-
Notifications
You must be signed in to change notification settings - Fork 42
/
r2convolution.py
507 lines (370 loc) · 21.1 KB
/
r2convolution.py
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
499
500
501
502
503
504
from torch.nn.functional import conv2d, pad
from escnn.nn import FieldType
from escnn.nn import GeometricTensor
import escnn.nn
from escnn.group import Representation, Group
from escnn.kernels import KernelBasis
from escnn.gspaces import GSpace2D
from .rd_convolution import _RdConv
from typing import Callable, Union, List
import torch
import numpy as np
import math
__all__ = ["R2Conv"]
class R2Conv(_RdConv):
def __init__(self,
in_type: FieldType,
out_type: FieldType,
kernel_size: int,
padding: int = 0,
stride: int = 1,
dilation: int = 1,
padding_mode: str = 'zeros',
groups: int = 1,
bias: bool = True,
sigma: Union[List[float], float] = None,
frequencies_cutoff: Union[float, Callable[[float], int]] = None,
rings: List[float] = None,
maximum_offset: int = None,
recompute: bool = False,
basis_filter: Callable[[dict], bool] = None,
initialize: bool = True,
):
r"""
G-steerable planar convolution mapping between the input and output :class:`~escnn.nn.FieldType` s specified by
the parameters ``in_type`` and ``out_type``.
This operation is equivariant under the action of :math:`\R^2\rtimes G` where :math:`G` is the
:attr:`escnn.nn.FieldType.fibergroup` of ``in_type`` and ``out_type``.
Specifically, let :math:`\rho_\text{in}: G \to \GL{\R^{c_\text{in}}}` and
:math:`\rho_\text{out}: G \to \GL{\R^{c_\text{out}}}` be the representations specified by the input and output
field types.
Then :class:`~escnn.nn.R2Conv` guarantees an equivariant mapping
.. math::
\kappa \star [\mathcal{T}^\text{in}_{g,u} . f] = \mathcal{T}^\text{out}_{g,u} . [\kappa \star f] \qquad\qquad \forall g \in G, u \in \R^2
where the transformation of the input and output fields are given by
.. math::
[\mathcal{T}^\text{in}_{g,u} . f](x) &= \rho_\text{in}(g)f(g^{-1} (x - u)) \\
[\mathcal{T}^\text{out}_{g,u} . f](x) &= \rho_\text{out}(g)f(g^{-1} (x - u)) \\
The equivariance of G-steerable convolutions is guaranteed by restricting the space of convolution kernels to an
equivariant subspace.
As proven in `3D Steerable CNNs <https://arxiv.org/abs/1807.02547>`_, this parametrizes the *most general
equivariant convolutional map* between the input and output fields.
For feature fields on :math:`\R^2` (e.g. images), the complete G-steerable kernel spaces for :math:`G \leq \O2`
is derived in `General E(2)-Equivariant Steerable CNNs <https://arxiv.org/abs/1911.08251>`_.
.. warning ::
This class implements a *discretized* convolution operator over a discrete grid.
This means that equivariance to continuous symmetries is *not* perfect.
In practice, by using sufficiently band-limited filters, the equivariance error introduced by the
discretization of the filters and the features is contained, but some design choices may have a negative
effect on the overall equivariance of the architecture.
During training, in each forward pass the module expands the basis of G-steerable kernels with learned weights
before calling :func:`torch.nn.functional.conv2d`.
When :meth:`~torch.nn.Module.eval()` is called, the filter is built with the current trained weights and stored
for future reuse such that no overhead of expanding the kernel remains.
.. warning ::
When :meth:`~torch.nn.Module.train()` is called, the attributes :attr:`~escnn.nn.R2Conv.filter` and
:attr:`~escnn.nn.R2Conv.expanded_bias` are discarded to avoid situations of mismatch with the
learnable expansion coefficients.
See also :meth:`escnn.nn.R2Conv.train`.
This behaviour can cause problems when storing the :meth:`~torch.nn.Module.state_dict` of a model while in
a mode and lately loading it in a model with a different mode, as the attributes of the class change.
To avoid this issue, we recommend converting the model to eval mode before storing or loading the state
dictionary.
The learnable expansion coefficients of the this module can be initialized with the methods in
:mod:`escnn.nn.init`.
By default, the weights are initialized in the constructors using :func:`~escnn.nn.init.generalized_he_init`.
.. warning ::
This initialization procedure can be extremely slow for wide layers.
In case initializing the model is not required (e.g. before loading the state dict of a pre-trained model)
or another initialization method is preferred (e.g. :func:`~escnn.nn.init.deltaorthonormal_init`), the
parameter ``initialize`` can be set to ``False`` to avoid unnecessary overhead.
The parameters ``basisexpansion``, ``sigma``, ``frequencies_cutoff``, ``rings`` and ``maximum_offset`` are
optional parameters used to control how the basis for the filters is built, how it is sampled on the filter
grid and how it is expanded to build the filter. We suggest to keep these default values.
.. warning ::
Even if the input tensor has a `coords` attribute, the output of this module will not have one.
Args:
in_type (FieldType): the type of the input field, specifying its transformation law
out_type (FieldType): the type of the output field, specifying its transformation law
kernel_size (int): the size of the (square) filter
padding(int, optional): implicit zero paddings on both sides of the input. Default: ``0``
stride(int, optional): the stride of the kernel. Default: ``1``
dilation(int, optional): the spacing between kernel elements. Default: ``1``
padding_mode(str, optional): ``zeros``, ``reflect``, ``replicate`` or ``circular``. Default: ``zeros``
groups (int, optional): number of blocked connections from input channels to output channels.
It allows depthwise convolution. When used, the input and output types need to be
divisible in ``groups`` groups, all equal to each other.
Default: ``1``.
bias (bool, optional): Whether to add a bias to the output (only to fields which contain a
trivial irrep) or not. Default ``True``
sigma (list or float, optional): width of each ring where the bases are sampled. If only one scalar
is passed, it is used for all rings.
frequencies_cutoff (callable or float, optional): function mapping the radii of the basis elements to the
maximum frequency accepted. If a float values is passed, the maximum frequency is equal to the
radius times this factor. By default (``None``), a more complex policy is used.
rings (list, optional): radii of the rings where to sample the bases
maximum_offset (int, optional): number of additional (aliased) frequencies in the intertwiners for finite
groups. By default (``None``), all additional frequencies allowed by the frequencies cut-off
are used.
recompute (bool, optional): if ``True``, recomputes a new basis for the equivariant kernels.
By Default (``False``), it caches the basis built or reuse a cached one, if it is found.
basis_filter (callable, optional): function which takes as input a descriptor of a basis element
(as a dictionary) and returns a boolean value: whether to preserve (``True``) or discard (``False``)
the basis element. By default (``None``), no filtering is applied.
initialize (bool, optional): initialize the weights of the model. Default: ``True``
Attributes:
~.weights (torch.Tensor): the learnable parameters which are used to expand the kernel
~.filter (torch.Tensor): the convolutional kernel obtained by expanding the parameters
in :attr:`~escnn.nn.R2Conv.weights`
~.bias (torch.Tensor): the learnable parameters which are used to expand the bias, if ``bias=True``
~.expanded_bias (torch.Tensor): the equivariant bias which is summed to the output, obtained by expanding
the parameters in :attr:`~escnn.nn.R2Conv.bias`
"""
assert isinstance(in_type.gspace, GSpace2D)
assert isinstance(out_type.gspace, GSpace2D)
basis_filter, self._rings, self._sigma, self._maximum_frequency = compute_basis_params(
kernel_size, frequencies_cutoff, rings, sigma, dilation, basis_filter
)
super(R2Conv, self).__init__(
in_type,
out_type,
2,
kernel_size,
padding,
stride,
dilation,
padding_mode,
groups,
bias,
basis_filter,
recompute,
)
if initialize:
# by default, the weights are initialized with a generalized form of He's weight initialization
escnn.nn.init.generalized_he_init(self.weights.data, self.basisexpansion)
def _build_kernel_basis(self, in_repr: Representation, out_repr: Representation) -> KernelBasis:
return self.space.build_kernel_basis(in_repr, out_repr, self._sigma, self._rings,
maximum_frequency=self._maximum_frequency
)
def forward(self, input: GeometricTensor):
r"""
Convolve the input with the expanded filter and bias.
Args:
input (GeometricTensor): input feature field transforming according to ``in_type``
Returns:
output feature field transforming according to ``out_type``
"""
assert input.type == self.in_type
if not self.training:
_filter = self.filter
_bias = self.expanded_bias
else:
# retrieve the filter and the bias
_filter, _bias = self.expand_parameters()
# use it for convolution and return the result
if self.padding_mode == 'zeros':
output = conv2d(input.tensor, _filter,
stride=self.stride,
padding=self.padding,
dilation=self.dilation,
groups=self.groups,
bias=_bias)
else:
output = conv2d(pad(input.tensor, self._reversed_padding_repeated_twice, self.padding_mode),
_filter,
stride=self.stride,
dilation=self.dilation,
groups=self.groups,
bias=_bias)
return GeometricTensor(output, self.out_type, coords=None)
def check_equivariance(self, atol: float = 0.1, rtol: float = 0.1, assertion: bool = True, verbose: bool = True):
# np.set_printoptions(precision=5, threshold=30 *self.in_type.size**2, suppress=False, linewidth=30 *self.in_type.size**2)
feature_map_size = 33
last_downsampling = 5
first_downsampling = 5
initial_size = (feature_map_size * last_downsampling - 1 + self.kernel_size) * first_downsampling
c = self.in_type.size
import matplotlib.image as mpimg
from skimage.measure import block_reduce
from skimage.transform import resize
# x = mpimg.imread('../group/testimage.jpeg').transpose((2, 0, 1))[np.newaxis, 0:c, :, :]
import scipy
x = scipy.datasets.face().transpose((2, 0, 1))[np.newaxis, 0:c, :, :]
x = resize(
x,
(x.shape[0], x.shape[1], initial_size, initial_size),
anti_aliasing=True
)
x = x / 255.0 - 0.5
if x.shape[1] < c:
to_stack = [x for i in range(c // x.shape[1])]
if c % x.shape[1] > 0:
to_stack += [x[:, :(c % x.shape[1]), ...]]
x = np.concatenate(to_stack, axis=1)
x = GeometricTensor(torch.FloatTensor(x), self.in_type)
def shrink(t: GeometricTensor, s) -> GeometricTensor:
return GeometricTensor(torch.FloatTensor(block_reduce(t.tensor.detach().numpy(), s, func=np.mean)), t.type)
errors = []
for el in self.space.testing_elements:
out1 = self(shrink(x, (1, 1, 5, 5))).transform(el).tensor.detach().numpy()
out2 = self(shrink(x.transform(el), (1, 1, 5, 5))).tensor.detach().numpy()
out1 = block_reduce(out1, (1, 1, 5, 5), func=np.mean)
out2 = block_reduce(out2, (1, 1, 5, 5), func=np.mean)
b, c, h, w = out2.shape
center_mask = np.zeros((2, h, w))
center_mask[1, :, :] = np.arange(0, w) - w / 2
center_mask[0, :, :] = np.arange(0, h) - h / 2
center_mask[0, :, :] = center_mask[0, :, :].T
center_mask = center_mask[0, :, :] ** 2 + center_mask[1, :, :] ** 2 < (h / 4) ** 2
out1 = out1[..., center_mask]
out2 = out2[..., center_mask]
out1 = out1.reshape(-1)
out2 = out2.reshape(-1)
errs = np.abs(out1 - out2)
esum = np.maximum(np.abs(out1), np.abs(out2))
esum[esum == 0.0] = 1
relerr = errs / esum
if verbose:
print(el, relerr.max(), relerr.mean(), relerr.var(), errs.max(), errs.mean(), errs.var())
tol = rtol * esum + atol
if np.any(errs > tol) and verbose:
print(out1[errs > tol])
print(out2[errs > tol])
print(tol[errs > tol])
if assertion:
assert np.all(
errs < tol), 'The error found during equivariance check with element "{}" is too high: max = {}, mean = {} var ={}'.format(
el, errs.max(), errs.mean(), errs.var())
errors.append((el, errs.mean()))
return errors
# init.deltaorthonormal_init(self.weights.data, self.basisexpansion)
# filter = self.basisexpansion()
# center = self.s // 2
# filter = filter[..., center, center]
# assert torch.allclose(torch.eye(filter.shape[1]), filter.t() @ filter, atol=3e-7)
def export(self):
r"""
Export this module to a normal PyTorch :class:`torch.nn.Conv2d` module and set to "eval" mode.
"""
# set to eval mode so the filter and the bias are updated with the current
# values of the weights
self.eval()
_filter = self.filter
_bias = self.expanded_bias
# build the PyTorch Conv2d module
has_bias = self.bias is not None
conv = torch.nn.Conv2d(self.in_type.size,
self.out_type.size,
self.kernel_size,
padding=self.padding,
stride=self.stride,
dilation=self.dilation,
groups=self.groups,
bias=has_bias)
# set the filter and the bias
conv.weight.data = _filter.data
if has_bias:
conv.bias.data = _bias.data
return conv
def bandlimiting_filter(frequency_cutoff: Union[float, Callable[[float], float]]) -> Callable[[dict], bool]:
r"""
Returns a method which takes as input the attributes (as a dictionary) of a basis element and returns a boolean
value: whether to preserve that element (true) or not (false)
If the parameter ``frequency_cutoff`` is a scalar value, the maximum frequency allowed at a certain radius is
proportional to the radius itself. in thi case, the parameter ``frequency_cutoff`` is the factor controlling this
proportionality relation.
If the parameter ``frequency_cutoff`` is a callable, it needs to take as input a radius (a scalar value) and return
the maximum frequency which can be sampled at that radius.
args:
frequency_cutoff (float): factor controlling the bandlimiting
returns:
a function which checks the attributes of individual basis elements and chooses whether to discard them or not
"""
if isinstance(frequency_cutoff, float):
frequency_cutoff = lambda r, fco=frequency_cutoff: r * frequency_cutoff
def bl_filter(attributes: dict) -> bool:
return math.fabs(attributes["irrep:frequency"]) <= frequency_cutoff(attributes["radius"])
return bl_filter
def compute_basis_params(
kernel_size: int,
frequencies_cutoff: Union[float, Callable[[float], float]] = None,
rings: List[float] = None,
sigma: List[float] = None,
dilation: int = 1,
custom_basis_filter: Callable[[dict], bool] = None,
):
width = dilation * (kernel_size - 1) / 2
max_radius = width * np.sqrt(2)
# by default, the number of rings equals half of the filter size
if rings is None:
n_rings = math.ceil(kernel_size / 2)
rings = torch.linspace(0, (kernel_size - 1) // 2, n_rings) * dilation
rings = rings.tolist()
assert all([max_radius >= r >= 0 for r in rings])
if sigma is None:
sigma = [0.6] * (len(rings) - 1) + [0.4]
for i, r in enumerate(rings):
if r == 0.:
sigma[i] = 0.005
elif isinstance(sigma, float):
sigma = [sigma] * len(rings)
if frequencies_cutoff is None:
frequencies_cutoff = -1.
if isinstance(frequencies_cutoff, float):
if frequencies_cutoff == -3:
frequencies_cutoff = _manual_fco3(kernel_size // 2)
elif frequencies_cutoff == -2:
frequencies_cutoff = _manual_fco2(kernel_size // 2)
elif frequencies_cutoff == -1:
frequencies_cutoff = _manual_fco1(kernel_size // 2)
else:
frequencies_cutoff = lambda r, fco=frequencies_cutoff: fco * r
# check if the object is a callable function
assert callable(frequencies_cutoff)
maximum_frequency = int(max(frequencies_cutoff(r) for r in rings))
fco_filter = bandlimiting_filter(frequencies_cutoff)
if custom_basis_filter is not None:
basis_filter = lambda d, custom_basis_filter=custom_basis_filter, fco_filter=fco_filter: (
custom_basis_filter(d) and fco_filter(d))
else:
basis_filter = fco_filter
return basis_filter, rings, sigma, maximum_frequency
def _manual_fco3(max_radius: float) -> Callable[[float], float]:
r"""
Returns a method which takes as input the radius of a ring and returns the maximum frequency which can be sampled
on that ring.
Args:
max_radius (float): radius of the last ring touching the border of the grid
Returns:
a function which checks the attributes of individual basis elements and chooses whether to discard them or not
"""
def bl_filter(r: float) -> float:
max_freq = 0 if r == 0. else 1 if r == max_radius else 2
return max_freq
return bl_filter
def _manual_fco2(max_radius: float) -> Callable[[float], float]:
r"""
Returns a method which takes as input the radius of a ring and returns the maximum frequency which can be sampled
on that ring.
Args:
max_radius (float): radius of the last ring touching the border of the grid
Returns:
a function which checks the attributes of individual basis elements and chooses whether to discard them or not
"""
def bl_filter(r: float) -> float:
max_freq = 0 if r == 0. else min(2 * r, 1 if r == max_radius else 2 * r - (r + 1) % 2)
return max_freq
return bl_filter
def _manual_fco1(max_radius: float) -> Callable[[float], float]:
r"""
Returns a method which takes as input the radius of a ring and returns the maximum frequency which can be sampled
on that ring.
Args:
max_radius (float): radius of the last ring touching the border of the grid
Returns:
a function which checks the attributes of individual basis elements and chooses whether to discard them or not
"""
def bl_filter(r: float) -> float:
max_freq = 0 if r == 0. else min(2 * r, 2 if r == max_radius else 2 * r - (r + 1) % 2)
return max_freq
return bl_filter