File size: 15,441 Bytes
cb18693
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Convolutional Neural Operator used by the CNO Navier--Stokes experiment.

The filename is fixed by the reproduction request.  This module implements a
CNO, not a Fourier Neural Operator.  The implementation is written from the
architecture and filter description in Sections 2 and C.1.4 of arXiv:2302.01178.
No source code from the authors' repository is included here.
"""

from __future__ import annotations

from typing import Literal

import torch
from torch import Tensor, nn
import torch.nn.functional as F


def _design_windowed_sinc(
    num_taps: int,
    resample_factor: int,
    cutoff_denominator: float,
    half_width: float,
) -> Tensor:
    """Construct a finite, symmetric low-pass windowed-sinc filter.

    The paper fixes ``N_tap=12``, ``c_h=0.8`` and a cutoff arbitrarily close
    to the target Nyquist frequency, ``s/2.0001``.  At a high-rate grid used
    for factor-r resampling, this corresponds to a normalized cutoff of
    ``1 / (r * 2.0001)`` cycles per sample.  The exact finite window is not
    specified in the paper; a Kaiser window is used as an explicit,
    configurable approximation to ``scipy.signal.firwin``.
    """
    if num_taps < 2:
        raise ValueError(f"num_taps must be at least 2, got {num_taps}")
    if resample_factor < 1:
        raise ValueError("resample_factor must be positive")
    if cutoff_denominator <= 2.0:
        raise ValueError("cutoff_denominator must be greater than 2")
    if half_width <= 0:
        raise ValueError("half_width must be positive")

    dtype = torch.float64
    positions = torch.arange(num_taps, dtype=dtype) - (num_taps - 1) / 2
    cutoff = 1.0 / (resample_factor * cutoff_denominator)
    ideal = 2.0 * cutoff * torch.sinc(2.0 * cutoff * positions)

    # c_h=0.8 maps to a conventional beta=8.6 window.  Keeping the relation
    # explicit makes the paper-unspecified window choice auditable.
    beta = 8.6 * half_width / 0.8
    window = torch.kaiser_window(num_taps, periodic=False, beta=beta, dtype=dtype)
    kernel = ideal * window
    kernel = kernel / kernel.sum()
    return kernel.to(torch.float32)


class FixedSincResample2d(nn.Module):
    """Separable periodic 2-D windowed-sinc up/downsampling."""

    def __init__(
        self,
        factor: int = 2,
        num_taps: int = 12,
        cutoff_denominator: float = 2.0001,
        half_width: float = 0.8,
    ) -> None:
        super().__init__()
        if factor < 1:
            raise ValueError("factor must be positive")
        self.factor = int(factor)
        kernel = _design_windowed_sinc(
            num_taps=num_taps,
            resample_factor=factor,
            cutoff_denominator=cutoff_denominator,
            half_width=half_width,
        )
        self.register_buffer("kernel", kernel, persistent=True)

    def _filter(self, x: Tensor, gain: float = 1.0) -> Tensor:
        if x.ndim != 4:
            raise ValueError(f"expected BCHW input, got shape {tuple(x.shape)}")
        channels = x.shape[1]
        kernel = self.kernel.to(device=x.device, dtype=x.dtype)
        taps = int(kernel.numel())
        pad_left = (taps - 1) // 2
        pad_right = taps - 1 - pad_left

        weight_x = (kernel * gain).view(1, 1, 1, taps).repeat(channels, 1, 1, 1)
        x = F.pad(x, (pad_left, pad_right, 0, 0), mode="circular")
        x = F.conv2d(x, weight_x, groups=channels)

        weight_y = kernel.view(1, 1, taps, 1).repeat(channels, 1, 1, 1)
        x = F.pad(x, (0, 0, pad_left, pad_right), mode="circular")
        return F.conv2d(x, weight_y, groups=channels)

    def upsample(self, x: Tensor) -> Tensor:
        if self.factor == 1:
            return x
        batch, channels, height, width = x.shape
        up = x.new_zeros(batch, channels, height * self.factor, width * self.factor)
        up[..., :: self.factor, :: self.factor] = x
        return self._filter(up, gain=float(self.factor * self.factor))

    def downsample(self, x: Tensor) -> Tensor:
        if self.factor == 1:
            return x
        if x.shape[-2] % self.factor or x.shape[-1] % self.factor:
            raise ValueError(
                f"spatial shape {tuple(x.shape[-2:])} is not divisible by {self.factor}"
            )
        return self._filter(x)[..., :: self.factor, :: self.factor]


class BandlimitedActivation(nn.Module):
    """Paper Eq. (2.6): upsample, activate, then low-pass/downsample."""

    def __init__(
        self,
        upsampling_factor: int = 2,
        num_taps: int = 12,
        cutoff_denominator: float = 2.0001,
        half_width: float = 0.8,
        negative_slope: float = 0.2,
    ) -> None:
        super().__init__()
        self.negative_slope = float(negative_slope)
        self.resampler = FixedSincResample2d(
            factor=upsampling_factor,
            num_taps=num_taps,
            cutoff_denominator=cutoff_denominator,
            half_width=half_width,
        )

    def forward(self, x: Tensor) -> Tensor:
        x = self.resampler.upsample(x)
        x = F.leaky_relu(x, negative_slope=self.negative_slope)
        return self.resampler.downsample(x)


def _periodic_conv(in_channels: int, out_channels: int, kernel_size: int) -> nn.Conv2d:
    if kernel_size % 2 != 1:
        raise ValueError("CNO convolution kernel_size must be odd")
    return nn.Conv2d(
        in_channels,
        out_channels,
        kernel_size=kernel_size,
        padding=kernel_size // 2,
        padding_mode="circular",
    )


class CNOBlock(nn.Module):
    """Physical-space convolution followed by bandlimited activation/resampling."""

    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        kernel_size: int,
        resample: Literal["same", "up", "down"],
        activation_kwargs: dict,
        batch_norm: bool = True,
    ) -> None:
        super().__init__()
        self.resample = resample
        self.conv = _periodic_conv(in_channels, out_channels, kernel_size)
        self.norm = nn.BatchNorm2d(out_channels) if batch_norm else nn.Identity()
        self.activation = BandlimitedActivation(**activation_kwargs)
        self.resampler = FixedSincResample2d(
            factor=activation_kwargs["upsampling_factor"],
            num_taps=activation_kwargs["num_taps"],
            cutoff_denominator=activation_kwargs["cutoff_denominator"],
            half_width=activation_kwargs["half_width"],
        )

    def forward(self, x: Tensor) -> Tensor:
        x = self.activation(self.norm(self.conv(x)))
        if self.resample == "down":
            return self.resampler.downsample(x)
        if self.resample == "up":
            return self.resampler.upsample(x)
        return x


class ResidualBlock(nn.Module):
    """Paper Eq. (2.7): identity plus K o Sigma o K."""

    def __init__(self, channels: int, kernel_size: int, activation_kwargs: dict) -> None:
        super().__init__()
        self.conv1 = _periodic_conv(channels, channels, kernel_size)
        self.norm1 = nn.BatchNorm2d(channels)
        self.activation = BandlimitedActivation(**activation_kwargs)
        self.conv2 = _periodic_conv(channels, channels, kernel_size)
        self.norm2 = nn.BatchNorm2d(channels)

    def forward(self, x: Tensor) -> Tensor:
        residual = self.norm1(self.conv1(x))
        residual = self.activation(residual)
        residual = self.norm2(self.conv2(residual))
        return x + residual


class InvariantBlock(nn.Module):
    """Paper Eq. (2.8): Sigma o K at an unchanged representation."""

    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        kernel_size: int,
        activation_kwargs: dict,
    ) -> None:
        super().__init__()
        self.conv = _periodic_conv(in_channels, out_channels, kernel_size)
        self.norm = nn.BatchNorm2d(out_channels)
        self.activation = BandlimitedActivation(**activation_kwargs)

    def forward(self, x: Tensor) -> Tensor:
        return self.activation(self.norm(self.conv(x)))


class LiftProjectBlock(nn.Module):
    """Two convolutions with no BatchNorm, as specified for lift/project."""

    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        latent_channels: int,
        kernel_size: int,
        activation_kwargs: dict,
    ) -> None:
        super().__init__()
        self.conv1 = _periodic_conv(in_channels, latent_channels, kernel_size)
        self.activation = BandlimitedActivation(**activation_kwargs)
        self.conv2 = _periodic_conv(latent_channels, out_channels, kernel_size)

    def forward(self, x: Tensor) -> Tensor:
        return self.conv2(self.activation(self.conv1(x)))


def _residual_stack(
    channels: int,
    count: int,
    kernel_size: int,
    activation_kwargs: dict,
) -> nn.Module:
    if count == 0:
        return nn.Identity()
    return nn.Sequential(
        *[
            ResidualBlock(channels, kernel_size, activation_kwargs)
            for _ in range(count)
        ]
    )


class CNO2d(nn.Module):
    """Operator U-Net CNO for the paper's 2-D Navier--Stokes experiment."""

    def __init__(
        self,
        in_channels: int = 1,
        out_channels: int = 1,
        base_width: int = 32,
        levels: int = 3,
        bottleneck_residual_blocks: int = 8,
        intermediate_residual_blocks: int = 1,
        kernel_size: int = 3,
        latent_channels: int = 64,
        activation_upsampling_factor: int = 2,
        filter_taps: int = 12,
        filter_half_width: float = 0.8,
        cutoff_denominator: float = 2.0001,
        leaky_relu_slope: float = 0.2,
    ) -> None:
        super().__init__()
        if levels < 1:
            raise ValueError("levels must be positive")
        if base_width % 2:
            raise ValueError("base_width must be even because lift width is d_e/2")

        self.in_channels = int(in_channels)
        self.out_channels = int(out_channels)
        self.base_width = int(base_width)
        self.levels = int(levels)
        self.required_divisor = 2**levels
        lift_width = base_width // 2
        encoder_widths = [base_width * (2**index) for index in range(levels)]
        activation_kwargs = {
            "upsampling_factor": activation_upsampling_factor,
            "num_taps": filter_taps,
            "cutoff_denominator": cutoff_denominator,
            "half_width": filter_half_width,
            "negative_slope": leaky_relu_slope,
        }

        self.lift = LiftProjectBlock(
            in_channels,
            lift_width,
            latent_channels,
            kernel_size,
            activation_kwargs,
        )

        down_blocks: list[nn.Module] = []
        intermediate_blocks: list[nn.Module] = []
        current_width = lift_width
        for level, next_width in enumerate(encoder_widths):
            down_blocks.append(
                CNOBlock(
                    current_width,
                    next_width,
                    kernel_size,
                    "down",
                    activation_kwargs,
                )
            )
            # The bottleneck has its own N_res,b stack; N_res,i belongs to the
            # genuinely intermediate resolutions only.
            count = intermediate_residual_blocks if level < levels - 1 else 0
            intermediate_blocks.append(
                _residual_stack(next_width, count, kernel_size, activation_kwargs)
            )
            current_width = next_width
        self.encoder = nn.ModuleList(down_blocks)
        self.encoder_residuals = nn.ModuleList(intermediate_blocks)

        self.bottleneck = _residual_stack(
            encoder_widths[-1],
            bottleneck_residual_blocks,
            kernel_size,
            activation_kwargs,
        )

        pre_patch: list[nn.Module] = []
        post_patch: list[nn.Module] = []
        up_blocks: list[nn.Module] = []
        decoder_current = encoder_widths[-1]
        output_widths = list(reversed([lift_width] + encoder_widths[:-1]))
        for next_width in output_widths:
            pre_patch.append(
                InvariantBlock(
                    decoder_current,
                    decoder_current,
                    kernel_size,
                    activation_kwargs,
                )
            )
            post_patch.append(
                InvariantBlock(
                    decoder_current * 2,
                    decoder_current,
                    kernel_size,
                    activation_kwargs,
                )
            )
            up_blocks.append(
                CNOBlock(
                    decoder_current,
                    next_width,
                    kernel_size,
                    "up",
                    activation_kwargs,
                )
            )
            decoder_current = next_width
        self.decoder_pre_patch = nn.ModuleList(pre_patch)
        self.decoder_post_patch = nn.ModuleList(post_patch)
        self.decoder = nn.ModuleList(up_blocks)

        self.project = LiftProjectBlock(
            lift_width * 2,
            out_channels,
            latent_channels,
            kernel_size,
            activation_kwargs,
        )

    def forward(self, x: Tensor) -> Tensor:
        if x.ndim != 4:
            raise ValueError(f"CNO2d expects BCHW input, got {tuple(x.shape)}")
        if x.shape[1] != self.in_channels:
            raise ValueError(
                f"expected {self.in_channels} channels, got {x.shape[1]}"
            )
        height, width = x.shape[-2:]
        if height % self.required_divisor or width % self.required_divisor:
            raise ValueError(
                f"spatial shape {(height, width)} must be divisible by "
                f"2**levels={self.required_divisor}"
            )

        lifted = self.lift(x)
        encoded = lifted
        skips: list[Tensor] = []
        for down, residuals in zip(self.encoder, self.encoder_residuals):
            encoded = residuals(down(encoded))
            skips.append(encoded)

        decoded = self.bottleneck(encoded)
        for pre, post, up, skip in zip(
            self.decoder_pre_patch,
            self.decoder_post_patch,
            self.decoder,
            reversed(skips),
        ):
            decoded = pre(decoded)
            if decoded.shape[-2:] != skip.shape[-2:]:
                raise RuntimeError(
                    "decoder/skip spatial mismatch before patching: "
                    f"{tuple(decoded.shape)} versus {tuple(skip.shape)}"
                )
            decoded = post(torch.cat((decoded, skip), dim=1))
            decoded = up(decoded)

        if decoded.shape[-2:] != lifted.shape[-2:]:
            raise RuntimeError(
                f"final decoder/lift mismatch: {decoded.shape} versus {lifted.shape}"
            )
        output = self.project(torch.cat((decoded, lifted), dim=1))
        if output.shape[-2:] != (height, width):
            raise RuntimeError(
                f"CNO changed output grid from {(height, width)} to {output.shape[-2:]}"
            )
        return output


def build_model(model_config: dict) -> CNO2d:
    """Build a CNO2d from the ``model`` section of config.yaml."""
    return CNO2d(**model_config)


def count_trainable_parameters(model: nn.Module) -> int:
    return sum(parameter.numel() for parameter in model.parameters() if parameter.requires_grad)


__all__ = ["CNO2d", "build_model", "count_trainable_parameters"]