-
Notifications
You must be signed in to change notification settings - Fork 18
/
model.py
635 lines (554 loc) · 21.8 KB
/
model.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
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
import argparse
from collections import OrderedDict
import copy
from albumentations.pytorch.transforms import ToTensorV2
import numpy as np
import pytorch_lightning as pl
from sklearn.decomposition import PCA
from sklearn.covariance import LedoitWolf
from sklearn.svm import OneClassSVM
from scipy.stats import chi2
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.nn.parameter import Parameter
from torch.utils.data import DataLoader, RandomSampler
from torchvision.utils import make_grid
from ..common.augmentation import (
dataset_aug,
Compose,
unnormalize,
MaskToSize,
)
from albumentations import Resize, Normalize
from ..common.dataset import (
FoldSplit,
StratifiedFoldSplit,
AnomalyDetectionDataset,
)
from ..common.evaluation import (
log_roc_figure,
latent_tsne_figure,
latent_pca_figure,
MinMaxSaver,
sample_covariance,
)
from ..common.utils import TensorList, batched_index_select, flatten
from .transparent import build_transparent, MODEL_NAMES
from tqdm import tqdm
from typing import NoReturn, Union, Iterable, Optional, Type
class Mahalanobis(pl.LightningModule):
def __init__(
self,
hparams: argparse.Namespace,
dataset_cls: Type[AnomalyDetectionDataset],
) -> NoReturn:
super(Mahalanobis, self).__init__()
self.hparams = hparams
self.normalization_epochs = 100 if self.hparams.augment else 1
self.model, self.input_size = build_transparent(
hparams.arch,
pretrained=True,
extract_blocks=hparams.extract_blocks,
freeze=True,
)
if hasattr(self.model, "fc"):
self.model.fc = nn.Identity()
# We do not want to augment in val/test
final_transform = [Normalize(), ToTensorV2()]
final_eval_transform = copy.deepcopy(final_transform)
if hparams.augment:
uncached = Compose(
[dataset_aug(hparams, dataset_cls), *final_transform]
)
else:
uncached = Compose(final_transform)
dataset = dataset_cls(
hparams=hparams,
train=True,
transform=Resize(self.input_size, self.input_size),
uncached_transform=uncached,
)
self.datasplit = FoldSplit(
dataset,
self.hparams.fold,
uncached_eval_transform=Compose(final_eval_transform),
)
self.testset = dataset_cls(
hparams=hparams,
train=False,
load_masks=True,
transform=Compose(
[
Resize(self.input_size, self.input_size),
# Reduce mask to size for max perf.
MaskToSize(),
]
),
uncached_transform=Compose(final_eval_transform),
)
self.test_saver = MinMaxSaver(unnormalize_fn=unnormalize)
if self.hparams.variance_threshold:
assert (
self.hparams.pca != self.hparams.npca
), "need to set either only pca or npca flag"
else:
assert (
not self.hparams.pca and not self.hparams.npca
), "cannot specify pca/npca and not specify variance_threshold"
self.statistics_computed = False
self._device = torch.device(
"cuda"
if torch.cuda.is_available() and self.hparams.gpus
else "cpu"
)
def forward(self, x: torch.Tensor) -> TensorList:
"""Output features of self.model."""
with torch.no_grad():
z = self.model(x)
return z
@staticmethod
def tensorlist_or_tensor(items: list) -> Union[torch.Tensor, TensorList]:
if len(items) == 1:
return items[0].unsqueeze(0)
return TensorList(items)
def compute_train_sed(self, features: TensorList) -> NoReturn:
"""Compute sed normalization mean & stddev.
This is the per feature independent gaussian assumption (only mean
and stddev).
"""
mean = features.mean(dim=1) # mean is level x features.
stddev = features.std(dim=1)
self.sed_mean = mean # equal to the mean of mvg, but not equal to pca_mean (as sed/maha may be performed on reduced dimensionalities)
self.sed_stddev = stddev
@staticmethod
def compute_mahalanobis_threshold(
k: int, p: float = 0.9973
) -> torch.Tensor:
"""Compute a threshold on the mahalanobis distance.
So that the probability of mahalanobis with k dimensions being less
than the returned threshold is p.
"""
# Mahalanobis² is Chi² distributed with k degrees of freedom.
# So t is square root of the inverse cdf at p.
return torch.Tensor([chi2.ppf(p, k)]).sqrt()
def compute_train_gaussian(self, features: TensorList) -> NoReturn:
"""
features: TensorList
"""
def fit_inv_covariance(samples):
return torch.Tensor(LedoitWolf().fit(samples.cpu()).precision_).to(
samples.device
)
print("Performing Covariance Estimation")
inv_covariance = TensorList(
[fit_inv_covariance(level) for level in features]
)
mean = features.mean(dim=1) # mean features.
self.mvg_mean = mean
self.mvg_inv_covariance = inv_covariance
# Also cache the number of features each level outputs for later.
feature_count = torch.cat(
[torch.Tensor([level.shape[-1]]) for level in mean]
)
self.feature_count = feature_count
def compute_train_ocsvm(self, features: TensorList) -> NoReturn:
def fit_ocsvm(samples):
ocsvm = OneClassSVM(kernel="rbf", gamma="scale")
ocsvm.fit(samples)
return ocsvm
print("Fitting OCSVM")
self.ocsvms = [fit_ocsvm(level) for level in features.cpu()]
def ocsvm_predict(self, features: TensorList) -> TensorList:
return TensorList(
[
torch.Tensor(
-ocsvm.decision_function(level.mean(dim=(-2, -1)).cpu())
)
for level, ocsvm in zip(features, self.ocsvms)
]
)
def compute_pca(
self, features: TensorList, variance_threshold: float = 0.95
) -> NoReturn:
"""Compute pca normalization of teacher features retaining variance.
Contrary to normal pca, this throws away the features with large
variance and only keeps the ones with small amounts of variance.
It is expected that those features will activate on the anomalies.
"""
mean = features.mean(dim=1) # mean is level x features.
def fit_level(features: torch.Tensor) -> torch.Tensor:
pca = PCA(n_components=None).fit(features)
# Select features above variance_threshold.
variances = pca.explained_variance_ratio_.cumsum()
last_component = (variances > variance_threshold).argmax()
# last_component is the index of the last value needed to reach at
# least the required explained variance.
# As the true variance lies somewhere in between [last_component - 1,
# last_component], we include the whole interval for both pca as
# well as NPCA based dimensionality reduction
if self.hparams.pca:
return torch.Tensor(pca.components_[: last_component + 1])
elif self.hparams.npca:
return torch.Tensor(pca.components_[last_component - 1 :])
else:
raise ValueError(
"either hparams.pca or hparams.npca need to be specified"
)
components = self.tensorlist_or_tensor(
[fit_level(level) for level in features.cpu()]
)
self.pca_mean = mean
self.pca_components = components
# TensorList if len(extract_blocks) > 1 else torch.Tensor
def l2_distance(
self, features: Union[torch.Tensor, TensorList]
) -> Union[torch.Tensor, TensorList]:
return (
(features - self.mvg_mean.unsqueeze(1).unsqueeze(-1).unsqueeze(-1))
.mean(dim=(-2, -1))
.pow(2)
.mean(dim=2)
.sqrt()
)
def sed_distance(
self, features: Union[torch.Tensor, TensorList]
) -> Union[torch.Tensor, TensorList]:
"""Return normalized features (using the computed normalization)."""
# Unsqueeze batch, height & width.
return (
(
(
features
- self.sed_mean.unsqueeze(1).unsqueeze(-1).unsqueeze(-1)
)
/ self.sed_stddev.unsqueeze(1).unsqueeze(-1).unsqueeze(-1)
)
.mean(dim=(-2, -1))
.pow(2)
.mean(dim=2)
.sqrt()
)
def pca_reduction(
self, features: Union[torch.Tensor, TensorList]
) -> Union[torch.Tensor, TensorList]:
"""Return pca-reduced features (using the computed PCA)."""
# Features is level x training_samples x features x height x width.
# Unsqueeze batch, height & width.
demeaned = features - self.pca_mean.unsqueeze(1).unsqueeze(
-1
).unsqueeze(-1)
def batched_mul_components(
level: torch.Tensor, components: torch.Tensor
) -> torch.Tensor:
# Cannot use einsum because of unsupported broadcasting.
# So do a permute to (samples x height x width x features).
reduced = torch.matmul( # Batched vector matrix multiply.
level.permute(0, 2, 3, 1).unsqueeze(-2),
components.t().unsqueeze(0).unsqueeze(0).unsqueeze(0),
).squeeze(
-2
) # Squeeze so this is vector matrix multiply.
return reduced.permute(0, 3, 1, 2) # Back to BCHW.
return self.tensorlist_or_tensor(
# This is (X - mean).dot(components.t()).
[
batched_mul_components(level, components)
for level, components in zip(
demeaned, self.pca_components.to(self._device)
)
]
)
@staticmethod
def mahalanobis_distance(
values: torch.Tensor, mean: torch.Tensor, inv_covariance: torch.Tensor
) -> torch.Tensor:
"""Compute the batched mahalanobis distance.
values is a batch of feature vectors.
mean is either the mean of the distribution to compare, or a second
batch of feature vectors.
inv_covariance is the inverse covariance of the target distribution.
"""
assert values.dim() == 2
assert 1 <= mean.dim() <= 2
assert inv_covariance.dim() == 2
assert values.shape[1] == mean.shape[-1]
assert mean.shape[-1] == inv_covariance.shape[0]
assert inv_covariance.shape[0] == inv_covariance.shape[1]
if mean.dim() == 1: # Distribution mean.
mean = mean.unsqueeze(0)
x_mu = values - mean # batch x features
# Same as dist = x_mu.t() * inv_covariance * x_mu batch wise
dist = torch.einsum("im,mn,in->i", x_mu, inv_covariance, x_mu)
return dist.sqrt()
def validation_step(self, batch: dict, batch_idx: int) -> dict:
return OrderedDict({"loss/val": torch.tensor(0)})
def validation_end(self, outputs: list) -> dict:
return {"progress_bar": {}, "log": {}}
def compute_train_statistics(self) -> NoReturn:
# Use same dataloader as training.
trainset = self.datasplit.train()
print(
"Computing training statistics with {} images".format(
len(trainset)
)
)
dataloader = DataLoader(
trainset, batch_size=self.hparams.batch_size, num_workers=2
)
with torch.no_grad():
outputs = []
for epoch in tqdm(range(self.normalization_epochs)):
for batch in dataloader:
z = self.model(batch["image"].to(self._device))
# Mean across locations (at this point to save GPU RAM).
z = z.mean(dim=(3, 4), keepdim=True)
outputs.append(z)
# Features is level x training_samples x features.
features = TensorList.cat(outputs, dim=1)
# TODO: Beautify this so pca reduction does not require singleton dimensions for hxw
if self.hparams.npca or self.hparams.pca:
self.compute_pca(
features.mean(dim=(3, 4)),
variance_threshold=self.hparams.variance_threshold,
)
outputs_reduced = []
for batch in outputs:
reduced = self.pca_reduction(batch)
outputs_reduced.append(reduced.mean(dim=(3, 4)))
features = TensorList.cat(outputs_reduced, dim=1)
else:
features = features.mean(dim=(3, 4))
self.compute_train_gaussian(features)
if self.hparams.ocsvm:
self.compute_train_ocsvm(features)
if self.hparams.sed:
self.compute_train_sed(features)
def test_step(self, batch: dict, batch_idx: int) -> dict:
x, target, mask = batch["image"], batch["target"], batch["mask"]
z = self(x)
if not self.statistics_computed:
self.compute_train_statistics()
self.statistics_computed = True
# reduce features also here
if self.hparams.npca or self.hparams.pca:
z = self.pca_reduction(z)
maha = TensorList(
[
self.mahalanobis_distance(
level.mean(dim=(-2, -1)), val_mean, val_icov
)
for level, val_mean, val_icov in zip(
z, self.mvg_mean, self.mvg_inv_covariance
)
]
)
if self.hparams.ocsvm:
ocsvm = self.ocsvm_predict(z)
if self.hparams.sed:
sed = self.sed_distance(z)
if self.hparams.l2:
l2 = self.l2_distance(z)
self.test_saver.update(x, target, maha.mean(dim=0))
result = OrderedDict({"target": target, "x": x.cpu(), "mask": mask})
for name in ("maha", "sed", "l2", "ocsvm"):
if name in locals():
result[name] = locals()[name].cpu()
return result
def evaluate_latent(
self,
outputs: Iterable[dict],
target: torch.Tensor,
key: str,
images: Optional[torch.Tensor] = None,
) -> dict:
"""Plot & evaluate the latent space from test output.
Return the new entries to the log dict as a dict.
"""
# Latent space is given as [batch x dimension] tensors for each level.
z = TensorList.cat([output[key] for output in outputs], dim=1)
# Iterate all levels of the latent space.
result = {}
for i, (z_i, feature_c) in enumerate(zip(z, self.feature_count)):
# Map index to correct attach_block level.
i = self.hparams.extract_blocks[i]
pred = z_i
auc = log_roc_figure(
self.logger,
target,
pred,
self.current_epoch,
kind="latent_{}/level_{}".format(key, i),
)
result["latent_{}/level_{}/auc/test".format(key, i)] = auc
if key == "maha":
# Mahalanobis TPR/FPR at n sigma.
sigmas = [
1,
2,
3,
4,
5,
6,
7,
8,
] # values > 8 are not computable.
for sigma in sigmas:
# Probability of a Gaussian at n sigma.
p = torch.erf(sigma / torch.DoubleTensor([2]).sqrt())
# Threshold on mahalanobis distance at p (n sigma).
threshold = self.compute_mahalanobis_threshold(
feature_c, p=p.item()
)
anomaly_pred = pred > threshold
tpr = (
anomaly_pred[target != 0].sum()
/ (target != 0).sum().float()
)
result[
"latent_{}/level_{}/sigma_{}/tpr/test".format(
key, i, sigma
)
] = tpr
fpr = (
anomaly_pred[target == 0].sum()
/ (target == 0).sum().float()
)
result[
"latent_{}/level_{}/sigma_{}/fpr/test".format(
key, i, sigma
)
] = fpr
return result
def test_end(self, outputs: Iterable[dict]) -> dict:
self.logger.experiment.add_image(
"min_max_good_pred_images/test",
self.test_saver.good_grid(),
self.current_epoch,
)
self.logger.experiment.add_image(
"min_max_anomaly_pred_images/test",
self.test_saver.anomaly_grid(),
self.current_epoch,
)
# compute & draw roc / connected component analysis
target = torch.cat([output["target"].cpu() for output in outputs], 0)
x = torch.cat([output["x"] for output in outputs], 0)
tqdm_dict = {}
# Classification evaluation.
for distance in ("maha", "sed", "l2", "ocsvm"):
# check if distance is in returned dict
if distance in outputs[0]:
pred = torch.cat(
[
output.get(distance).mean(dim=0).cpu()
for output in outputs
],
0,
)
tqdm_dict[
"{}/full_auc/test".format(distance)
] = log_roc_figure(
self.logger,
target,
pred,
self.current_epoch,
kind="full/{}".format(distance),
)
print(
"{} test auc: {}".format(
distance,
tqdm_dict.get("{}/full_auc/test".format(distance)),
)
)
# Latent space evaluation / visualisation.
tqdm_dict.update(
self.evaluate_latent(outputs, target, distance, images=x)
)
if self.hparams.npca:
# Log how many features where removed.
for i, level in enumerate(self.pca_components):
# Map index to correct attach_block level.
i = self.hparams.extract_blocks[i]
percentage = level.shape[0] / level.shape[1]
print(
"npca features level {}: {} ({})".format(
i, level.shape[0], percentage
)
)
tqdm_dict["npca_feat/level_{}/test".format(i)] = level.shape[0]
tqdm_dict["npca_perc/level_{}/test".format(i)] = percentage
tqdm_dict = flatten(tqdm_dict)
return {"progress_bar": tqdm_dict, "log": tqdm_dict}
# required by pytorch lightning
def configure_optimizers(self) -> list:
return []
@pl.data_loader
def train_dataloader(self) -> DataLoader:
# already shuffled in datasplit
trainset = self.datasplit.train()
sampler = RandomSampler(trainset)
print("Training with {} images".format(len(trainset)))
return DataLoader(
trainset,
batch_size=self.hparams.batch_size,
num_workers=2,
sampler=sampler,
)
@pl.data_loader
def val_dataloader(self) -> DataLoader:
valset = self.datasplit.val()
print("Validating with {} images".format(len(valset)))
return DataLoader(
valset, batch_size=self.hparams.batch_size, num_workers=2
)
@pl.data_loader
def test_dataloader(self) -> DataLoader:
print("Testing with {} images".format(len(self.testset)))
return DataLoader(self.testset, batch_size=self.hparams.batch_size)
@staticmethod
def add_model_specific_args(
parser: argparse.ArgumentParser,
) -> argparse.ArgumentParser:
"""Specify the hyperparams for this LightningModule."""
# MODEL specific
parser.add_argument(
"-a",
"--arch",
metavar="ARCH",
default="resnet18",
choices=MODEL_NAMES,
help="model architecture: "
+ " | ".join(MODEL_NAMES)
+ " (default: resnet18)",
)
parser.add_argument(
"--extract_blocks",
type=int,
nargs="+",
default=[5],
help="Blocks to extract the features from "
"Compared to the paper, we have an index"
" offset of 1 (we star 0 based here but"
" with base 1 in the paper)",
)
parser.add_argument(
"--batch_size", type=int, default=64, help="size of the batches"
)
parser.add_argument(
"--variance_threshold",
type=float,
default=None,
help="variance threshold to apply",
)
parser.add_argument("--pca", action="store_true", help="enable pca")
parser.add_argument("--npca", action="store_true", help="Enable npca")
parser.add_argument(
"--augment", action="store_true", help="Enable data augmentation"
)
parser.add_argument(
"--l2", action="store_true", help="Evaluate l2 distance"
)
parser.add_argument("--sed", action="store_true", help="SED distance")
parser.add_argument("--ocsvm", action="store_true", help="train ocsvm")
return parser