchNorm2d(input_channels)
        self.bn2 = torch.nn.BatchNorm2d(input_channels)
        self.relu = torch.nn.ReLU()
        # Upsample in the 1x1
        conv2d = get_conv_layer(conv_type)
        self.conv_1x1 = spectral_norm(
            conv2d(
                in_channels=input_channels,
                out_channels=output_channels,
                kernel_size=1,
            ),
            eps=spectral_normalized_eps,
        )
        # Upsample 2D conv
        self.first_conv_3x3 = spectral_norm(
            conv2d(
                in_channels=input_channels,
                out_channels=input_channels,
                kernel_size=3,
                padding=1,
            ),
            eps=spectral_normalized_eps,
        )
        self.last_conv_3x3 = spectral_norm(
            conv2d(
                in_channels=input_channels, out_channels=output_channels, kernel_size=3, padding=1
            ),
            eps=spectral_normalized_eps,
        )

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        """Apply the forward function."""
        # Optionally spectrally normalized 1x1 convolution
        if x.shape[1] != self.output_channels:
            sc = self.conv_1x1(x)
        else:
            sc = x

        x2 = self.bn1(x)
        x2 = self.relu(x2)
        x2 = self.first_conv_3x3(x2)  # Make sure size is doubled
        x2 = self.bn2(x2)
        x2 = self.relu(x2)
        x2 = self.last_conv_3x3(x2)
        # Sum combine, residual connection
        x = x2 + sc
        return x


class UpsampleGBlock(torch.nn.Module):
    """Residual generator block with upsampling."""

    def __init__(
        self,
        input_channels: int = 12,
        output_channels: int = 12,
        conv_type: str = "standard",
        spectral_normalized_eps=0.0001,
    ):
        """
        G Block from Skillful Nowcasting, see https://arxiv.org/pdf/2104.00954.pdf.

        Args:
            input_channels: Number of input channels.
            output_channels: Number of output channels.
            conv_type: Type of convolution desired, see satflow/models/utils.py for options.
            spectral_normalized_eps: constrains the spectral norm of the weights.
        """
        super().__init__()
        self.output_channels = output_channels
        self.bn1 = torch.nn.BatchNorm2d(input_channels)
        self.bn2 = torch.nn.BatchNorm2d(input_channels)
        self.relu = torch.nn.ReLU()
        # Upsample in the 1x1
        conv2d = get_conv_layer(conv_type)
        self.conv_1x1 = spectral_norm(
            conv2d(
                in_channels=input_channels,
                out_channels=output_channels,
                kernel_size=1,
            ),
            eps=spectral_normalized_eps,
        )
        self.upsample = torch.nn.Upsample(scale_factor=2, mode="nearest")
        # Upsample 2D conv
        self.first_conv_3x3 = spectral_norm(
            conv2d(
                in_channels=input_channels,
                out_channels=input_channels,
                kernel_size=3,
                padding=1,
            ),
            eps=spectral_normalized_eps,
        )
        self.last_conv_3x3 = spectral_norm(
            conv2d(
                in_channels=input_channels, out_channels=output_channels, kernel_size=3, padding=1
            ),
            eps=spectral_normalized_eps,
        )

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        """Apply the forward function."""
        # Spectrally nsormalized 1x1 convolution
        sc = self.upsample(x)
        sc = self.conv_1x1(sc)

        x2 = self.bn1(x)
        x2 = self.relu(x2)
        # Upsample
        x2 = self.upsample(x2)
        x2 = self.first_conv_3x3(x2)  # Make sure size is doubled
        x2 = self.bn2(x2)
        x2 = self.relu(x2)
        x2 = self.last_conv_3x3(x2)
        # Sum combine, residual connection
        x = x2 + sc
        return x


class DBlock(torch.nn.Module):
    """D block class."""

    def __init__(
        self,
        input_channels: int = 12,
        output_channels: int = 12,
        conv_type: str = "standard",
        first_relu: bool = True,
        keep_same_output: bool = False,
    ):
        """
        D and 3D Block from Skillful Nowcasting, see https://arxiv.org/pdf/2104.00954.pdf.

        Args:
            input_channels: Number of input channels
            output_channels: Number of output channels
            conv_type: Convolution type, see satflow/models/utils.py for options
            first_relu: Whether to have an ReLU before the first 3x3 convolution
            keep_same_output: Whether the output should have the same spatial dimensions
            as input, if False, downscales by 2
        """
        super().__init__()
        self.input_channels = input_channels
        self.output_channels = output_channels
        self.first_relu = first_relu
        self.keep_same_output = keep_same_output
        self.conv_type = conv_type
        conv2d = get_conv_layer(conv_type)
        if conv_type == "3d":
            # 3D Average pooling
            self.pooling = torch.nn.AvgPool3d(kernel_size=2, stride=2)
        else:
            self.pooling = torch.nn.AvgPool2d(kernel_size=2, stride=2)
        self.conv_1x1 = spectral_norm(
            conv2d(
                in_channels=input_channels,
                out_channels=output_channels,
                kernel_size=1,
            )
        )
        self.first_conv_3x3 = spectral_norm(
            conv2d(
                in_channels=input_channels,
                out_channels=output_channels,
                kernel_size=3,
                padding=1,
            )
        )
        self.last_conv_3x3 = spectral_norm(
            conv2d(
                in_channels=output_channels,
                out_channels=output_channels,
                kernel_size=3,
                padding=1,
                stride=1,
            )
        )
        # Downsample at end of 3x3
        self.relu = torch.nn.ReLU()
        # Concatenate to double final channels and keep reduced spatial extent

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        """Apply the D residual block."""
        if self.input_channels != self.output_channels:
            x1 = self.conv_1x1(x)
            if not self.keep_same_output:
                x1 = self.pooling(x1)
        else:
            x1 = x

        if self.first_relu:
            x = self.relu(x)
        x = self.first_conv_3x3(x)
        x = self.relu(x)
        x = self.last_conv_3x3(x)

        if not self.keep_same_output:
            x = self.pooling(x)
        x = x1 + x  # Sum the outputs should be half spatial and double channels
        return x


class LBlock(torch.nn.Module):
    """Residual block for the Latent Stack."""

    def __init__(
        self,
        input_channels: int = 12,
        output_channels: int = 12,
        kernel_size: int = 3,
        conv_type: str = "standard",
    ):
        """
        Initialize the L-block.

        L-Block for increasing the number of channels in the input
         from Skillful Nowcasting, see https://arxiv.org/pdf/2104.00954.pdf
        Args:
            input_channels: Number of input channels
            output_channels: Number of output channels
            conv_type: Which type of convolution desired, see satflow/models/utils.py for options
        """
        super().__init__()
        # Output size should be channel_out - channel_in
        self.input_channels = input_channels
        self.output_channels = output_channels
        conv2d = get_conv_layer(conv_type)
        self.conv_1x1 = conv2d(
            in_channels=input_channels,
            out_channels=output_channels - input_channels,
            kernel_size=1,
        )

        self.first_conv_3x3 = conv2d(
            input_channels,
            out_channels=output_channels,
            kernel_size=kernel_size,
            padding=1,
            stride=1,
        )
        self.relu = torch.nn.ReLU()
        self.last_conv_3x3 = conv2d(
            in_channels=output_channels,
            out_channels=output_channels,
            kernel_size=kernel_size,
            padding=1,
            stride=1,
        )

    def forward(self, x) -> torch.Tensor:
        """Apply the L residual block to this tensor."""
        if self.input_channels < self.output_channels:
            sc = self.conv_1x1(x)
            sc = torch.cat([x, sc], dim=1)
        else:
            sc = x

        x2 = self.relu(x)
        x2 = self.first_conv_3x3(x2)
        x2 = self.relu(x2)
        x2 = self.last_conv_3x3(x2)
        return x2 + sc


class ContextConditioningStack(torch.nn.Module):
    """Context conditioning stack."""

    def __init__(
        self,
        input_channels: int = 1,
        output_channels: int = 768,
        num_context_steps: int = 4,
        conv_type: str = "standard",
    ):
        """
        Conditioning Stack using the context images from Skillful Nowcasting, see https://arxiv.org/pdf/2104.00954.pdf.

        Args:
            input_channels: Number of input channels per timestep
            output_channels: Number of output channels for the lowest block
            num_context_steps: number of context steps (int)
            conv_type: Type of 2D convolution to use, see satflow/models/utils.py for options
            **kwargs: Allow initialize of the parameters above through key pairs
        """
        super().__init__()

        conv2d = get_conv_layer(conv_type)
        self.space2depth = PixelUnshuffle(downscale_factor=2)
        # Process each observation processed separately with 4 downsample blocks
        # Concatenate across channel dimension, and for each output, 3x3 spectrally
        # normalized convolution to reduce number of channels by 2, followed by ReLU
        self.d1 = DBlock(
            input_channels=4 * input_channels,
            output_channels=((output_channels // 4) * input_channels) // num_context_steps,
            conv_type=conv_type,
        )
        self.d2 = DBlock(
            input_channels=((output_channels // 4) * input_channels) // num_context_steps,
            output_channels=((output_channels // 2) * input_channels) // num_context_steps,
            conv_type=conv_type,
        )
        self.d3 = DBlock(
            input_channels=((output_channels // 2) * input_channels) // num_context_steps,
            output_channels=(output_channels * input_channels) // num_context_steps,
            conv_type=conv_type,
        )
        self.d4 = DBlock(
            input_channels=(output_channels * input_channels) // num_context_steps,
            output_channels=(output_channels * 2 * input_channels) // num_context_steps,
            conv_type=conv_type,
        )
        self.conv1 = spectral_norm(
            conv2d(
                in_channels=(output_channels // 4) * input_channels,
                out_channels=(output_channels // 8) * input_channels,
                kernel_size=3,
                padding=1,
            )
        )

        self.conv2 = spectral_norm(
            conv2d(
                in_channels=(output_channels // 2) * input_channels,
                out_channels=(output_channels // 4) * input_channels,
                kernel_size=3,
                padding=1,
            )
        )

        self.conv3 = spectral_norm(
            conv2d(
                in_channels=output_channels * input_channels,
                out_channels=(output_channels // 2) * input_channels,
                kernel_size=3,
                padding=1,
            )
        )

        self.conv4 = spectral_norm(
            conv2d(
                in_channels=output_channels * 2 * input_channels,
                out_channels=output_channels * input_channels,
                kernel_size=3,
                padding=1,
            )
        )

        self.relu = torch.nn.ReLU()

    def forward(
        self, x: torch.Tensor
    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
        """Generate the condition representation."""
        # Each timestep processed separately
        x = self.space2depth(x)
        steps = x.size(1)  # Number of timesteps
        scale_1 = []
        scale_2 = []
        scale_3 = []
        scale_4 = []
        for i in range(steps):
            s1 = self.d1(x[:, i, :, :, :])
            s2 = self.d2(s1)
            s3 = self.d3(s2)
            s4 = self.d4(s3)
            scale_1.append(s1)
            scale_2.append(s2)
            scale_3.append(s3)
            scale_4.append(s4)
        scale_1 = torch.stack(scale_1, dim=1)  # B, T, C, H, W and want along C dimension
        scale_2 = torch.stack(scale_2, dim=1)  # B, T, C, H, W and want along C dimension
        scale_3 = torch.stack(scale_3, dim=1)  # B, T, C, H, W and want along C dimension
        scale_4 = torch.stack(scale_4, dim=1)  # B, T, C, H, W and want along C dimension
        # Mixing layer
        scale_1 = self._mixing_layer(scale_1, self.conv1)
        scale_2 = self._mixing_layer(scale_2, self.conv2)
        scale_3 = self._mixing_layer(scale_3, self.conv3)
        scale_4 = self._mixing_layer(scale_4, self.conv4)
        return scale_1, scale_2, scale_3, scale_4

    def _mixing_layer(self, inputs, conv_block):
        """Combine the inputs and then passed into the convolution stack."""
        # Convert from [batch_size, time, h, w, c] -> [batch_size, h, w, c * time]
        # then perform convolution on the output while preserving number of c.
        stacked_inputs = einops.rearrange(inputs, "b t c h w -> b (c t) h w")
        return F.relu(conv_block(stacked_inputs))


class LatentConditioningStack(torch.nn.Module):
    """Latent conditioning stack class."""

    def __init__(
        self,
        shape: (int, int, int) = (8, 8, 8),
        output_channels: int = 768,
        use_attention: bool = True,
    ):
        """
        Latent conditioning stack from Skillful Nowcasting, see https://arxiv.org/pdf/2104.00954.pdf.

        Args:
            shape: Shape of the latent space, Should be (H/32,W/32,x) of the final image shape
            output_channels: Number of output channels for the conditioning stack
            use_attention: Whether to have a self-attention block or not
            **kwargs: allow initialize of the parameters above through key pairs
        """
        super().__init__()

        self.shape = shape
        self.use_attention = use_attention
        self.distribution = normal.Normal(loc=torch.Tensor([0.0]), scale=torch.Tensor([1.0]))

        self.conv_3x3 = spectral_norm(
            torch.nn.Conv2d(
                in_channels=shape[0], out_channels=shape[0], kernel_size=(3, 3), padding=1
            )
        )
        self.l_block1 = LBlock(input_channels=shape[0], output_channels=output_channels // 32)
        self.l_block2 = LBlock(
            input_channels=output_channels // 32, output_channels=output_channels // 16
        )
        self.l_block3 = LBlock(
            input_channels=output_channels // 16, output_channels=output_channels // 4
        )
        if self.use_attention:
            self.att_block = AttentionLayer(
                input_channels=output_channels // 4, output_channels=output_channels // 4
            )
        self.l_block4 = LBlock(input_channels=output_channels // 4, output_channels=output_channels)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        """
        Apply convolution, l blocks and spatial attention module to the tensor.

        Args:
            x: tensor on the correct device, to move over the latent distribution

        Returns:
               tensor

        """
        # Independent draws from Norma ldistribution
        z = self.distribution.sample(self.shape)
        # Batch is at end for some reason, reshape
        z = torch.permute(z, (3, 0, 1, 2)).type_as(x)

        # 3x3 Convolution
        z = self.conv_3x3(z)

        # 3 L Blocks to increase number of channels
        z = self.l_block1(z)
        z = self.l_block2(z)
        z = self.l_block3(z)
        # Spatial attention module
        z = self.att_block(z)

        # L block to increase number of channel to 768
        z = self.l_block4(z)
        return z