Skip to content

Reversion

iris.reversion

Trainable neural networks for imagery reversion.

Reversion is defined as a constrained inverse to observation of a physical tensor. The initial experiments detailed in the IRIS paper do not yet reconstruct the entire physical tensor from the entire observed PPV cube. Rather, a top-down density image is reconstructed from a latitude-reduced observation. In particular, the IRIS paper considers mean-reductions over the latitude dimension of the cube--yielding longitude-velocity PV images-- although max reductions are also allowed by the code.

Reversion is accomplished by a neural network trained on a dataset. This module provides the neural network architecture used in the IRIS paper. The core design is a convolutional neural network (CNN) with pixelwise attention, structured as an encoder-decoder. The encoder maps a reduced observation into a latent featural space. The decoder maps the latent featural object to a top-down density image. The entire neural network has about ~14M trainable parameters.

Authors

B.L. DuBois (brendan@bldubois.com)

Reverter

Bases: Module

A trainable neural network for imagery reversion.

Reverts either a mean- or max-reduced PPV cube in a single spectral line to a top-down density image. The core architecture of Reverter is an encoder-decoder convolutional neural network (CNN) with pixelwise self-attention, which we describe in detail in the IRIS paper (subsec: Implementation of Reversion: Architecture).

For convenience, the module expects a full PPV cube and automatically applies the hyper.cube_hyper.reduction specified in hyperparameters, unless the keyword arg reduce=False is passed to the module forward method, in which case the expected input is a reduced PV observation. Use reduce=False when applying to a PreObservedDataset. The Reverter also keeps track of its training physical units, which are saved as non-trainable parameters in its model state dict. A utility multi_unit_call automatically performs the unit conversions necessary when applied to an input or output space in different physical units than those on which the Reverter is trained.

Attributes:

Name Type Description
temperature Parameter

The brightness temperature units of the Reverter, as a conversion factor from K.
intensity Parameter

The intensity units of the Reverter, as a conversion factor from \(\text{kg}/\text{s}^2\).
v_density Parameter

The velocity-density units of the Reverter, as a conversion factor from \(\text{kg} \cdot \text{s}/\text{m}^3\), for use if training on simple observations. See also SimplyObservedDataset.
density Parameter

The density units of the Reverter, as a conversion factor from \(\text{kg}/\text{m}^3\).
reduction Callable[[Tensor], Tensor]

The latitude reduction performed on an input PPV cube. Either a mean or max reduction. Set by hyper.cube_hyper.reduction.
encoder Encoder

The encoder CNN.
decoder Decoder

The decoder CNN.

Parameters:

Name Type Description Default
hyper Hyper

A hyperparameters object.

 required
units_hyper Hyper | None

An optional separate hyperparameters object from which to adopt units but not other configurations. If None, units are taken from hyper.

 None

Raises:

Type Description
ValueError

If hyper.cube_hyper.reduction is not one of 'mean', 'max'.
Source code in iris/reversion.py 
 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
class Reverter(torch.nn.Module):
    r"""
    A trainable neural network for imagery reversion.

    Reverts either a mean- or max-reduced PPV cube in a single spectral line to a
    [top-down density image][iris.arepo_processing.columnize_physical_tensor].
    The core architecture of `Reverter` is an encoder-decoder convolutional neural network (CNN)
    with pixelwise self-attention, which we describe in detail in the IRIS paper
    (subsec: Implementation of Reversion: Architecture).

    For convenience,
    the module expects a full PPV cube and automatically applies the
    `hyper.cube_hyper.reduction` specified in hyperparameters, unless the
    keyword arg `reduce=False` is passed to the module [`forward`][iris.reversion.Reverter.forward]
    method, in which case the expected input is a reduced PV observation. Use `reduce=False` when
    applying to a [`PreObservedDataset`][iris.arepo_processing.PreObservedDataset]. The `Reverter`
    also keeps track of its training physical units, which are saved as non-trainable parameters
    in its model state dict. A utility
    [`multi_unit_call`][iris.reversion.Reverter.multi_unit_call] automatically performs the
    unit conversions necessary when applied to an input or output space in different physical
    units than those on which the `Reverter` is trained.

    Attributes:
        temperature: The brightness temperature units of the `Reverter`, as a conversion factor from K.
        intensity: The intensity units of the `Reverter`, as a conversion factor from $\text{kg}/\text{s}^2$.
        v_density: The velocity-density units of the `Reverter`, as a conversion factor from
            $\text{kg} \cdot \text{s}/\text{m}^3$, for use if training on
            [simple observations][iris.observation.SimpleObserver]. See also
            [`SimplyObservedDataset`][iris.arepo_processing.SimplyObservedDataset].
        density: The density units of the `Reverter`, as a conversion factor from $\text{kg}/\text{m}^3$.
        reduction: The latitude reduction performed on an input PPV cube. Either a mean or max reduction.
            Set by `hyper.cube_hyper.reduction`.
        encoder: The encoder CNN.
        decoder: The decoder CNN.

    Args:
        hyper: A hyperparameters object.
        units_hyper: An optional separate hyperparameters object from which to adopt units but not
            other configurations. If `None`, units are taken from `hyper`.

    Raises:
        ValueError: If `hyper.cube_hyper.reduction` is not one of `'mean', 'max'`.
    """

    temperature: torch.nn.Parameter
    intensity: torch.nn.Parameter
    v_density: torch.nn.Parameter
    density: torch.nn.Parameter
    reduction: typing.Callable[[torch.Tensor], torch.Tensor]
    encoder: Encoder
    decoder: Decoder

    def __init__(self, hyper: hp.Hyper, units_hyper: hp.Hyper | None = None) -> None:
        super().__init__()
        if units_hyper is None:
            units_hyper = hyper
        temperature = units_hyper.dataset_hyper._temperature_iris_per_SI
        if temperature is None:
            temperature = 1.0
        length = units_hyper.dataset_hyper._length_iris_per_SI
        if length is None:
            length = 1.0
        mass = units_hyper.dataset_hyper._mass_iris_per_SI
        if mass is None:
            mass = 1.0
        time = units_hyper.dataset_hyper._time_iris_per_SI
        if time is None:
            time = 1.0
        density = mass / length / length / length
        intensity = mass / time / time
        v_density = density * time
        self.temperature = torch.nn.Parameter(
            torch.tensor(temperature, dtype=torch.float32), requires_grad=False)
        self.intensity = torch.nn.Parameter(
            torch.tensor(intensity, dtype=torch.float32), requires_grad=False)
        self.v_density = torch.nn.Parameter(
            torch.tensor(v_density, dtype=torch.float32), requires_grad=False)
        self.density = torch.nn.Parameter(
            torch.tensor(density, dtype=torch.float32), requires_grad=False)

        reduction = hyper.cube_hyper.reduction
        if reduction == 'mean':
            self.reduction = self._reduce_mean
        elif reduction == 'max':
            self.reduction = self._reduce_max
        else:
            raise ValueError("Cube reduction must be one of: 'mean', 'max'.")
        self.encoder = Encoder()
        self.decoder = Decoder()
        return

    def forward(self, inputs: torch.Tensor, reduce: bool = True) -> torch.Tensor:
        """
        The `Reverter` forward pass.

        If `reduce`, applies `self.reduction`. Then applies `self.encoder` and `self.decoder`.

        Args:
            inputs: The input observations. A batch of either full PPV cubes or latitude-reduced PV images.
            reduce: If `True`, applies `self.reduction`.

        Returns:
            A batch of [top-down density images][iris.arepo_processing.columnize_physical_tensor].
        """
        if reduce:
            inputs = self.reduction(inputs)
        latent = self.encoder(inputs)
        outputs = self.decoder(latent)
        return outputs

    def _reduce_mean(self, cube: torch.Tensor) -> torch.Tensor:
        """
        Performs a mean-reduction over the latitude dimension of a PPV cube.

        Args:
            cube: The PPV cube to reduce.

        Returns:
            The latitude-meaned PV image.
        """
        return torch.mean(cube, dim=3)

    def _reduce_max(self, cube: torch.Tensor) -> torch.Tensor:
        """
        Performs a max-reduction over the latitude dimension of a PPV cube.

        Args:
            cube: The PPV cube to reduce.

        Returns:
            The latitude-reduced peak-value image (PV).
        """
        return torch.max(cube, dim=3)[0]

    def multi_unit_call(self,
                        inputs: torch.Tensor,
                        in_units: hp.Hyper,
                        out_units: hp.Hyper,
                        in_space: str = 'T',
                        reduce: bool = True) -> torch.Tensor:
        """
        Wraps the model forward call in input and output unit conversions.

        Used when applying the `Reverter` to an input or output space in different physical
        units than those on which it was trained.

        Args:
            inputs: The input observations. A batch of either full PPV cubes or latitude-reduced PV images.
            in_units: A hyperparameters object specifying the units of the input space.
            out_units: A hyperparameters object specifying the units of the output space.
            in_space: The inputs space. If processing a
                [synthetic observation][iris.observation.SyntheticObserver], options are
                temperature (brightness or Raleigh-Jeans, specify `'T'`) or intensity (specify `'I'`).
                If processing a
                [simple observation][iris.observation.SimpleObserver],
                specify velocity-density (`'vrho'`).
            reduce: If `True`, applies `self.reduction`.

        Returns:
            A batch of [top-down density images][iris.arepo_processing.columnize_physical_tensor].

        Raises:
            ValueError: If `in_space` is not one of `'T', 'I', 'vrho'`.
        """
        if in_space == 'T':
            in_temperature = in_units.dataset_hyper._temperature_iris_per_SI
            inputs = inputs * (self.temperature / in_temperature)
        elif in_space == 'I':
            in_time = in_units.dataset_hyper._time_iris_per_SI
            in_mass = in_units.dataset_hyper._mass_iris_per_SI
            in_intensity = in_mass / in_time / in_time
            inputs = inputs * (self.intensity / in_intensity)
        elif in_space == 'vrho':
            in_length = in_units.dataset_hyper._length_iris_per_SI
            in_time = in_units.dataset_hyper._time_iris_per_SI
            in_mass = in_units.dataset_hyper._mass_iris_per_SI
            in_v_density = in_mass * in_time / in_length / in_length / in_length
            inputs = inputs * (self.v_density / in_v_density)
        else:
            raise ValueError("Arg in_space must be one of 'T', 'I', 'vrho'.")

        out_length = out_units.dataset_hyper._length_iris_per_SI
        out_mass = out_units.dataset_hyper._mass_iris_per_SI
        out_density = out_mass / out_length / out_length / out_length

        outputs = self(inputs, reduce=reduce)
        outputs = outputs * (out_density / self.density)
        return outputs

forward(inputs, reduce=True)

The Reverter forward pass.

If reduce, applies self.reduction. Then applies self.encoder and self.decoder.

Parameters:

Name Type Description Default
inputs Tensor

The input observations. A batch of either full PPV cubes or latitude-reduced PV images.

 required
reduce bool

If True, applies self.reduction.

 True

Returns:

Type Description
Tensor

A batch of top-down density images.
Source code in iris/reversion.py 
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
def forward(self, inputs: torch.Tensor, reduce: bool = True) -> torch.Tensor:
    """
    The `Reverter` forward pass.

    If `reduce`, applies `self.reduction`. Then applies `self.encoder` and `self.decoder`.

    Args:
        inputs: The input observations. A batch of either full PPV cubes or latitude-reduced PV images.
        reduce: If `True`, applies `self.reduction`.

    Returns:
        A batch of [top-down density images][iris.arepo_processing.columnize_physical_tensor].
    """
    if reduce:
        inputs = self.reduction(inputs)
    latent = self.encoder(inputs)
    outputs = self.decoder(latent)
    return outputs

_reduce_mean(cube)

Performs a mean-reduction over the latitude dimension of a PPV cube.

Parameters:

Name Type Description Default
cube Tensor

The PPV cube to reduce.

 required

Returns:

Type Description
Tensor

The latitude-meaned PV image.
Source code in iris/reversion.py 
146
147
148
149
150
151
152
153
154
155
156
def _reduce_mean(self, cube: torch.Tensor) -> torch.Tensor:
    """
    Performs a mean-reduction over the latitude dimension of a PPV cube.

    Args:
        cube: The PPV cube to reduce.

    Returns:
        The latitude-meaned PV image.
    """
    return torch.mean(cube, dim=3)

_reduce_max(cube)

Performs a max-reduction over the latitude dimension of a PPV cube.

Parameters:

Name Type Description Default
cube Tensor

The PPV cube to reduce.

 required

Returns:

Type Description
Tensor

The latitude-reduced peak-value image (PV).
Source code in iris/reversion.py 
158
159
160
161
162
163
164
165
166
167
168
def _reduce_max(self, cube: torch.Tensor) -> torch.Tensor:
    """
    Performs a max-reduction over the latitude dimension of a PPV cube.

    Args:
        cube: The PPV cube to reduce.

    Returns:
        The latitude-reduced peak-value image (PV).
    """
    return torch.max(cube, dim=3)[0]

multi_unit_call(inputs, in_units, out_units, in_space='T', reduce=True)

Wraps the model forward call in input and output unit conversions.

Used when applying the Reverter to an input or output space in different physical units than those on which it was trained.

Parameters:

Name Type Description Default
inputs Tensor

The input observations. A batch of either full PPV cubes or latitude-reduced PV images.

 required
in_units Hyper

A hyperparameters object specifying the units of the input space.

 required
out_units Hyper

A hyperparameters object specifying the units of the output space.

 required
in_space str

The inputs space. If processing a synthetic observation, options are temperature (brightness or Raleigh-Jeans, specify 'T') or intensity (specify 'I'). If processing a simple observation, specify velocity-density ('vrho').

 'T'
reduce bool

If True, applies self.reduction.

 True

Returns:

Type Description
Tensor

A batch of top-down density images.

Raises:

Type Description
ValueError

If in_space is not one of 'T', 'I', 'vrho'.
Source code in iris/reversion.py 
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
def multi_unit_call(self,
                    inputs: torch.Tensor,
                    in_units: hp.Hyper,
                    out_units: hp.Hyper,
                    in_space: str = 'T',
                    reduce: bool = True) -> torch.Tensor:
    """
    Wraps the model forward call in input and output unit conversions.

    Used when applying the `Reverter` to an input or output space in different physical
    units than those on which it was trained.

    Args:
        inputs: The input observations. A batch of either full PPV cubes or latitude-reduced PV images.
        in_units: A hyperparameters object specifying the units of the input space.
        out_units: A hyperparameters object specifying the units of the output space.
        in_space: The inputs space. If processing a
            [synthetic observation][iris.observation.SyntheticObserver], options are
            temperature (brightness or Raleigh-Jeans, specify `'T'`) or intensity (specify `'I'`).
            If processing a
            [simple observation][iris.observation.SimpleObserver],
            specify velocity-density (`'vrho'`).
        reduce: If `True`, applies `self.reduction`.

    Returns:
        A batch of [top-down density images][iris.arepo_processing.columnize_physical_tensor].

    Raises:
        ValueError: If `in_space` is not one of `'T', 'I', 'vrho'`.
    """
    if in_space == 'T':
        in_temperature = in_units.dataset_hyper._temperature_iris_per_SI
        inputs = inputs * (self.temperature / in_temperature)
    elif in_space == 'I':
        in_time = in_units.dataset_hyper._time_iris_per_SI
        in_mass = in_units.dataset_hyper._mass_iris_per_SI
        in_intensity = in_mass / in_time / in_time
        inputs = inputs * (self.intensity / in_intensity)
    elif in_space == 'vrho':
        in_length = in_units.dataset_hyper._length_iris_per_SI
        in_time = in_units.dataset_hyper._time_iris_per_SI
        in_mass = in_units.dataset_hyper._mass_iris_per_SI
        in_v_density = in_mass * in_time / in_length / in_length / in_length
        inputs = inputs * (self.v_density / in_v_density)
    else:
        raise ValueError("Arg in_space must be one of 'T', 'I', 'vrho'.")

    out_length = out_units.dataset_hyper._length_iris_per_SI
    out_mass = out_units.dataset_hyper._mass_iris_per_SI
    out_density = out_mass / out_length / out_length / out_length

    outputs = self(inputs, reduce=reduce)
    outputs = outputs * (out_density / self.density)
    return outputs

Encoder

Bases: Module

The encoder module used by Reverter.

Maps an input PV image of dimensions batch, channel=1, lon=512, v=512 into a latent featural space of dimensions batch, channel=2048, lon=1, v=1, implying a total size reduction factor of \(512 \cdot 512 / 2048 = 128\). See the IRIS paper for architectural details and discussion (subsec: Implementation of Reversion: Architecture).

Attributes:

Name Type Description
_1_1_convolution Conv2d

A convolution with in_channels=1, out_channels=4, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1), groups=1, bias=True, dtype=torch.float32.
_1_2_batch_norm BatchNorm2d

A batch normalization.
_1_3_leaky_relu LeakyReLU

A leaky ReLU.
_2_1_convolution Conv2d

A convolution with in_channels=4, out_channels=16, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1), groups=1, bias=True, dtype=torch.float32.
_2_2_batch_norm BatchNorm2d

A batch normalization.
_2_3_leaky_relu LeakyReLU

A leaky ReLU.
_3_1_convolution Conv2d

A convolution with in_channels=16, out_channels=32, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1), groups=1, bias=True, dtype=torch.float32.
_3_2_batch_norm BatchNorm2d

A batch normalization.
_3_3_leaky_relu LeakyReLU

A leaky ReLU.
_4_1_convolution Conv2d

A convolution with in_channels=32, out_channels=64, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1), groups=1, bias=True, dtype=torch.float32.
_4_2_batch_norm BatchNorm2d

A batch normalization.
_4_3_leaky_relu LeakyReLU

A leaky ReLU.
_4_4_attention PixelSelfAttention2d

A pixelwise attention layer.
_5_1_convolution Conv2d

A convolution with in_channels=64, out_channels=128, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1), groups=1, bias=True, dtype=torch.float32.
_5_2_batch_norm BatchNorm2d

A batch normalization.
_5_3_leaky_relu LeakyReLU

A leaky ReLU.
_5_4_attention PixelSelfAttention2d

A pixelwise attention layer.
_6_1_convolution Conv2d

A convolution with in_channels=128, out_channels=256, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1), groups=1, bias=True, dtype=torch.float32.
_6_2_batch_norm BatchNorm2d

A batch normalization.
_6_3_leaky_relu LeakyReLU

A leaky ReLU.
_6_4_attention PixelSelfAttention2d

A pixelwise attention layer.
_7_1_convolution Conv2d

A convolution with in_channels=256, out_channels=512, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1), groups=1, bias=True, dtype=torch.float32.
_7_2_batch_norm BatchNorm2d

A batch normalization.
_7_3_leaky_relu LeakyReLU

A leaky ReLU.
_8_1_convolution Conv2d

A convolution with in_channels=512, out_channels=1024, kernel_size=(2, 2), stride=(2, 2), padding=(0, 0), groups=1, bias=True, dtype=torch.float32.
_8_2_batch_norm BatchNorm2d

A batch normalization.
_8_3_leaky_relu LeakyReLU

A leaky ReLU.
_9_1_convolution Conv2d

A convolution with in_channels=1024, out_channels=2048, kernel_size=(2, 2), stride=(1, 1), padding=(0, 0), groups=4, bias=True, dtype=torch.float32.
_9_2_leaky_relu LeakyReLU

A batch normalization.
Source code in iris/reversion.py 
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
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
class Encoder(torch.nn.Module):
    r"""
    The encoder module used by [`Reverter`][iris.reversion.Reverter].

    Maps an input PV image of dimensions `batch, channel=1, lon=512, v=512` into a latent featural
    space of dimensions `batch, channel=2048, lon=1, v=1`, implying a total size reduction factor
    of $512 \cdot 512 / 2048 = 128$. See the IRIS paper for architectural details and discussion
    (subsec: Implementation of Reversion: Architecture).

    Attributes:
        _1_1_convolution: A convolution with `in_channels=1,
                                              out_channels=4,
                                              kernel_size=(4, 4),
                                              stride=(2, 2),
                                              padding=(1, 1),
                                              groups=1,
                                              bias=True`,
                                              dtype=torch.float32.
        _1_2_batch_norm: A batch normalization.
        _1_3_leaky_relu: A leaky ReLU.

        _2_1_convolution: A convolution with `in_channels=4,
                                              out_channels=16,
                                              kernel_size=(4, 4),
                                              stride=(2, 2),
                                              padding=(1, 1),
                                              groups=1,
                                              bias=True`,
                                              dtype=torch.float32.
        _2_2_batch_norm: A batch normalization.
        _2_3_leaky_relu: A leaky ReLU.

        _3_1_convolution: A convolution with `in_channels=16,
                                              out_channels=32,
                                              kernel_size=(4, 4),
                                              stride=(2, 2),
                                              padding=(1, 1),
                                              groups=1,
                                              bias=True`,
                                              dtype=torch.float32.
        _3_2_batch_norm: A batch normalization.
        _3_3_leaky_relu: A leaky ReLU.

        _4_1_convolution: A convolution with `in_channels=32,
                                              out_channels=64,
                                              kernel_size=(4, 4),
                                              stride=(2, 2),
                                              padding=(1, 1),
                                              groups=1,
                                              bias=True`,
                                              dtype=torch.float32.
        _4_2_batch_norm: A batch normalization.
        _4_3_leaky_relu: A leaky ReLU.
        _4_4_attention: A pixelwise attention layer.

        _5_1_convolution: A convolution with `in_channels=64,
                                              out_channels=128,
                                              kernel_size=(4, 4),
                                              stride=(2, 2),
                                              padding=(1, 1),
                                              groups=1,
                                              bias=True`,
                                              dtype=torch.float32.
        _5_2_batch_norm: A batch normalization.
        _5_3_leaky_relu: A leaky ReLU.
        _5_4_attention: A pixelwise attention layer.

        _6_1_convolution: A convolution with `in_channels=128,
                                              out_channels=256,
                                              kernel_size=(4, 4),
                                              stride=(2, 2),
                                              padding=(1, 1),
                                              groups=1,
                                              bias=True`,
                                              dtype=torch.float32.
        _6_2_batch_norm: A batch normalization.
        _6_3_leaky_relu: A leaky ReLU.
        _6_4_attention: A pixelwise attention layer.

        _7_1_convolution: A convolution with `in_channels=256,
                                              out_channels=512,
                                              kernel_size=(4, 4),
                                              stride=(2, 2),
                                              padding=(1, 1),
                                              groups=1,
                                              bias=True`,
                                              dtype=torch.float32.
        _7_2_batch_norm: A batch normalization.
        _7_3_leaky_relu: A leaky ReLU.

        _8_1_convolution: A convolution with `in_channels=512,
                                              out_channels=1024,
                                              kernel_size=(2, 2),
                                              stride=(2, 2),
                                              padding=(0, 0),
                                              groups=1,
                                              bias=True`,
                                              dtype=torch.float32.
        _8_2_batch_norm: A batch normalization.
        _8_3_leaky_relu: A leaky ReLU.

        _9_1_convolution: A convolution with `in_channels=1024,
                                              out_channels=2048,
                                              kernel_size=(2, 2),
                                              stride=(1, 1),
                                              padding=(0, 0),
                                              groups=4,
                                              bias=True`,
                                              dtype=torch.float32.
        _9_2_leaky_relu: A batch normalization.
    """

    _1_1_convolution: torch.nn.Conv2d
    _1_2_batch_norm: torch.nn.BatchNorm2d
    _1_3_leaky_relu: torch.nn.LeakyReLU

    _2_1_convolution: torch.nn.Conv2d
    _2_2_batch_norm: torch.nn.BatchNorm2d
    _2_3_leaky_relu: torch.nn.LeakyReLU

    _3_1_convolution: torch.nn.Conv2d
    _3_2_batch_norm: torch.nn.BatchNorm2d
    _3_3_leaky_relu: torch.nn.LeakyReLU

    _4_1_convolution: torch.nn.Conv2d
    _4_2_batch_norm: torch.nn.BatchNorm2d
    _4_3_leaky_relu: torch.nn.LeakyReLU
    _4_4_attention: PixelSelfAttention2d

    _5_1_convolution: torch.nn.Conv2d
    _5_2_batch_norm: torch.nn.BatchNorm2d
    _5_3_leaky_relu: torch.nn.LeakyReLU
    _5_4_attention: PixelSelfAttention2d

    _6_1_convolution: torch.nn.Conv2d
    _6_2_batch_norm: torch.nn.BatchNorm2d
    _6_3_leaky_relu: torch.nn.LeakyReLU
    _6_4_attention: PixelSelfAttention2d

    _7_1_convolution: torch.nn.Conv2d
    _7_2_batch_norm: torch.nn.BatchNorm2d
    _7_3_leaky_relu: torch.nn.LeakyReLU

    _8_1_convolution: torch.nn.Conv2d
    _8_2_batch_norm: torch.nn.BatchNorm2d
    _8_3_leaky_relu: torch.nn.LeakyReLU

    _9_1_convolution: torch.nn.Conv2d
    _9_2_leaky_relu: torch.nn.LeakyReLU

    def __init__(self) -> None:
        super().__init__()
        self._1_1_convolution = torch.nn.Conv2d(in_channels=1,
                                                out_channels=4,
                                                kernel_size=(4, 4),
                                                stride=(2, 2),
                                                padding=(1, 1),
                                                groups=1,
                                                bias=True,
                                                dtype=torch.float32)
        torch.nn.init.kaiming_uniform_(self._1_1_convolution.weight,
                                       a=0.01,
                                       nonlinearity='leaky_relu')
        self._1_2_batch_norm = torch.nn.BatchNorm2d(4)
        self._1_3_leaky_relu = torch.nn.LeakyReLU(negative_slope=0.01)

        self._2_1_convolution = torch.nn.Conv2d(in_channels=4,
                                                out_channels=16,
                                                kernel_size=(4, 4),
                                                stride=(2, 2),
                                                padding=(1, 1),
                                                groups=1,
                                                bias=True,
                                                dtype=torch.float32)
        torch.nn.init.kaiming_uniform_(self._2_1_convolution.weight,
                                       a=0.01,
                                       nonlinearity='leaky_relu')
        self._2_2_batch_norm = torch.nn.BatchNorm2d(16)
        self._2_3_leaky_relu = torch.nn.LeakyReLU(negative_slope=0.01)

        self._3_1_convolution = torch.nn.Conv2d(in_channels=16,
                                                out_channels=32,
                                                kernel_size=(4, 4),
                                                stride=(2, 2),
                                                padding=(1, 1),
                                                groups=1,
                                                bias=True,
                                                dtype=torch.float32)
        torch.nn.init.kaiming_uniform_(self._3_1_convolution.weight,
                                       a=0.01,
                                       nonlinearity='leaky_relu')
        self._3_2_batch_norm = torch.nn.BatchNorm2d(32)
        self._3_3_leaky_relu = torch.nn.LeakyReLU(negative_slope=0.01)

        self._4_1_convolution = torch.nn.Conv2d(in_channels=32,
                                                out_channels=64,
                                                kernel_size=(4, 4),
                                                stride=(2, 2),
                                                padding=(1, 1),
                                                groups=1,
                                                bias=True,
                                                dtype=torch.float32)
        torch.nn.init.kaiming_uniform_(self._4_1_convolution.weight,
                                       a=0.01,
                                       nonlinearity='leaky_relu')
        self._4_2_batch_norm = torch.nn.BatchNorm2d(64)
        self._4_3_leaky_relu = torch.nn.LeakyReLU(negative_slope=0.01)
        self._4_4_attention = PixelSelfAttention2d(channels=64, num_heads=2)

        self._5_1_convolution = torch.nn.Conv2d(in_channels=64,
                                                out_channels=128,
                                                kernel_size=(4, 4),
                                                stride=(2, 2),
                                                padding=(1, 1),
                                                groups=1,
                                                bias=True,
                                                dtype=torch.float32)
        torch.nn.init.kaiming_uniform_(self._5_1_convolution.weight,
                                       a=0.01,
                                       nonlinearity='leaky_relu')
        self._5_2_batch_norm = torch.nn.BatchNorm2d(128)
        self._5_3_leaky_relu = torch.nn.LeakyReLU(negative_slope=0.01)
        self._5_4_attention = PixelSelfAttention2d(channels=128, num_heads=4)

        self._6_1_convolution = torch.nn.Conv2d(in_channels=128,
                                                out_channels=256,
                                                kernel_size=(4, 4),
                                                stride=(2, 2),
                                                padding=(1, 1),
                                                groups=1,
                                                bias=True,
                                                dtype=torch.float32)
        torch.nn.init.kaiming_uniform_(self._6_1_convolution.weight,
                                       a=0.01,
                                       nonlinearity='leaky_relu')
        self._6_2_batch_norm = torch.nn.BatchNorm2d(256)
        self._6_3_leaky_relu = torch.nn.LeakyReLU(negative_slope=0.01)
        self._6_4_attention = PixelSelfAttention2d(channels=256, num_heads=8)

        self._7_1_convolution = torch.nn.Conv2d(in_channels=256,
                                                out_channels=512,
                                                kernel_size=(4, 4),
                                                stride=(2, 2),
                                                padding=(1, 1),
                                                groups=1,
                                                bias=True,
                                                dtype=torch.float32)
        torch.nn.init.kaiming_uniform_(self._7_1_convolution.weight,
                                       a=0.01,
                                       nonlinearity='leaky_relu')
        self._7_2_batch_norm = torch.nn.BatchNorm2d(512)
        self._7_3_leaky_relu = torch.nn.LeakyReLU(negative_slope=0.01)

        self._8_1_convolution = torch.nn.Conv2d(in_channels=512,
                                                out_channels=1024,
                                                kernel_size=(2, 2),
                                                stride=(2, 2),
                                                padding=(0, 0),
                                                groups=1,
                                                bias=True,
                                                dtype=torch.float32)
        torch.nn.init.kaiming_uniform_(self._8_1_convolution.weight,
                                       a=0.01,
                                       nonlinearity='leaky_relu')
        self._8_2_batch_norm = torch.nn.BatchNorm2d(1024)
        self._8_3_leaky_relu = torch.nn.LeakyReLU(negative_slope=0.01)

        self._9_1_convolution = torch.nn.Conv2d(in_channels=1024,
                                                out_channels=2048,
                                                kernel_size=(2, 2),
                                                stride=(1, 1),
                                                padding=(0, 0),
                                                groups=4,
                                                bias=True,
                                                dtype=torch.float32)
        torch.nn.init.kaiming_uniform_(self._9_1_convolution.weight,
                                       a=0.01,
                                       nonlinearity='leaky_relu')
        self._9_2_leaky_relu = torch.nn.LeakyReLU(negative_slope=0.01)
        return

    def forward(self, inputs: torch.Tensor) -> torch.Tensor:
        """
        The encoder forward pass.

        Args:
            inputs: A batch of latitude-reduced PV images.

        Returns:
            A batch of latent featural encodings.
        """
        x = self._1_1_convolution(inputs)
        x = self._1_2_batch_norm(x)
        x = self._1_3_leaky_relu(x)

        x = self._2_1_convolution(x)
        x = self._2_2_batch_norm(x)
        x = self._2_3_leaky_relu(x)

        x = self._3_1_convolution(x)
        x = self._3_2_batch_norm(x)
        x = self._3_3_leaky_relu(x)

        x = self._4_1_convolution(x)
        x = self._4_2_batch_norm(x)
        x = self._4_3_leaky_relu(x)
        x = self._4_4_attention(x)

        x = self._5_1_convolution(x)
        x = self._5_2_batch_norm(x)
        x = self._5_3_leaky_relu(x)
        x = self._5_4_attention(x)

        x = self._6_1_convolution(x)
        x = self._6_2_batch_norm(x)
        x = self._6_3_leaky_relu(x)
        x = self._6_4_attention(x)

        x = self._7_1_convolution(x)
        x = self._7_2_batch_norm(x)
        x = self._7_3_leaky_relu(x)

        x = self._8_1_convolution(x)
        x = self._8_2_batch_norm(x)
        x = self._8_3_leaky_relu(x)

        x = self._9_1_convolution(x)
        outputs = self._9_2_leaky_relu(x)
        return outputs

forward(inputs)

The encoder forward pass.

Parameters:

Name Type Description Default
inputs Tensor

A batch of latitude-reduced PV images.

 required

Returns:

Type Description
Tensor

A batch of latent featural encodings.
Source code in iris/reversion.py 
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
def forward(self, inputs: torch.Tensor) -> torch.Tensor:
    """
    The encoder forward pass.

    Args:
        inputs: A batch of latitude-reduced PV images.

    Returns:
        A batch of latent featural encodings.
    """
    x = self._1_1_convolution(inputs)
    x = self._1_2_batch_norm(x)
    x = self._1_3_leaky_relu(x)

    x = self._2_1_convolution(x)
    x = self._2_2_batch_norm(x)
    x = self._2_3_leaky_relu(x)

    x = self._3_1_convolution(x)
    x = self._3_2_batch_norm(x)
    x = self._3_3_leaky_relu(x)

    x = self._4_1_convolution(x)
    x = self._4_2_batch_norm(x)
    x = self._4_3_leaky_relu(x)
    x = self._4_4_attention(x)

    x = self._5_1_convolution(x)
    x = self._5_2_batch_norm(x)
    x = self._5_3_leaky_relu(x)
    x = self._5_4_attention(x)

    x = self._6_1_convolution(x)
    x = self._6_2_batch_norm(x)
    x = self._6_3_leaky_relu(x)
    x = self._6_4_attention(x)

    x = self._7_1_convolution(x)
    x = self._7_2_batch_norm(x)
    x = self._7_3_leaky_relu(x)

    x = self._8_1_convolution(x)
    x = self._8_2_batch_norm(x)
    x = self._8_3_leaky_relu(x)

    x = self._9_1_convolution(x)
    outputs = self._9_2_leaky_relu(x)
    return outputs

Decoder

Bases: Module

The decoder module used by Reverter.

Maps a latent featural object of dimensions batch, channel=2048, r=1, lon=1 to a top-down density image of dimensions batch, channel=1, r=512, lon=512, implying a total size expansion factor of \(512 \cdot 512 / 2048 = 128\). See the IRIS paper for architectural details and discussion (subsec: Implementation of Reversion: Architecture).

Attributes:

Name Type Description
_1_1_transpose_convolution ConvTranspose2d

A transpose convolution with in_channels=2048, out_channels=1024, kernel_size=(2, 2), stride=(1, 1), padding=(0, 0), groups=4, bias=True, dtype=torch.float32.
_1_2_batch_norm BatchNorm2d

A batch normalization.
_1_3_leaky_relu LeakyReLU

A leaky ReLU.
_2_1_transpose_convolution ConvTranspose2d

A transpose convolution with in_channels=1024, out_channels=512, kernel_size=(2, 2), stride=(2, 2), padding=(0, 0), groups=1, bias=True, dtype=torch.float32.
_2_2_batch_norm BatchNorm2d

A batch normalization.
_2_3_leaky_relu LeakyReLU

A leaky ReLU.
_3_1_transpose_convolution ConvTranspose2d

A transpose convolution with in_channels=512, out_channels=256, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1), groups=1, bias=True, dtype=torch.float32.
_3_2_batch_norm BatchNorm2d

A batch normalization.
_3_3_leaky_relu LeakyReLU

A leaky ReLU.
_4_1_transpose_convolution ConvTranspose2d

A transpose convolution with in_channels=256, out_channels=128, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1), groups=1, bias=True, dtype=torch.float32.
_4_2_batch_norm BatchNorm2d

A batch normalization.
_4_3_leaky_relu LeakyReLU

A leaky ReLU.
_5_1_transpose_convolution ConvTranspose2d

A transpose convolution with in_channels=128, out_channels=64, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1), groups=1, bias=True, dtype=torch.float32.
_5_2_batch_norm BatchNorm2d

A batch normalization.
_5_3_leaky_relu LeakyReLU

A leaky ReLU.
_6_1_transpose_convolution ConvTranspose2d

A transpose convolution with in_channels=64, out_channels=32, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1), groups=1, bias=True, dtype=torch.float32.
_6_2_batch_norm BatchNorm2d

A batch normalization.
_6_3_leaky_relu LeakyReLU

A leaky ReLU.
_7_1_transpose_convolution ConvTranspose2d

A transpose convolution with in_channels=32, out_channels=16, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1), groups=1, bias=True, dtype=torch.float32.
_7_2_batch_norm BatchNorm2d

A batch normalization.
_7_3_leaky_relu LeakyReLU

A leaky ReLU.
_8_1_transpose_convolution ConvTranspose2d

A transpose convolution with in_channels=16, out_channels=4, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1), groups=1, bias=True, dtype=torch.float32.
_8_2_batch_norm BatchNorm2d

A batch normalization.
_8_3_leaky_relu LeakyReLU

A leaky ReLU.
_9_1_transpose_convolution ConvTranspose2d

A transpose convolution with in_channels=4, out_channels=1, kernel_size=(4, 4), stride=(2, 2), padding=(1, 1), groups=1, bias=True, dtype=torch.float32.
_9_2_relu ReLU

The output hard ReLU used to prevent negative density predictions.
Source code in iris/reversion.py 
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
class Decoder(torch.nn.Module):
    r"""
    The decoder module used by [`Reverter`][iris.reversion.Reverter].

    Maps a latent featural object of dimensions `batch, channel=2048, r=1, lon=1` to a
    [top-down density image][iris.arepo_processing.columnize_physical_tensor] of dimensions
    `batch, channel=1, r=512, lon=512`, implying a total size expansion factor of
    $512 \cdot 512 / 2048 = 128$. See the IRIS paper for architectural details and discussion
    (subsec: Implementation of Reversion: Architecture).

    Attributes:
        _1_1_transpose_convolution: A transpose convolution with `in_channels=2048,
                                                                  out_channels=1024,
                                                                  kernel_size=(2, 2),
                                                                  stride=(1, 1),
                                                                  padding=(0, 0),
                                                                  groups=4,
                                                                  bias=True,
                                                                  dtype=torch.float32`.
        _1_2_batch_norm: A batch normalization.
        _1_3_leaky_relu: A leaky ReLU.

        _2_1_transpose_convolution: A transpose convolution with `in_channels=1024,
                                                                  out_channels=512,
                                                                  kernel_size=(2, 2),
                                                                  stride=(2, 2),
                                                                  padding=(0, 0),
                                                                  groups=1,
                                                                  bias=True,
                                                                  dtype=torch.float32`.
        _2_2_batch_norm: A batch normalization.
        _2_3_leaky_relu: A leaky ReLU.

        _3_1_transpose_convolution: A transpose convolution with `in_channels=512,
                                                                  out_channels=256,
                                                                  kernel_size=(4, 4),
                                                                  stride=(2, 2),
                                                                  padding=(1, 1),
                                                                  groups=1,
                                                                  bias=True,
                                                                  dtype=torch.float32`.
        _3_2_batch_norm: A batch normalization.
        _3_3_leaky_relu: A leaky ReLU.

        _4_1_transpose_convolution: A transpose convolution with `in_channels=256,
                                                                  out_channels=128,
                                                                  kernel_size=(4, 4),
                                                                  stride=(2, 2),
                                                                  padding=(1, 1),
                                                                  groups=1,
                                                                  bias=True,
                                                                  dtype=torch.float32`.
        _4_2_batch_norm: A batch normalization.
        _4_3_leaky_relu: A leaky ReLU.

        _5_1_transpose_convolution: A transpose convolution with `in_channels=128,
                                                                  out_channels=64,
                                                                  kernel_size=(4, 4),
                                                                  stride=(2, 2),
                                                                  padding=(1, 1),
                                                                  groups=1,
                                                                  bias=True,
                                                                  dtype=torch.float32`.
        _5_2_batch_norm: A batch normalization.
        _5_3_leaky_relu: A leaky ReLU.

        _6_1_transpose_convolution: A transpose convolution with `in_channels=64,
                                                                  out_channels=32,
                                                                  kernel_size=(4, 4),
                                                                  stride=(2, 2),
                                                                  padding=(1, 1),
                                                                  groups=1,
                                                                  bias=True,
                                                                  dtype=torch.float32`.
        _6_2_batch_norm: A batch normalization.
        _6_3_leaky_relu: A leaky ReLU.

        _7_1_transpose_convolution: A transpose convolution with `in_channels=32,
                                                                  out_channels=16,
                                                                  kernel_size=(4, 4),
                                                                  stride=(2, 2),
                                                                  padding=(1, 1),
                                                                  groups=1,
                                                                  bias=True,
                                                                  dtype=torch.float32`.
        _7_2_batch_norm: A batch normalization.
        _7_3_leaky_relu: A leaky ReLU.

        _8_1_transpose_convolution: A transpose convolution with `in_channels=16,
                                                                  out_channels=4,
                                                                  kernel_size=(4, 4),
                                                                  stride=(2, 2),
                                                                  padding=(1, 1),
                                                                  groups=1,
                                                                  bias=True,
                                                                  dtype=torch.float32`.
        _8_2_batch_norm: A batch normalization.
        _8_3_leaky_relu: A leaky ReLU.

        _9_1_transpose_convolution: A transpose convolution with `in_channels=4,
                                                                  out_channels=1,
                                                                  kernel_size=(4, 4),
                                                                  stride=(2, 2),
                                                                  padding=(1, 1),
                                                                  groups=1,
                                                                  bias=True,
                                                                  dtype=torch.float32`.
        _9_2_relu: The output hard ReLU used to prevent negative density predictions.
    """

    _1_1_transpose_convolution: torch.nn.ConvTranspose2d
    _1_2_batch_norm: torch.nn.BatchNorm2d
    _1_3_leaky_relu: torch.nn.LeakyReLU

    _2_1_transpose_convolution: torch.nn.ConvTranspose2d
    _2_2_batch_norm: torch.nn.BatchNorm2d
    _2_3_leaky_relu: torch.nn.LeakyReLU

    _3_1_transpose_convolution: torch.nn.ConvTranspose2d
    _3_2_batch_norm: torch.nn.BatchNorm2d
    _3_3_leaky_relu: torch.nn.LeakyReLU

    _4_1_transpose_convolution: torch.nn.ConvTranspose2d
    _4_2_batch_norm: torch.nn.BatchNorm2d
    _4_3_leaky_relu: torch.nn.LeakyReLU

    _5_1_transpose_convolution: torch.nn.ConvTranspose2d
    _5_2_batch_norm: torch.nn.BatchNorm2d
    _5_3_leaky_relu: torch.nn.LeakyReLU

    _6_1_transpose_convolution: torch.nn.ConvTranspose2d
    _6_2_batch_norm: torch.nn.BatchNorm2d
    _6_3_leaky_relu: torch.nn.LeakyReLU

    _7_1_transpose_convolution: torch.nn.ConvTranspose2d
    _7_2_batch_norm: torch.nn.BatchNorm2d
    _7_3_leaky_relu: torch.nn.LeakyReLU

    _8_1_transpose_convolution: torch.nn.ConvTranspose2d
    _8_2_batch_norm: torch.nn.BatchNorm2d
    _8_3_leaky_relu: torch.nn.LeakyReLU

    _9_1_transpose_convolution: torch.nn.ConvTranspose2d
    _9_2_relu: torch.nn.ReLU

    def __init__(self) -> None:
        super().__init__()
        self._1_1_transpose_convolution = torch.nn.ConvTranspose2d(in_channels=2048,
                                                                   out_channels=1024,
                                                                   kernel_size=(2, 2),
                                                                   stride=(1, 1),
                                                                   padding=(0, 0),
                                                                   groups=4,
                                                                   bias=True,
                                                                   dtype=torch.float32)
        torch.nn.init.kaiming_uniform_(self._1_1_transpose_convolution.weight,
                                       a=0.01,
                                       nonlinearity='leaky_relu')
        self._1_2_batch_norm = torch.nn.BatchNorm2d(1024)
        self._1_3_leaky_relu = torch.nn.LeakyReLU(negative_slope=0.01)

        self._2_1_transpose_convolution = torch.nn.ConvTranspose2d(in_channels=1024,
                                                                   out_channels=512,
                                                                   kernel_size=(2, 2),
                                                                   stride=(2, 2),
                                                                   padding=(0, 0),
                                                                   groups=1,
                                                                   bias=True,
                                                                   dtype=torch.float32)
        torch.nn.init.kaiming_uniform_(self._2_1_transpose_convolution.weight,
                                       a=0.01,
                                       nonlinearity='leaky_relu')
        self._2_2_batch_norm = torch.nn.BatchNorm2d(512)
        self._2_3_leaky_relu = torch.nn.LeakyReLU(negative_slope=0.01)

        self._3_1_transpose_convolution = torch.nn.ConvTranspose2d(in_channels=512,
                                                                   out_channels=256,
                                                                   kernel_size=(4, 4),
                                                                   stride=(2, 2),
                                                                   padding=(1, 1),
                                                                   groups=1,
                                                                   bias=True,
                                                                   dtype=torch.float32)
        torch.nn.init.kaiming_uniform_(self._3_1_transpose_convolution.weight,
                                       a=0.01,
                                       nonlinearity='leaky_relu')
        self._3_2_batch_norm = torch.nn.BatchNorm2d(256)
        self._3_3_leaky_relu = torch.nn.LeakyReLU(negative_slope=0.01)

        self._4_1_transpose_convolution = torch.nn.ConvTranspose2d(in_channels=256,
                                                                   out_channels=128,
                                                                   kernel_size=(4, 4),
                                                                   stride=(2, 2),
                                                                   padding=(1, 1),
                                                                   groups=1,
                                                                   bias=True,
                                                                   dtype=torch.float32)
        torch.nn.init.kaiming_uniform_(self._4_1_transpose_convolution.weight,
                                       a=0.01,
                                       nonlinearity='leaky_relu')
        self._4_2_batch_norm = torch.nn.BatchNorm2d(128)
        self._4_3_leaky_relu = torch.nn.LeakyReLU(negative_slope=0.01)

        self._5_1_transpose_convolution = torch.nn.ConvTranspose2d(in_channels=128,
                                                                   out_channels=64,
                                                                   kernel_size=(4, 4),
                                                                   stride=(2, 2),
                                                                   padding=(1, 1),
                                                                   groups=1,
                                                                   bias=True,
                                                                   dtype=torch.float32)
        torch.nn.init.kaiming_uniform_(self._5_1_transpose_convolution.weight,
                                       a=0.01,
                                       nonlinearity='leaky_relu')
        self._5_2_batch_norm = torch.nn.BatchNorm2d(64)
        self._5_3_leaky_relu = torch.nn.LeakyReLU(negative_slope=0.01)

        self._6_1_transpose_convolution = torch.nn.ConvTranspose2d(in_channels=64,
                                                                   out_channels=32,
                                                                   kernel_size=(4, 4),
                                                                   stride=(2, 2),
                                                                   padding=(1, 1),
                                                                   groups=1,
                                                                   bias=True,
                                                                   dtype=torch.float32)
        torch.nn.init.kaiming_uniform_(self._6_1_transpose_convolution.weight,
                                       a=0.01,
                                       nonlinearity='leaky_relu')
        self._6_2_batch_norm = torch.nn.BatchNorm2d(32)
        self._6_3_leaky_relu = torch.nn.LeakyReLU(negative_slope=0.01)

        self._7_1_transpose_convolution = torch.nn.ConvTranspose2d(in_channels=32,
                                                                   out_channels=16,
                                                                   kernel_size=(4, 4),
                                                                   stride=(2, 2),
                                                                   padding=(1, 1),
                                                                   groups=1,
                                                                   bias=True,
                                                                   dtype=torch.float32)
        torch.nn.init.kaiming_uniform_(self._7_1_transpose_convolution.weight,
                                       a=0.01,
                                       nonlinearity='leaky_relu')
        self._7_2_batch_norm = torch.nn.BatchNorm2d(16)
        self._7_3_leaky_relu = torch.nn.LeakyReLU(negative_slope=0.01)

        self._8_1_transpose_convolution = torch.nn.ConvTranspose2d(in_channels=16,
                                                                   out_channels=4,
                                                                   kernel_size=(4, 4),
                                                                   stride=(2, 2),
                                                                   padding=(1, 1),
                                                                   groups=1,
                                                                   bias=True,
                                                                   dtype=torch.float32)
        torch.nn.init.kaiming_uniform_(self._8_1_transpose_convolution.weight,
                                       a=0.01,
                                       nonlinearity='leaky_relu')
        self._8_2_batch_norm = torch.nn.BatchNorm2d(4)
        self._8_3_leaky_relu = torch.nn.LeakyReLU(negative_slope=0.01)

        self._9_1_transpose_convolution = torch.nn.ConvTranspose2d(in_channels=4,
                                                                   out_channels=1,
                                                                   kernel_size=(4, 4),
                                                                   stride=(2, 2),
                                                                   padding=(1, 1),
                                                                   groups=1,
                                                                   bias=True,
                                                                   dtype=torch.float32)
        torch.nn.init.kaiming_uniform_(self._9_1_transpose_convolution.weight,
                                       nonlinearity='relu')
        self._9_2_relu = torch.nn.ReLU()
        return

    def forward(self, inputs: torch.Tensor) -> torch.Tensor:
        """
        The decoder forward pass.

        Args:
            inputs: A batch of latent featural encodings.

        Returns:
            A batch of top-down density images.
        """
        x = self._1_1_transpose_convolution(inputs)
        x = self._1_2_batch_norm(x)
        x = self._1_3_leaky_relu(x)

        x = self._2_1_transpose_convolution(x)
        x = self._2_2_batch_norm(x)
        x = self._2_3_leaky_relu(x)

        x = self._3_1_transpose_convolution(x)
        x = self._3_2_batch_norm(x)
        x = self._3_3_leaky_relu(x)

        x = self._4_1_transpose_convolution(x)
        x = self._4_2_batch_norm(x)
        x = self._4_3_leaky_relu(x)

        x = self._5_1_transpose_convolution(x)
        x = self._5_2_batch_norm(x)
        x = self._5_3_leaky_relu(x)

        x = self._6_1_transpose_convolution(x)
        x = self._6_2_batch_norm(x)
        x = self._6_3_leaky_relu(x)

        x = self._7_1_transpose_convolution(x)
        x = self._7_2_batch_norm(x)
        x = self._7_3_leaky_relu(x)

        x = self._8_1_transpose_convolution(x)
        x = self._8_2_batch_norm(x)
        x = self._8_3_leaky_relu(x)

        x = self._9_1_transpose_convolution(x)
        outputs = self._9_2_relu(x)
        return outputs

forward(inputs)

The decoder forward pass.

Parameters:

Name Type Description Default
inputs Tensor

A batch of latent featural encodings.

 required

Returns:

Type Description
Tensor

A batch of top-down density images.
Source code in iris/reversion.py 
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
def forward(self, inputs: torch.Tensor) -> torch.Tensor:
    """
    The decoder forward pass.

    Args:
        inputs: A batch of latent featural encodings.

    Returns:
        A batch of top-down density images.
    """
    x = self._1_1_transpose_convolution(inputs)
    x = self._1_2_batch_norm(x)
    x = self._1_3_leaky_relu(x)

    x = self._2_1_transpose_convolution(x)
    x = self._2_2_batch_norm(x)
    x = self._2_3_leaky_relu(x)

    x = self._3_1_transpose_convolution(x)
    x = self._3_2_batch_norm(x)
    x = self._3_3_leaky_relu(x)

    x = self._4_1_transpose_convolution(x)
    x = self._4_2_batch_norm(x)
    x = self._4_3_leaky_relu(x)

    x = self._5_1_transpose_convolution(x)
    x = self._5_2_batch_norm(x)
    x = self._5_3_leaky_relu(x)

    x = self._6_1_transpose_convolution(x)
    x = self._6_2_batch_norm(x)
    x = self._6_3_leaky_relu(x)

    x = self._7_1_transpose_convolution(x)
    x = self._7_2_batch_norm(x)
    x = self._7_3_leaky_relu(x)

    x = self._8_1_transpose_convolution(x)
    x = self._8_2_batch_norm(x)
    x = self._8_3_leaky_relu(x)

    x = self._9_1_transpose_convolution(x)
    outputs = self._9_2_relu(x)
    return outputs

PixelSelfAttention2d

Bases: Module

Implements a pixelwise self-attention layer.

Applies a layer norm, followed by a multi-head attention, followed by a layer norm.

Attributes:

Name Type Description
channels int

The number of input/output channels.
num_heads int

The number of attention heads used by the attention block.
pre_norm LayerNorm

The layer norm applied before attention.
attention MultiheadAttention

The attention block.
post_norm LayerNorm

The layer norm applied after attention.

Parameters:

Name Type Description Default
channels int

Sets self.channels.

 required
num_heads int

Sets self.num_heads.

 required
Source code in iris/reversion.py 
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
class PixelSelfAttention2d(torch.nn.Module):
    """
    Implements a pixelwise self-attention layer.

    Applies a layer norm, followed by a multi-head attention, followed by a layer norm.

    Attributes:
        channels: The number of input/output channels.
        num_heads: The number of attention heads used by the attention block.
        pre_norm: The layer norm applied before attention.
        attention: The attention block.
        post_norm: The layer norm applied after attention.

    Args:
        channels: Sets `self.channels`.
        num_heads: Sets `self.num_heads`.
    """

    channels: int
    num_heads: int
    pre_norm: torch.nn.LayerNorm
    attention: torch.nn.MultiheadAttention
    post_norm: torch.nn.LayerNorm

    def __init__(self, channels: int, num_heads: int) -> None:
        super().__init__()
        self.channels = channels
        self.num_heads = num_heads
        self.pre_norm = torch.nn.LayerNorm(channels)
        self.attention = torch.nn.MultiheadAttention(embed_dim=channels,
                                                     num_heads=num_heads,
                                                     batch_first=True)
        self.post_norm = torch.nn.LayerNorm(channels)
        return

    def forward(self, inputs: torch.Tensor) -> torch.Tensor:
        """
        The forward pass of the attention layer.

        Args:
            inputs: A batch of multi-channeled images.

        Returns:
            A batch of self-attended multi-channeled images.
        """
        batch, channel, height, width = inputs.shape
        x = inputs.flatten(2).transpose(1, 2)

        y = self.pre_norm(x)
        a, _ = self.attention(y, y, y)
        x = self.post_norm(x + a)

        outputs = x.transpose(1, 2).view(batch, channel, height, width)
        return outputs

forward(inputs)

The forward pass of the attention layer.

Parameters:

Name Type Description Default
inputs Tensor

A batch of multi-channeled images.

 required

Returns:

Type Description
Tensor

A batch of self-attended multi-channeled images.
Source code in iris/reversion.py 
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
def forward(self, inputs: torch.Tensor) -> torch.Tensor:
    """
    The forward pass of the attention layer.

    Args:
        inputs: A batch of multi-channeled images.

    Returns:
        A batch of self-attended multi-channeled images.
    """
    batch, channel, height, width = inputs.shape
    x = inputs.flatten(2).transpose(1, 2)

    y = self.pre_norm(x)
    a, _ = self.attention(y, y, y)
    x = self.post_norm(x + a)

    outputs = x.transpose(1, 2).view(batch, channel, height, width)
    return outputs