File size: 19,287 Bytes
482ab8a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
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
555
556
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
"""
The MIT License (MIT)

Copyright (c) 2017 Marvin Teichmann
"""

from __future__ import absolute_import, division, print_function

import logging
import math
import os
import sys
import warnings

import numpy as np
import scipy as scp

logging.basicConfig(
    format="%(asctime)s %(levelname)s %(message)s",
    level=logging.INFO,
    stream=sys.stdout,
)

try:
    import pyinn as P

    has_pyinn = True
except ImportError:
    #  PyInn is required to use our cuda based message-passing implementation
    #  Torch 0.4 provides a im2col operation, which will be used instead.
    #  It is ~15% slower.
    has_pyinn = False
    pass

import gc

import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.autograd import Variable
from torch.nn import functional as nnfun
from torch.nn.parameter import Parameter

# Default config as proposed by Philipp Kraehenbuehl and Vladlen Koltun,
default_conf = {
    "filter_size": 11,
    "blur": 4,
    "merge": True,
    "norm": "none",
    "weight": "vector",
    "unary_weight": 1,
    "weight_init": 0.2,
    "trainable": False,
    "convcomp": False,
    "logsoftmax": True,  # use logsoftmax for numerical stability
    "softmax": True,
    "skip_init_softmax": False,
    "final_softmax": False,
    "pos_feats": {
        "sdims": 3,
        "compat": 3,
    },
    "col_feats": {
        "sdims": 80,
        "schan": 13,  # schan depend on the input scale.
        # use schan = 13 for images in [0, 255]
        # for normalized images in [-0.5, 0.5] try schan = 0.1
        "compat": 10,
        "use_bias": False,
    },
    "trainable_bias": False,
    "pyinn": False,
}

# Config used for test cases on 10 x 10 pixel greyscale inpu
test_config = {
    "filter_size": 5,
    "blur": 1,
    "merge": False,
    "norm": "sym",
    "trainable": False,
    "weight": "scalar",
    "unary_weight": 1,
    "weight_init": 0.5,
    "convcomp": False,
    "trainable": False,
    "convcomp": False,
    "logsoftmax": True,  # use logsoftmax for numerical stability
    "softmax": True,
    "pos_feats": {
        "sdims": 1.5,
        "compat": 3,
    },
    "col_feats": {"sdims": 2, "schan": 2, "compat": 3, "use_bias": True},
    "trainable_bias": False,
}


class GaussCRF(nn.Module):
    """Implements ConvCRF with hand-crafted features.

    It uses the more generic ConvCRF class as basis and utilizes a config
    dict to easily set hyperparameters and follows the design choices of:
    Philipp Kraehenbuehl and Vladlen Koltun, "Efficient Inference in Fully
    "Connected CRFs with Gaussian Edge Pots" (arxiv.org/abs/1210.5644)
    """

    def __init__(self, conf, shape, nclasses=None, use_gpu=True):
        super(GaussCRF, self).__init__()

        self.conf = conf
        self.shape = shape
        self.nclasses = nclasses

        self.trainable = conf["trainable"]

        if not conf["trainable_bias"]:
            self.register_buffer("mesh", self._create_mesh())
        else:
            self.register_parameter("mesh", Parameter(self._create_mesh()))

        if self.trainable:

            def register(name, tensor):
                self.register_parameter(name, Parameter(tensor))

        else:

            def register(name, tensor):
                self.register_buffer(name, Variable(tensor))

        register("pos_sdims", torch.Tensor([1 / conf["pos_feats"]["sdims"]]))

        if conf["col_feats"]["use_bias"]:
            register("col_sdims", torch.Tensor([1 / conf["col_feats"]["sdims"]]))
        else:
            self.col_sdims = None

        register("col_schan", torch.Tensor([1 / conf["col_feats"]["schan"]]))
        register("col_compat", torch.Tensor([conf["col_feats"]["compat"]]))
        register("pos_compat", torch.Tensor([conf["pos_feats"]["compat"]]))

        if conf["weight"] is None:
            weight = None
        elif conf["weight"] == "scalar":
            val = conf["weight_init"]
            weight = torch.Tensor([val])
        elif conf["weight"] == "vector":
            val = conf["weight_init"]
            weight = val * torch.ones(1, nclasses, 1, 1)

        self.CRF = ConvCRF(
            shape,
            nclasses,
            mode="col",
            conf=conf,
            use_gpu=use_gpu,
            filter_size=conf["filter_size"],
            norm=conf["norm"],
            blur=conf["blur"],
            trainable=conf["trainable"],
            convcomp=conf["convcomp"],
            weight=weight,
            final_softmax=conf["final_softmax"],
            unary_weight=conf["unary_weight"],
            pyinn=conf["pyinn"],
        )

        return

    def forward(self, unary, img, num_iter=5):
        """Run a forward pass through ConvCRF.

        Arguments:
            unary: torch.Tensor with shape [bs, num_classes, height, width].
                The unary predictions. Logsoftmax is applied to the unaries
                during inference. When using CNNs don't apply softmax,
                use unnormalized output (logits) instead.

            img: torch.Tensor with shape [bs, 3, height, width]
                The input image. Default config assumes image
                data in [0, 255]. For normalized images adapt
                `schan`. Try schan = 0.1 for images in [-0.5, 0.5]
        """

        conf = self.conf

        bs, c, x, y = img.shape

        pos_feats = self.create_position_feats(sdims=self.pos_sdims, bs=bs)
        col_feats = self.create_colour_feats(
            img,
            sdims=self.col_sdims,
            schan=self.col_schan,
            bias=conf["col_feats"]["use_bias"],
            bs=bs,
        )

        compats = [self.pos_compat, self.col_compat]

        self.CRF.add_pairwise_energies([pos_feats, col_feats], compats, conf["merge"])

        prediction = self.CRF.inference(unary, num_iter=num_iter)

        self.CRF.clean_filters()
        return prediction

    def _create_mesh(self, requires_grad=False):
        hcord_range = [range(s) for s in self.shape]
        mesh = np.array(np.meshgrid(*hcord_range, indexing="ij"), dtype=np.float32)

        return torch.from_numpy(mesh)

    def create_colour_feats(self, img, schan, sdims=0.0, bias=True, bs=1):
        norm_img = img * schan

        if bias:
            norm_mesh = self.create_position_feats(sdims=sdims, bs=bs)
            feats = torch.cat([norm_mesh, norm_img], dim=1)
        else:
            feats = norm_img
        return feats

    def create_position_feats(self, sdims, bs=1):
        if type(self.mesh) is Parameter:
            return torch.stack(bs * [self.mesh * sdims])
        else:
            return torch.stack(bs * [Variable(self.mesh) * sdims])


def show_memusage(device=0, name=""):
    import gpustat

    gc.collect()
    gpu_stats = gpustat.GPUStatCollection.new_query()
    item = gpu_stats.jsonify()["gpus"][device]

    logging.info(
        "{:>5}/{:>5} MB Usage at {}".format(
            item["memory.used"], item["memory.total"], name
        )
    )


def exp_and_normalize(features, dim=0):
    """
    Aka "softmax" in deep learning literature
    """
    normalized = torch.nn.functional.softmax(features, dim=dim)
    return normalized


def _get_ind(dz):
    if dz == 0:
        return 0, 0
    if dz < 0:
        return 0, -dz
    if dz > 0:
        return dz, 0


def _negative(dz):
    """
    Computes -dz for numpy indexing. Goal is to use as in array[i:-dz].

    However, if dz=0 this indexing does not work.
    None needs to be used instead.
    """
    if dz == 0:
        return None
    else:
        return -dz


class MessagePassingCol:
    """Perform the Message passing of ConvCRFs.

    The main magic happens here.
    """

    def __init__(
        self,
        feat_list,
        compat_list,
        merge,
        npixels,
        nclasses,
        norm="sym",
        filter_size=5,
        clip_edges=0,
        use_gpu=False,
        blur=1,
        matmul=False,
        verbose=False,
        pyinn=False,
    ):

        if not norm == "sym" and not norm == "none":
            raise NotImplementedError

        span = filter_size // 2
        assert filter_size % 2 == 1
        self.span = span
        self.filter_size = filter_size
        self.use_gpu = use_gpu
        self.verbose = verbose
        self.blur = blur
        self.pyinn = pyinn

        self.merge = merge

        self.npixels = npixels

        if not self.blur == 1 and self.blur % 2:
            raise NotImplementedError

        self.matmul = matmul

        self._gaus_list = []
        self._norm_list = []

        for feats, compat in zip(feat_list, compat_list):
            gaussian = self._create_convolutional_filters(feats)
            if not norm == "none":
                mynorm = self._get_norm(gaussian)
                self._norm_list.append(mynorm)
            else:
                self._norm_list.append(None)

            gaussian = compat * gaussian
            self._gaus_list.append(gaussian)

        if merge:
            self.gaussian = sum(self._gaus_list)
            if not norm == "none":
                raise NotImplementedError

    def _get_norm(self, gaus):
        norm_tensor = torch.ones([1, 1, self.npixels[0], self.npixels[1]])
        normalization_feats = torch.autograd.Variable(norm_tensor)
        if self.use_gpu:
            normalization_feats = normalization_feats.cuda()

        norm_out = self._compute_gaussian(normalization_feats, gaussian=gaus)
        return 1 / torch.sqrt(norm_out + 1e-20)

    def _create_convolutional_filters(self, features):

        span = self.span

        bs = features.shape[0]

        if self.blur > 1:
            off_0 = (self.blur - self.npixels[0] % self.blur) % self.blur
            off_1 = (self.blur - self.npixels[1] % self.blur) % self.blur
            pad_0 = math.ceil(off_0 / 2)
            pad_1 = math.ceil(off_1 / 2)
            if self.blur == 2:
                assert pad_0 == self.npixels[0] % 2
                assert pad_1 == self.npixels[1] % 2

            features = torch.nn.functional.avg_pool2d(
                features,
                kernel_size=self.blur,
                padding=(pad_0, pad_1),
                count_include_pad=False,
            )

            npixels = [
                math.ceil(self.npixels[0] / self.blur),
                math.ceil(self.npixels[1] / self.blur),
            ]
            assert npixels[0] == features.shape[2]
            assert npixels[1] == features.shape[3]
        else:
            npixels = self.npixels

        gaussian_tensor = features.data.new(
            bs, self.filter_size, self.filter_size, npixels[0], npixels[1]
        ).fill_(0)

        gaussian = Variable(gaussian_tensor)

        for dx in range(-span, span + 1):
            for dy in range(-span, span + 1):

                dx1, dx2 = _get_ind(dx)
                dy1, dy2 = _get_ind(dy)

                feat_t = features[:, :, dx1 : _negative(dx2), dy1 : _negative(dy2)]
                feat_t2 = features[
                    :, :, dx2 : _negative(dx1), dy2 : _negative(dy1)
                ]  # NOQA

                diff = feat_t - feat_t2
                diff_sq = diff * diff
                exp_diff = torch.exp(torch.sum(-0.5 * diff_sq, dim=1))

                gaussian[
                    :, dx + span, dy + span, dx2 : _negative(dx1), dy2 : _negative(dy1)
                ] = exp_diff

        return gaussian.view(
            bs, 1, self.filter_size, self.filter_size, npixels[0], npixels[1]
        )

    def compute(self, input):
        if self.merge:
            pred = self._compute_gaussian(input, self.gaussian)
        else:
            assert len(self._gaus_list) == len(self._norm_list)
            pred = 0
            for gaus, norm in zip(self._gaus_list, self._norm_list):
                pred += self._compute_gaussian(input, gaus, norm)

        return pred

    def _compute_gaussian(self, input, gaussian, norm=None):

        if norm is not None:
            input = input * norm

        shape = input.shape
        num_channels = shape[1]
        bs = shape[0]

        if self.blur > 1:
            off_0 = (self.blur - self.npixels[0] % self.blur) % self.blur
            off_1 = (self.blur - self.npixels[1] % self.blur) % self.blur
            pad_0 = int(math.ceil(off_0 / 2))
            pad_1 = int(math.ceil(off_1 / 2))
            input = torch.nn.functional.avg_pool2d(
                input,
                kernel_size=self.blur,
                padding=(pad_0, pad_1),
                count_include_pad=False,
            )
            npixels = [
                math.ceil(self.npixels[0] / self.blur),
                math.ceil(self.npixels[1] / self.blur),
            ]
            assert npixels[0] == input.shape[2]
            assert npixels[1] == input.shape[3]
        else:
            npixels = self.npixels

        if self.verbose:
            show_memusage(name="Init")

        if self.pyinn:
            input_col = P.im2col(input, self.filter_size, 1, self.span)
        else:
            # An alternative implementation of num2col.
            #
            # This has implementation uses the torch 0.4 im2col operation.
            # This implementation was not avaible when we did the experiments
            # published in our paper. So less "testing" has been done.
            #
            # It is around ~20% slower then the pyinn implementation but
            # easier to use as it removes a dependency.
            input_unfold = F.unfold(input, self.filter_size, 1, self.span)
            input_unfold = input_unfold.view(
                bs,
                num_channels,
                self.filter_size,
                self.filter_size,
                npixels[0],
                npixels[1],
            )
            input_col = input_unfold

        k_sqr = self.filter_size * self.filter_size

        if self.verbose:
            show_memusage(name="Im2Col")

        product = gaussian * input_col
        if self.verbose:
            show_memusage(name="Product")

        product = product.view([bs, num_channels, k_sqr, npixels[0], npixels[1]])

        message = product.sum(2)

        if self.verbose:
            show_memusage(name="FinalNorm")

        if self.blur > 1:
            in_0 = self.npixels[0]
            in_1 = self.npixels[1]
            message = message.view(bs, num_channels, npixels[0], npixels[1])
            with warnings.catch_warnings():
                warnings.simplefilter("ignore")
                # Suppress warning regarding corner alignment
                message = torch.nn.functional.upsample(
                    message, scale_factor=self.blur, mode="bilinear"
                )

            message = message[:, :, pad_0 : pad_0 + in_0, pad_1 : in_1 + pad_1]
            message = message.contiguous()

            message = message.view(shape)
            assert message.shape == shape

        if norm is not None:
            message = norm * message

        return message


class ConvCRF(nn.Module):
    """
        Implements a generic CRF class.

    This class provides tools to build
    your own ConvCRF based model.
    """

    def __init__(
        self,
        npixels,
        nclasses,
        conf,
        mode="conv",
        filter_size=5,
        clip_edges=0,
        blur=1,
        use_gpu=False,
        norm="sym",
        merge=False,
        verbose=False,
        trainable=False,
        convcomp=False,
        weight=None,
        final_softmax=True,
        unary_weight=10,
        pyinn=False,
        skip_init_softmax=False,
        eps=1e-8,
    ):

        super(ConvCRF, self).__init__()
        self.nclasses = nclasses

        self.filter_size = filter_size
        self.clip_edges = clip_edges
        self.use_gpu = use_gpu
        self.mode = mode
        self.norm = norm
        self.merge = merge
        self.kernel = None
        self.verbose = verbose
        self.blur = blur
        self.final_softmax = final_softmax
        self.pyinn = pyinn
        self.skip_init_softmax = skip_init_softmax
        self.eps = eps

        self.conf = conf

        self.unary_weight = unary_weight

        if self.use_gpu:
            if not torch.cuda.is_available():
                logging.error("GPU mode requested but not avaible.")
                logging.error("Please run using use_gpu=False.")
                raise ValueError

        self.npixels = npixels

        if type(npixels) is tuple or type(npixels) is list:
            self.height = npixels[0]
            self.width = npixels[1]
        else:
            self.npixels = npixels

        if trainable:

            def register(name, tensor):
                self.register_parameter(name, Parameter(tensor))

        else:

            def register(name, tensor):
                self.register_buffer(name, Variable(tensor))

        if weight is None:
            self.weight = None
        else:
            register("weight", weight)

        if convcomp:
            self.comp = nn.Conv2d(
                nclasses, nclasses, kernel_size=1, stride=1, padding=0, bias=False
            )

            self.comp.weight.data.fill_(0.1 * math.sqrt(2.0 / nclasses))
        else:
            self.comp = None

    def clean_filters(self):
        self.kernel = None

    def add_pairwise_energies(self, feat_list, compat_list, merge):
        assert len(feat_list) == len(compat_list)

        self.kernel = MessagePassingCol(
            feat_list=feat_list,
            compat_list=compat_list,
            merge=merge,
            npixels=self.npixels,
            filter_size=self.filter_size,
            nclasses=self.nclasses,
            use_gpu=self.use_gpu,
            norm=self.norm,
            verbose=self.verbose,
            blur=self.blur,
            pyinn=self.pyinn,
        )

    def inference(self, unary, num_iter=5):

        if not self.skip_init_softmax:
            if not self.conf["logsoftmax"]:
                lg_unary = torch.log(unary)
                prediction = exp_and_normalize(lg_unary, dim=1)
            else:
                lg_unary = nnfun.log_softmax(unary, dim=1, _stacklevel=5)
                prediction = lg_unary
        else:
            unary = unary + self.eps
            unary = unary.clamp(0, 1)
            lg_unary = torch.log(unary)
            prediction = lg_unary

        for i in range(num_iter):
            message = self.kernel.compute(prediction)

            if self.comp is not None:
                # message_r = message.view(tuple([1]) + message.shape)
                comp = self.comp(message)
                message = message + comp

            if self.weight is None:
                prediction = lg_unary + message
            else:
                prediction = (
                    self.unary_weight - self.weight
                ) * lg_unary + self.weight * message

            if not i == num_iter - 1 or self.final_softmax:
                if self.conf["softmax"]:
                    prediction = exp_and_normalize(prediction, dim=1)

        return prediction

    def start_inference(self):
        pass

    def step_inference(self):
        pass


def get_test_conf():
    return test_config.copy()


def get_default_conf():
    return default_conf.copy()