-
Notifications
You must be signed in to change notification settings - Fork 2
/
experiment.py
419 lines (349 loc) · 14 KB
/
experiment.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
import os
import numpy as np
import time
from argparse import Namespace
import torch
from summarytracker import SummaryTracker
from monolab.data_loader import prepare_dataloader
from monolab.loss import MonodepthLoss
from test import run_test
from utils import get_model, setup_logging, notify_mail, to_device, time_delta_now
import logging
logger = logging.getLogger(__name__)
class Experiment:
""" A class for training and testing a model that contains the actual network as self.model
The arguments are defined in the arg_parse()-function above.
Important:
- args.data_dir is the root directory
- args.filenames_file should be different depending on whether you want to train or test the model
- args.val_filenames_file is only used during training
"""
def __init__(self, args: Namespace, base_dir: str):
# Set seed for reproducibility
torch.manual_seed(args.seed)
np.random.seed(args.seed)
if torch.cuda.is_available():
torch.cuda.manual_seed_all(args.seed)
self.args = args
self.base_dir = base_dir
self.loss_names = dict(
full="monodepth-loss",
images="image-loss",
disp_gp="disparity-gradient-loss",
lr_consistency="lr-consistency-loss",
)
# Setup summary tracker
self.summary = SummaryTracker(
metric_names=list(self.loss_names.values()), args=args, base_dir=base_dir
)
# Get the model
self.model = self._get_model(args)
# Setup loss, optimizer and validation set
self.loss_function = MonodepthLoss(
device=self.device,
SSIM_w=args.weight_ssim,
disp_gradient_w=args.weight_disp_gradient,
lr_w=args.weight_lr_consistency,
).to(self.device)
logger.debug(f"Using loss function: {self.loss_function}")
self.optimizer = torch.optim.Adam(
self.model.parameters(), lr=args.learning_rate
)
logger.debug(f"Using optimizer: {self.optimizer}")
self.dataset_val = args.dataset_name_val
# the validation loader is a train loader but without data augmentation!
self.val_n_img, self.val_loader = prepare_dataloader(
root_dir=args.data_dir,
filenames_file=args.val_filenames_file,
mode="val",
augment_parameters=None,
do_augmentation=False,
shuffle=False,
shuffle_before=True,
batch_size=args.batch_size,
size=(args.input_height, args.input_width),
num_workers=args.num_workers,
dataset=self.dataset_val,
pin_memory=args.pin_memory,
)
logging.info(f"Using a validation set with {self.val_n_img} images")
# Load data
self.output_dir = args.output_dir
self.input_height = args.input_height
self.input_width = args.input_width
self.dataset_train = args.dataset_name_train
self.n_img, self.loader = prepare_dataloader(
root_dir=args.data_dir,
filenames_file=args.filenames_file,
mode="train",
augment_parameters=args.augment_parameters,
do_augmentation=args.do_augmentation,
shuffle=not args.overfit,
shuffle_before=False,
batch_size=args.batch_size,
size=(args.input_height, args.input_width),
num_workers=args.num_workers,
dataset=self.dataset_train,
pin_memory=args.pin_memory,
)
logging.info(
f"Using a training data set from {self.dataset_train} with {self.n_img} images"
)
if "cuda" in self.device:
torch.cuda.synchronize()
def _get_model(self, args) -> torch.nn.Module:
"""
Get the model. Automatically moves model to specified device(s).
Args:
args: Experiment args
Returns:
Model
"""
# Determine device
if args.cuda_device_ids[0] == -2:
self.device = "cpu"
logger.info("Running experiment on the CPU ...")
else:
self.device = f"cuda:{args.cuda_device_ids[0]}"
# Get model
self.model = get_model(
model=args.model,
n_input_channels=args.input_channels,
pretrained=args.imagenet_pretrained,
args=self.args,
)
logger.info(
"Training a {} model with {} parameters".format(
args.model, sum(p.numel() for p in self.model.parameters())
)
)
# Load pretrained model
if args.checkpoint:
logging.info(f"Loading pretrained model from {args.checkpoint}")
self.load(args.checkpoint)
self.multi_gpu = len(args.cuda_device_ids) > 1 or args.cuda_device_ids[0] == -1
# Check if multiple cuda devices are selected
if self.multi_gpu:
num_cuda_devices = torch.cuda.device_count()
if args.cuda_device_ids[0] == -1:
# Select all devices
cuda_device_ids = list(range(num_cuda_devices))
else:
cuda_device_ids = args.cuda_device_ids
# Check if multiple cuda devices are available
if num_cuda_devices > 1:
logger.info(
f"Running experiment on the following GPUs: {cuda_device_ids}"
)
# Transform model into data parallel model on all selected cuda deviecs
self.model = torch.nn.DataParallel(
self.model, device_ids=cuda_device_ids
)
else:
logger.warning(
f"Attempted to run the experiment on multiple GPUs while only {num_cuda_devices} GPU was available"
)
logger.debug(f"Sending model to device: {self.device}")
return self.model.to(self.device)
def train(self) -> None:
""" Train the model for self.args.epochs epochs
Returns:
None
"""
train_start_time = time.time()
# Store the best validation loss
best_val_loss = float("Inf")
# Start training
logger.info(
f"Starting training for {self.args.epochs} epochs on {self.n_img} images"
)
for epoch in range(1, self.args.epochs + 1):
# Adjust learning rate if flag is set
if self.args.adjust_lr:
adjust_learning_rate(self.optimizer, epoch, self.args.learning_rate)
epoch_time = time.time()
# Init training loss
running_loss = 0.0
running_image_loss = 0.0
running_disp_gradient_loss = 0.0
running_lr_loss = 0.0
# Init validation loss
running_val_loss = 0.0
running_val_image_loss = 0.0
running_val_disp_gradient_loss = 0.0
running_val_lr_loss = 0.0
self.model.train()
#################
# Training loop #
#################
for iteration, data in enumerate(self.loader):
# Load data
data = to_device(data, self.device)
left = data["left_image"]
right = data["right_image"]
# One optimization iteration
self.optimizer.zero_grad()
disps = self.model(left)
loss, image_loss, disp_gradient_loss, lr_loss = self.loss_function(
disps, [left, right]
)
loss.backward()
self.optimizer.step()
# Collect training loss
running_loss += loss.item()
running_image_loss += image_loss.item()
running_disp_gradient_loss += disp_gradient_loss.item()
running_lr_loss += lr_loss.item()
# Stop after 10 batches if overfitting is enabled
if self.args.overfit and iteration >= 5:
break
# Training finished #
logger.info(
f"Epoch [{epoch}/{self.args.epochs}] time: {time_delta_now(epoch_time)} s"
)
###################
# Validation loop #
###################
self.model.eval()
val_time = time.time()
with torch.no_grad():
for iteration, data in enumerate(self.val_loader):
data = to_device(data, self.device)
left = data["left_image"]
right = data["right_image"]
disps = self.model(left)
loss, image_loss, disp_gradient_loss, lr_loss = self.loss_function(
disps, [left, right]
)
# Collect validation loss
running_val_loss += loss.item()
running_val_image_loss += image_loss.item()
running_val_disp_gradient_loss += disp_gradient_loss.item()
running_val_lr_loss += lr_loss.item()
# Stop after 10 batches if overfitting is enabled
if self.args.overfit and iteration >= 5:
break
logger.info(f"Validation took {time_delta_now(val_time)}s")
#################
# Track results #
#################
running_val_loss /= len(self.val_loader)
running_val_image_loss /= len(self.val_loader)
running_val_disp_gradient_loss /= len(self.val_loader)
running_val_lr_loss /= len(self.val_loader)
# Generate 10 random disparity map predictions
self.gen_val_disp_maps(epoch)
# Estimate loss per image
running_loss /= len(self.loader)
running_image_loss /= len(self.loader)
running_disp_gradient_loss /= len(self.loader)
running_lr_loss /= len(self.loader)
# Update best loss
if running_val_loss < best_val_loss:
best_val_loss = running_val_loss
self.summary.add_epoch_metric(
epoch=epoch,
train_metric=running_loss,
val_metric=running_val_loss,
metric_name=self.loss_names["full"],
)
self.summary.add_epoch_metric(
epoch=epoch,
train_metric=running_image_loss,
val_metric=running_val_image_loss,
metric_name=self.loss_names["images"],
)
self.summary.add_epoch_metric(
epoch=epoch,
train_metric=running_disp_gradient_loss,
val_metric=running_val_disp_gradient_loss,
metric_name=self.loss_names["disp_gp"],
)
self.summary.add_epoch_metric(
epoch=epoch,
train_metric=running_lr_loss,
val_metric=running_val_lr_loss,
metric_name=self.loss_names["lr_consistency"],
)
self.summary.add_checkpoint(
model=self.model, val_loss=running_val_loss, multi_gpu=self.multi_gpu
)
logging.info(f"Finished Training. Best loss: {best_val_loss}")
self.summary.save()
# Store best validation loss
self.best_val_loss = best_val_loss
self.training_time = time_delta_now(train_start_time)
def gen_val_disp_maps(self, epoch: int):
"""
Generate n validation disparity maps
Args:
epoch (int): Current epoch
"""
gen_count = 0
n_gen_images = 25
with torch.no_grad():
for (i, data) in enumerate(self.val_loader):
# Stop after n_gen_images
if gen_count >= n_gen_images:
break
# Get the inputs
data = to_device(data, self.device)
left = data["left_image"]
# Do a forward pass
disps = self.model(left)
while gen_count < n_gen_images and gen_count < self.args.batch_size:
batch_idx = gen_count % self.args.batch_size
largest_disp_map = disps[0]
image_in_batch = largest_disp_map[batch_idx]
left_disp = image_in_batch[0]
d = left_disp.cpu().numpy()
self.summary.add_disparity_map(
epoch=epoch,
disp=torch.Tensor(d),
idx=gen_count,
input_img=left[batch_idx],
)
gen_count += 1
def save(self, path: str) -> None:
""" Save a .pth state dict from self.model
Args:
path: path to .pth state dict file
Returns:
None
"""
if self.multi_gpu:
torch.save(self.model.module.state_dict(), path)
else:
torch.save(self.model.state_dict(), path)
def load(self, path: str) -> None:
""" Load a .pth state dict into self.model
Args:
path: path to .pth state dict file
Returns:
None
"""
self.model.load_state_dict(torch.load(path, map_location=self.device))
def test(self):
# Run test
self.test_result, self.test_result_pp = run_test(
model=self.model,
args=self.args,
device=self.device,
result_dir=os.path.join(self.base_dir, "test"),
)
def adjust_learning_rate(
optimizer: torch.optim.Optimizer, epoch: int, learning_rate: float
):
""" Sets the learning rate to the initial LR\
decayed by factor 0.5 every 10 epochs after 30 epochs
Args:
optimizer: torch.optim type optimizer
epoch: current epoch
learning_rate: current learning rate
"""
if epoch >= 30:
lr = learning_rate * (0.5 ** (epoch // 10))
else:
lr = learning_rate
for param_group in optimizer.param_groups:
param_group["lr"] = lr