-
Notifications
You must be signed in to change notification settings - Fork 5
/
train_main.py
375 lines (325 loc) · 12 KB
/
train_main.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
"""
Copyright (c) 2022 MPI-Dortmund
SPDX-License-Identifier: MPL-2.0
This file is subject to the terms of the Mozilla Public License, Version 2.0 (MPL-2.0).
The full text of the MPL-2.0 can be found at http://mozilla.org/MPL/2.0/.
For files that are Incompatible With Secondary Licenses, as defined under the MPL-2.0,
additional notices are required. Refer to the MPL-2.0 license for more details on your
obligations and rights under this license and for instructions on how secondary licenses
may affect the distribution and modification of this software.
"""
import os
import random
import sys
from typing import List, Tuple, Dict, Callable
import numpy as np
import tqdm
from pytorch_metric_learning import miners, losses
from tomotwin.modules.common import distances, exceptions
from tomotwin.modules.common.distances import DistanceManager
from tomotwin.modules.common.preprocess import label_filename
from tomotwin.modules.common.utils import check_for_updates
from tomotwin.modules.networks.networkmanager import NetworkManager
from tomotwin.modules.training.LossPyML import LossPyML
from tomotwin.modules.training.argparse_ui import (
TrainingArgParseUI,
TrainingConfiguration,
)
from tomotwin.modules.training.filenamematchingtripletprovidernopdb import (
FilenameMatchingTripletProviderNoPDB,
)
from tomotwin.modules.training.filenametripletprovider import (
FilenameMatchingTripletProvider,
)
from tomotwin.modules.training.filepathtriplet import FilePathTriplet
from tomotwin.modules.training.mrctriplethandler import MRCTripletHandler
from tomotwin.modules.training.torchtrainer import TorchTrainer, TripletDataset
from tomotwin.modules.training.transforms import (
AugmentationPipeline,
AddNoise,
Rotate,
Shift,
RotateFull,
VoxelDropout,
)
try:
from importlib_metadata import version
except ModuleNotFoundError:
from importlib.metadata import version
def get_augmentations(
aug_train_shift_distance: int = 2, use_pdb_as_anchor: bool = True
) -> Tuple[AugmentationPipeline, AugmentationPipeline]:
"""
:param aug_train_shift_distance: Augmentation shifts
:param use_pdb_as_anchor: If true, only pdbs are used as anchors.
:return: Training and validation pipeline
"""
aug_anchor = None
if use_pdb_as_anchor:
aug_anchor = AugmentationPipeline(
# [(CenterCrop, 0.3), (AddNoise, 0.3), (Blur, 0.3), (Rotate, 0.5), (Shift, 0.3)]
augs=[
Rotate(axes=(0, 1)),
VoxelDropout(ratio=(0.05, 0.2)),
RotateFull(axes=(1, 2)), # x-y rotations
],
probs=[
0.3,
0.3,
0.3,
],
)
aug_volumes = AugmentationPipeline(
augs=[
VoxelDropout(ratio=(0.05, 0.2)),
RotateFull(axes=(1, 2)),
Shift(
axis=0,
min_shift=-aug_train_shift_distance,
max_shift=aug_train_shift_distance,
),
Shift(
axis=1,
min_shift=-aug_train_shift_distance,
max_shift=aug_train_shift_distance,
),
Shift(
axis=2,
min_shift=-aug_train_shift_distance,
max_shift=aug_train_shift_distance,
),
AddNoise(sigma=(0, 0.3)),
],
probs=[0.3, 0.3, 0.3, 0.3, 0.3, 0.9],
)
if aug_anchor is None:
aug_anchor = aug_volumes
return aug_anchor, aug_volumes
def train_test_split_anchor_positive(
data: List[FilePathTriplet], split: float = 0.8
) -> Tuple[List, List]:
"""
Create a train/test split of the data. It will make sure, that the anchor-positives pairs are
unique in train and test split.
"""
unique_pdbs = set()
for fpt in data:
unique_pdbs.add(fpt.anchor)
unique_pdbs = list(unique_pdbs)
unique_pdbs.sort()
unique_volumes = set()
for fpt in data:
unique_volumes.add(fpt.positive)
unique_volumes = list(unique_volumes)
unique_volumes.sort()
anchor_positives_pairs = []
for pdb in unique_pdbs:
for vol in unique_volumes:
if os.path.splitext(os.path.basename(pdb))[0].upper() in vol.upper():
anchor_positives_pairs.append((pdb, vol))
np.random.shuffle(anchor_positives_pairs)
split = int(split * len(anchor_positives_pairs))
train_anchor_positive = anchor_positives_pairs[:split]
valid_anchor_positive = anchor_positives_pairs[split:]
train_triplets = []
valid_triplets = []
def is_in_subset(triplet, pairs: List):
for p in pairs:
if triplet.anchor == p[0] and triplet.positive == p[1]:
return True
return False
for triplet in tqdm.tqdm(data, "Train/Valid split"):
if is_in_subset(triplet, train_anchor_positive):
train_triplets.append(triplet)
continue
if is_in_subset(triplet, valid_anchor_positive):
valid_triplets.append(triplet)
return train_triplets, valid_triplets
def generate_triplets(
tconf: TrainingConfiguration,
) -> Tuple[List[FilePathTriplet], List[FilePathTriplet]]:
"""
Generates the train/test triplets
:param tconf: Training configuration
:return: Two lists of FilePathTriplets. One for training, one for testing.
"""
if tconf.pdb_path:
tripletprov = FilenameMatchingTripletProvider(
path_pdb=tconf.pdb_path,
path_volume=tconf.volume_path,
max_neg=tconf.max_neg,
mask_pdb="**/*.mrc",
mask_volumes="**/*.mrc",
)
else:
tripletprov = FilenameMatchingTripletProviderNoPDB(
path_volume=tconf.volume_path, mask_volumes="**/*.mrc"
)
master_list = tripletprov.get_triplets()
if tconf.validvolumes is None:
if len(master_list) == 0:
print("No training data could be found")
sys.exit(0)
# Get triplets and make dataframes for sampling during training
train_triplets, test_triplets = train_test_split_anchor_positive(
master_list, split=0.8
)
else:
print("Use validvolumes path for validation data.")
train_triplets = master_list
if tconf.pdb_path:
tripletprov = FilenameMatchingTripletProvider(
path_pdb=tconf.pdb_path,
path_volume=tconf.validvolumes,
max_neg=tconf.max_neg,
mask_pdb="**/*.mrc",
mask_volumes="**/*.mrc",
)
else:
tripletprov = FilenameMatchingTripletProviderNoPDB(
path_volume=tconf.validvolumes, mask_volumes="**/*.mrc"
)
test_triplets = tripletprov.get_triplets()
return train_triplets, test_triplets
def get_loss_func(
net_conf: Dict, train_conf: Dict, distance: distances.Distance
) -> Callable:
"""
:param net_conf: Network confiiguration dict
:param train_conf: Training configuration dict
:param distance: Distance measure
:return: Loss function
"""
if train_conf["loss"] == "ArcFaceLoss":
loss_func = losses.ArcFaceLoss(
num_classes=train_conf["num_classes"],
margin=train_conf["af_margin"],
distance=distance,
scale=train_conf["af_scale"],
embedding_size=net_conf["output_channels"],
)
elif train_conf["loss"] == "SphereFaceLoss":
loss_func = losses.SphereFaceLoss(
num_classes=train_conf["num_classes"],
margin=train_conf["sf_margin"],
distance=distance,
scale=train_conf["sf_scale"],
embedding_size=net_conf["output_channels"],
)
elif train_conf["loss"] == "TripletLoss":
loss_func = losses.TripletMarginLoss(
margin=train_conf["tl_margin"], distance=distance
)
else:
raise exceptions.UnknownLoss("Specified loss not known")
return loss_func
def _main_():
seed = 17 # seed value
np.random.seed(seed)
random.seed(seed)
########################
# Get configuration from user interface
########################
ui = TrainingArgParseUI()
ui.run()
check_for_updates()
tconf = ui.get_training_configuration()
os.makedirs(tconf.output_path, exist_ok=True)
pth_log_out = os.path.join(tconf.output_path, "out.txt")
pth_log_err = os.path.join(tconf.output_path, "err.txt")
print("Redirecting stdout to", pth_log_out)
print("Redirecting stderr to", pth_log_err)
f = open(pth_log_out, "a", encoding="utf-8")
sys.stdout = f
f = open(pth_log_err, "a", encoding="utf-8")
sys.stderr = f
print("TomoTwin Version:", version("tomotwin-cryoet"))
########################
# Generate Triplets
########################
train_triplets, test_triplets = generate_triplets(tconf)
print("Number of train triplets:", len(train_triplets))
print("Number of validation triplets:", len(test_triplets))
########################
# Load config
########################
nw = NetworkManager()
config = nw.load_configuration(tconf.netconfig)
########################
# Configure datasets
########################
aug_args = {"use_pdb_as_anchor": tconf.pdb_path is not None}
if "aug_train_shift_distance" in config["train_config"]:
print("Use shift:", config["train_config"]["aug_train_shift_distance"])
aug_args["aug_train_shift_distance"] = config["train_config"][
"aug_train_shift_distance"
]
aug_anchor, aug_volumes = get_augmentations(**aug_args)
train_ds = TripletDataset(
training_data=train_triplets,
handler=MRCTripletHandler(),
augmentation_anchors=aug_anchor,
augmentation_volumes=aug_volumes,
label_ext_func=label_filename,
)
test_ds = TripletDataset(
training_data=test_triplets,
handler=MRCTripletHandler(),
label_ext_func=label_filename,
)
########################
# Init distance function
########################
dm = DistanceManager()
distance = dm.get_distance(tconf.distance)
print("Use distance function", distance.name())
########################
# Setup network
########################
config["distance"] = distance.name()
network = nw.create_network(config)
########################
# Setup miners and loss
########################
miner = None
if config["train_config"]["miner"]:
miner = miners.TripletMarginMiner(
margin=config["train_config"]["miner_margin"], type_of_triplets="semihard"
)
loss_func = get_loss_func(
net_conf=config["network_config"],
train_conf=config["train_config"],
distance=distance,
)
########################
# Create trainer and start training
########################
only_negative_labels = []
if "only_negative_labels" in config["train_config"]:
only_negative_labels = config["train_config"]["only_negative_labels"]
trainer = TorchTrainer(
epochs=tconf.num_epochs,
batchsize=int(config["train_config"]["batchsize"]),
learning_rate=config["train_config"]["learning_rate"],
network=network,
criterion=LossPyML(
loss_func=loss_func, miner=miner, only_negative_labels=only_negative_labels
),
workers=12,
log_dir=os.path.join(tconf.output_path, "tensorboard"),
training_data=train_ds,
test_data=test_ds,
output_path=tconf.output_path,
checkpoint=tconf.checkpoint,
optimizer=config["train_config"]["optimizer"],
weight_decay=config["train_config"]["weight_decay"],
patience=config["train_config"]["patience"],
save_epoch_seperately=tconf.save_after_improvement,
)
trainer.set_seed(seed)
config["window_size"] = tuple(train_ds.get_triplet_dimension())
trainer.set_network_config(config)
trainer.train()
trainer.write_results_to_disk(tconf.output_path)
if __name__ == "__main__":
_main_()