-
Notifications
You must be signed in to change notification settings - Fork 3
/
semantic_segmentation.py
456 lines (392 loc) · 23.2 KB
/
semantic_segmentation.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
import os
import urllib.request
import tarfile
import numpy as np
import torch
import torch.nn
import torch.nn.functional
import pytorch_lightning as pl
from pytorch_lightning import loggers as pl_loggers
from typing import Sequence, Tuple, Union, Optional
from torch.utils.data import DataLoader
import networks
import datasets
def calculate_binary_iou_batch(prediction, label):
"""
Calculates an iou core for binary tensors, maintaining the batch dimension.
Defaults to a value of 0 if the union is 0
:param prediction: binary tensor of shape (batch, 1, x, y)
:param label: binary tensor of shape (batch, 1, x, y)
:return: tensor of shape (batch)
"""
assert prediction.shape[1] == 1
assert label.shape[1] == 1
intersection = torch.sum(torch.logical_and(prediction, label), dim=[1, 2, 3])
union = torch.sum(torch.logical_or(prediction, label), dim=[1, 2, 3])
iou = torch.div(intersection, union)
iou[torch.logical_or(torch.isinf(iou), torch.isnan(iou))] = 0
return iou
class ClassificationDataModule(pl.LightningDataModule):
"""
A DataModule that loads all the datasets with defaults preset for classification purposes.
It outputs the relevant training dataloader, and for validation and testing a tuple of dataloaders in the following
order: (synthetic, real, original)
"""
def __init__(self,
dataset_identifier: str,
cross_validation_i: int,
):
"""
:param dataset_identifier: Which dataset to train on. One of the following:
'original' 'new_random', 'new_position', 'new_orientation', 'new_length'
:param cross_validation_i: An integer in [0, 4) specifying which cross-validation split to select.
"""
super().__init__()
self.dataset_identifier = dataset_identifier
self.cross_validation_i = cross_validation_i
if self.dataset_identifier == 'original':
self.train_on_original = True
self.category_order = 'random'
elif self.dataset_identifier == 'new_random':
self.train_on_original = False
self.category_order = 'random'
elif self.dataset_identifier == 'new_position':
self.train_on_original = False
self.category_order = 'position'
elif self.dataset_identifier == 'new_orientation':
self.train_on_original = False
self.category_order = 'orientation'
elif self.dataset_identifier == 'new_length':
self.train_on_original = False
self.category_order = 'length'
else:
raise ValueError(f'{self.dataset_identifier} is not a valid dataset identifier')
# Default params
self.batchsize = 64
self.dtype = np.float32
self.num_workers = 4
# new synthetic dataset params
self.imsize_synthetic = (128, 128) # half resolution will half this
self.train_batches_per_epoch = 128
self.val_batches_per_epoch = 8
self.test_batches_per_epoch = 8
self.filepath_real = os.path.join('data')
self.filepath_new_synthetic = os.path.join('data', 'separate.pickle')
self.filepath_original = os.path.join('data', 'Cleaned_LowRes_13434_overlapping_pairs.h5')
# placeholders
self.dataset_original_train = None
self.dataset_original_val = None
self.dataset_original_test = None
self.dataset_synthetic_train = None
self.dataset_synthetic_val = None
self.dataset_synthetic_test = None
self.dataset_real_val = None
self.dataset_real_test = None
def prepare_data(self):
if not os.path.isfile(self.filepath_original):
tar_path = os.path.join('data', 'Cleaned_LowRes_13434_overlapping_pairs.tar.xz')
if not os.path.isfile(tar_path):
url = "https://github.com/jeanpat/DeepFISH/blob/master/dataset/" \
"Cleaned_LowRes_13434_overlapping_pairs.tar.xz?raw=true"
filename, headers = urllib.request.urlretrieve(url, tar_path)
with tarfile.open(tar_path, 'r') as f:
f.extractall('data')
def setup(self, stage=None):
# original dataset
train_subsets_cv = {0: [(0.0, 0.6)],
1: [(0.0, 0.4), (0.6, 0.8)],
2: [(0.0, 0.2), (0.4, 0.8)],
3: [(0.2, 0.8)]}
val_subsets_cv = {0: [(0.6, 0.8)],
1: [(0.4, 0.6)],
2: [(0.2, 0.4)],
3: [(0.0, 0.2)]}
train_subset = train_subsets_cv[self.cross_validation_i]
val_subset = val_subsets_cv[self.cross_validation_i]
test_subset = [(0.8, 1.0)]
if self.train_on_original:
self.dataset_original_train = datasets.OriginalChromosomeDataset(self.filepath_original,
train_subset,
True,
False,
self.batchsize,
dtype=self.dtype)
self.dataset_original_val = datasets.OriginalChromosomeDataset(self.filepath_original,
val_subset,
True,
False,
self.batchsize,
fix_random_seed=True,
dtype=self.dtype)
self.dataset_original_test = datasets.OriginalChromosomeDataset(self.filepath_original,
test_subset,
True,
False,
self.batchsize,
fix_random_seed=True,
dtype=self.dtype)
# new synthetic dataset
train_slides_cv = {0: (0, 1, 2, 3, 4, 5, 6, 7, 8),
1: (0, 1, 2, 3, 4, 5, 9, 10, 11),
2: (0, 1, 2, 6, 7, 8, 9, 10, 11),
3: (3, 4, 5, 6, 7, 8, 9, 10, 11)}
val_slides_cv = {0: (9, 10, 11),
1: (6, 7, 8),
2: (3, 4, 5),
3: (0, 1, 2)}
train_slides = train_slides_cv[self.cross_validation_i]
val_slides = val_slides_cv[self.cross_validation_i]
test_slides = (12, 13, 14)
output_channels_list = ['dapi_cy3', '4_channel']
if not self.train_on_original:
self.dataset_synthetic_train = datasets.SyntheticChromosomeDataset(self.filepath_new_synthetic,
self.imsize_synthetic,
train_slides,
True,
self.batchsize,
self.train_batches_per_epoch,
output_channels_list,
self.category_order,
fix_random_seed=False,
dtype=self.dtype)
self.dataset_synthetic_val = datasets.SyntheticChromosomeDataset(self.filepath_new_synthetic,
self.imsize_synthetic,
val_slides,
True,
self.batchsize,
self.val_batches_per_epoch,
output_channels_list,
self.category_order,
fix_random_seed=True,
dtype=self.dtype)
self.dataset_synthetic_test = datasets.SyntheticChromosomeDataset(self.filepath_new_synthetic,
self.imsize_synthetic,
test_slides,
True,
self.batchsize,
self.test_batches_per_epoch,
output_channels_list,
self.category_order,
fix_random_seed=True,
dtype=self.dtype)
# real dataset
self.dataset_real_val = datasets.RealOverlappingChromosomes(self.filepath_real,
True,
(0, 0.5),
separate_channels=False,
half_resolution=True,
output_categories=4,
dtype=self.dtype)
self.dataset_real_test = datasets.RealOverlappingChromosomes(self.filepath_real,
True,
(0.5, 1),
separate_channels=False,
half_resolution=True,
output_categories=4,
dtype=self.dtype)
def train_dataloader(self):
if self.train_on_original:
return DataLoader(self.dataset_original_train,
batch_size=None,
num_workers=self.num_workers,
pin_memory=self.num_workers > 0)
else:
return DataLoader(self.dataset_synthetic_train,
batch_size=None,
num_workers=self.num_workers,
pin_memory=self.num_workers > 0)
def val_dataloader(self):
dataloader_synthetic = DataLoader(self.dataset_synthetic_val,
batch_size=None,
num_workers=self.num_workers,
pin_memory=self.num_workers > 0)
dataloader_real = DataLoader(self.dataset_real_val,
batch_size=None,
num_workers=self.num_workers,
pin_memory=self.num_workers > 0)
dataloader_original = DataLoader(self.dataset_original_val,
batch_size=None,
num_workers=self.num_workers,
pin_memory=self.num_workers > 0)
return dataloader_synthetic, dataloader_real, dataloader_original
def test_dataloader(self):
dataloader_synthetic = DataLoader(self.dataset_synthetic_test,
batch_size=None,
num_workers=self.num_workers,
pin_memory=self.num_workers > 0)
dataloader_real = DataLoader(self.dataset_real_test,
batch_size=None,
num_workers=self.num_workers,
pin_memory=self.num_workers > 0)
dataloader_original = DataLoader(self.dataset_original_test,
batch_size=None,
num_workers=self.num_workers,
pin_memory=self.num_workers > 0)
return dataloader_synthetic, dataloader_real, dataloader_original
class ClassificationModule(pl.LightningModule):
def __init__(self, smaller_network: bool):
"""
Module with hard coded parameters for everything.
:param smaller_network: Whether to use the smaller network (Hu et al) or larger (Saleh et al)
"""
super().__init__()
self.save_hyperparameters()
# hardcoded parameters
if smaller_network:
backbone_net = networks.FullyConv(n_channels_in=128,
ns_channels_layers=[256, 128],
activation=torch.nn.functional.relu,
kernel_size=3,
groups=1,
norm_layer=torch.nn.BatchNorm2d,
raw_output=False)
self.net = networks.Unet(n_channels_in=1,
n_channels_out=4,
n_channels_start=64,
depth_encoder=2,
depth_decoder=2,
n_resolutions=3,
backbone_net=backbone_net,
input_net=None,
output_net=None,
mode_add=False)
else:
backbone_net = networks.FullyConv(n_channels_in=256,
ns_channels_layers=[512, 256],
activation=torch.nn.functional.relu,
kernel_size=3,
groups=1,
norm_layer=torch.nn.BatchNorm2d,
raw_output=False)
self.net = networks.Unet(n_channels_in=1,
n_channels_out=4,
n_channels_start=64,
depth_encoder=2,
depth_decoder=2,
n_resolutions=4,
backbone_net=backbone_net,
input_net=None,
output_net=None,
mode_add=False)
def forward(self, x):
return self.net(x)
def training_step(self, batch, batch_step):
batch_prediction, batch_label = self._step_core(batch)
loss = torch.nn.functional.cross_entropy(batch_prediction, batch_label[:, 0, ...])
metrics = self._calculate_metrics(batch_prediction, batch_label, 'training')
metrics['loss'] = loss
self.log_dict(metrics, on_step=True)
return loss
def validation_step(self, batch, batch_step, dataloader_idx):
batch_prediction, batch_label = self._step_core(batch)
if dataloader_idx == 0:
dataset_name = 'val_synthetic'
elif dataloader_idx == 1:
dataset_name = 'val_real'
else:
dataset_name = 'val_original'
metrics = self._calculate_metrics(batch_prediction, batch_label, dataset_name)
self.log_dict(metrics, on_epoch=True)
return metrics
def test_step(self, batch, batch_step, dataloader_idx):
batch_prediction, batch_label = self._step_core(batch)
if dataloader_idx == 0:
dataset_name = 'test_synthetic'
elif dataloader_idx == 1:
dataset_name = 'test_real'
else:
dataset_name = 'test_original'
metrics = self._calculate_metrics(batch_prediction, batch_label, dataset_name)
self.log_dict(metrics, on_epoch=True)
return metrics
def configure_optimizers(self):
optimizer = torch.optim.Adam(self.parameters())
return optimizer
def _step_core(self, batch):
batch_in = batch[:, 0:1, ...]
batch_label = batch[:, 1:2, ...].long()
batch_prediction = self.net(batch_in)
return batch_prediction, batch_label
def _calculate_metrics(self, batch_prediction, batch_label, dataset_name):
batch_prediction = batch_prediction.detach()
batch_label = batch_label.detach()
batch_label_background = torch.eq(batch_label, 0)
batch_label_ch0 = torch.eq(batch_label, 1)
batch_label_ch1 = torch.eq(batch_label, 2)
batch_label_overlap = torch.eq(batch_label, 3)
batch_prediction_class = torch.argmax(batch_prediction, dim=1, keepdim=True)
batch_prediction_background = torch.eq(batch_prediction_class, 0)
batch_prediction_ch0 = torch.eq(batch_prediction_class, 1)
batch_prediction_ch1 = torch.eq(batch_prediction_class, 2)
batch_prediction_overlap = torch.eq(batch_prediction_class, 3)
batch_iou_background = calculate_binary_iou_batch(batch_prediction_background, batch_label_background)
batch_iou_ch0 = calculate_binary_iou_batch(batch_prediction_ch0, batch_label_ch0)
batch_iou_ch1 = calculate_binary_iou_batch(batch_prediction_ch1, batch_label_ch1)
batch_iou_overlap = calculate_binary_iou_batch(batch_prediction_overlap, batch_label_overlap)
batch_iou_ch0_switched = calculate_binary_iou_batch(batch_prediction_ch1, batch_label_ch0)
batch_iou_ch1_switched = calculate_binary_iou_batch(batch_prediction_ch0, batch_label_ch1)
switched_better = torch.lt(
torch.mean(torch.stack([batch_iou_ch0, batch_iou_ch1], dim=0), dim=0),
torch.mean(torch.stack([batch_iou_ch0_switched, batch_iou_ch1_switched], dim=0), dim=0)
)
standard_better = torch.logical_not(switched_better)
batch_prediction_ch0_best = batch_prediction_ch0 * standard_better[:, None, None, None] + \
batch_prediction_ch1 * switched_better[:, None, None, None]
batch_prediction_ch1_best = batch_prediction_ch1 * standard_better[:, None, None, None] + \
batch_prediction_ch0 * switched_better[:, None, None, None]
batch_iou_ch0_best = calculate_binary_iou_batch(batch_prediction_ch0_best, batch_label_ch0)
batch_iou_ch1_best = calculate_binary_iou_batch(batch_prediction_ch1_best, batch_label_ch1)
batch_iou_ch0_or_ch1 = calculate_binary_iou_batch(
torch.logical_or(batch_prediction_ch0, batch_prediction_ch1),
torch.logical_or(batch_label_ch0, batch_label_ch1)
) # no need to use best, because OR is the same
batch_iou_separated_ch0 = calculate_binary_iou_batch(
torch.logical_or(batch_prediction_ch0_best, batch_prediction_overlap),
torch.logical_or(batch_label_ch0, batch_label_overlap)
)
batch_iou_separated_ch1 = calculate_binary_iou_batch(
torch.logical_or(batch_prediction_ch1_best, batch_prediction_overlap),
torch.logical_or(batch_label_ch1, batch_label_overlap)
)
iou_background = torch.mean(batch_iou_background)
iou_ch0_best = torch.mean(batch_iou_ch0_best)
iou_ch1_best = torch.mean(batch_iou_ch1_best)
iou_overlap = torch.mean(batch_iou_overlap)
average_iou_classes = torch.mean(torch.stack([iou_background, iou_ch0_best, iou_ch1_best, iou_overlap]))
average_iou_separated = torch.mean(torch.stack([batch_iou_separated_ch0, batch_iou_separated_ch1]))
metrics = {dataset_name + '_iou_background': torch.mean(batch_iou_background),
dataset_name + '_iou_ch0': torch.mean(batch_iou_ch0_best),
dataset_name + '_iou_ch1': torch.mean(batch_iou_ch1_best),
dataset_name + '_iou_overlap': torch.mean(batch_iou_overlap),
dataset_name + '_average_iou_classes': average_iou_classes,
dataset_name + '_average_iou_separated': average_iou_separated,
dataset_name + '_iou_ch0_or_ch1': torch.mean(batch_iou_ch0_or_ch1)}
return metrics
def train(dataset_identifier: str,
smaller_network: bool,
cross_validation_i: int):
run_name = os.path.join(f"{dataset_identifier}_{'snet' if smaller_network else 'lnet'}", f"cv{cross_validation_i}")
max_epochs = 128
early_stopping_patience = 8
classification_module = ClassificationModule(smaller_network)
classification_data_module = ClassificationDataModule(dataset_identifier, cross_validation_i)
logger = pl_loggers.TensorBoardLogger('results/semantic_segmentation', name=run_name, default_hp_metric=False)
if dataset_identifier == 'original':
main_metric = 'val_original_average_iou_classes/dataloader_idx_2'
else:
main_metric = 'val_synthetic_average_iou_classes/dataloader_idx_0'
early_stopping_callback = pl.callbacks.EarlyStopping(main_metric,
patience=early_stopping_patience,
mode='max')
checkpoint_callback = pl.callbacks.ModelCheckpoint(monitor=main_metric, mode='max',
filename='best_{epoch}_{step}')
trainer = pl.Trainer(logger=logger, gpus=1, max_epochs=max_epochs,
callbacks=[early_stopping_callback, checkpoint_callback])
trainer.fit(classification_module, datamodule=classification_data_module)
def train_all():
for smaller_network in (False, True):
for dataset_identifier in ('original', 'new_random', 'new_position', 'new_orientation', 'new_length'):
for cross_validation_i in (0, 1, 2, 3):
train(dataset_identifier, smaller_network, cross_validation_i)
if __name__ == '__main__':
train_all()