-
Notifications
You must be signed in to change notification settings - Fork 0
/
train_odometry.py
361 lines (287 loc) · 13.5 KB
/
train_odometry.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
import os
import shutil
import argparse
from collections import defaultdict
from datetime import datetime
from habitat_baselines.rl.ddppo.algo.ddp_utils import get_distrib_size, init_distrib_slurm, rank0_only
from tqdm import tqdm
import torch
from torch.utils.tensorboard import SummaryWriter
from odometry.models.models import init_distributed
from odometry.utils.early_stopping import EarlyStopping
from odometry.config.default import get_config
from odometry.models import make_model
from odometry.dataset import make_dataset, make_data_loader
from odometry.losses import make_loss
from odometry.optims import make_optimizer
from odometry.metrics import make_metrics, action_id_to_action_name
from odometry.utils import set_random_seed, transform_batch
@rank0_only
def print_metrics(phase, metrics):
metrics_log_str = ' '.join([
'\t{}: {:.6f}\n'.format(k, v)
for k, v in metrics.items()
])
print(f'{phase}:\n {metrics_log_str}')
@rank0_only
def write_metrics(epoch, metrics, writer):
for metric_name, value in metrics.items():
key = 'losses' if 'loss' in metric_name else 'metrics'
writer.add_scalar(f'{key}/{metric_name}', value, epoch)
@rank0_only
def init_experiment(config):
if os.path.exists(config.experiment_dir):
def ask():
return input(f'Experiment "{config.experiment_name}" already exists. Delete (y/n)?')
answer = 'y' #ask()
while answer not in ('y', 'n'):
answer = 'y'# ask()
delete = answer == 'y'
if not delete:
exit(1)
shutil.rmtree(config.experiment_dir)
os.makedirs(config.experiment_dir)
with open(config.config_save_path, 'w') as dest_file:
config.dump(stream=dest_file)
def _all_reduce(t: torch.Tensor, device) -> torch.Tensor:
orig_device = t.device
t = t.to(device)
torch.distributed.all_reduce(t)
return t.to(orig_device)
def coalesce_post_step(metrics, device):
metric_name_ordering = sorted(metrics.keys())
stats = torch.tensor(
[metrics[k] for k in metric_name_ordering],
device="cpu",
dtype=torch.float32,
)
stats = _all_reduce(stats, device)
stats /= torch.distributed.get_world_size()
return {
k: stats[i].item() for i, k in enumerate(metric_name_ordering)
}
def train(model, optimizer, train_loader, loss_f, metric_fns, device, disable_tqdm=False, compute_metrics_per_action=True):
model.train()
num_items = 0
num_items_per_action = defaultdict(lambda: 0)
metrics = defaultdict(lambda: 0)
for data in tqdm(train_loader, disable=disable_tqdm):
data, embeddings, target = transform_batch(data)
data = data.float().to(device)
target = target.float().to(device)
for k, v in embeddings.items():
embeddings[k] = v.to(device)
output = model(data, **embeddings)
loss, loss_components = loss_f(output, target)
optimizer.zero_grad()
loss.backward()
optimizer.step()
batch_size = target.shape[0]
metrics['loss'] += loss.item() * batch_size
for loss_component, value in loss_components.items():
metrics[loss_component] += value.item() * batch_size
for metric_f in metric_fns:
metrics[metric_f.__name__] += metric_f(output, target).item() * batch_size
if compute_metrics_per_action:
for action_id in embeddings['action'].unique():
action_name = action_id_to_action_name[action_id.item()]
action_mask = embeddings['action'] == action_id
action_metric_name = f'{metric_f.__name__}_{action_name}'
num_action_items = action_mask.sum()
metrics[action_metric_name] += metric_f(output[action_mask], target[action_mask]).item() * num_action_items
num_items_per_action[action_metric_name] += num_action_items
num_items += batch_size
for metric_name in metrics:
metrics[metric_name] /= num_items_per_action.get(metric_name, num_items)
return metrics
def val(model, val_loader, loss_f, metric_fns, device, disable_tqdm=False, compute_metrics_per_action=True):
model.eval()
num_items = 0
num_items_per_action = defaultdict(lambda: 0)
metrics = defaultdict(lambda: 0)
with torch.no_grad():
for data in tqdm(val_loader, disable=disable_tqdm):
data, embeddings, target = transform_batch(data)
data = data.float().to(device)
target = target.float().to(device)
for k, v in embeddings.items():
embeddings[k] = v.to(device)
output = model(data, **embeddings)
loss, loss_components = loss_f(output, target)
batch_size = target.shape[0]
metrics['loss'] += loss.item() * batch_size
for loss_component, value in loss_components.items():
metrics[loss_component] += value.item() * batch_size
for metric_f in metric_fns:
metrics[metric_f.__name__] += metric_f(output, target).item() * batch_size
if compute_metrics_per_action:
for action_id in embeddings['action'].unique():
action_name = action_id_to_action_name[action_id.item()]
action_mask = embeddings['action'] == action_id
action_metric_name = f'{metric_f.__name__}_{action_name}'
num_action_items = action_mask.sum()
metrics[action_metric_name] += metric_f(output[action_mask], target[action_mask]).item() * num_action_items
num_items_per_action[action_metric_name] += num_action_items
num_items += batch_size
for metric_name in metrics:
metrics[metric_name] /= num_items_per_action.get(metric_name, num_items)
return metrics
def parse_args():
parser = argparse.ArgumentParser()
parser.add_argument(
'--config-file',
required=True,
type=str,
help='path to the configuration file'
)
parser.add_argument(
'--num-dataset-items',
required=False,
type=int,
default=None,
help='number of items to form a dataset'
)
parser.add_argument(
'--invert-rotations-train',
action='store_true',
help='indicates whether to invert rotation actions'
)
parser.add_argument(
'--invert-rotations-val',
action='store_true',
help='indicates whether to invert rotation actions'
)
parser.add_argument(
'--invert-collisions',
action='store_true',
help='indicates whether to invert rotation actions when the agent has collided with something'
)
parser.add_argument(
'--not-use-turn-left',
action='store_true',
)
parser.add_argument(
'--not-use-turn-right',
action='store_true',
)
parser.add_argument(
'--not-use-move-forward',
action='store_true',
)
parser.add_argument(
'--not-use-rgb',
action='store_true',
)
args = parser.parse_args()
return args
def main():
args = parse_args()
config_path = args.config_file
config = get_config(config_path, new_keys_allowed=True)
config.defrost()
config.experiment_dir = os.path.join(config.log_dir, config.experiment_name)
config.tb_dir = os.path.join(config.experiment_dir, 'tb')
config.model.best_checkpoint_path = os.path.join(config.experiment_dir, 'best_checkpoint.pt')
config.model.last_checkpoint_path = os.path.join(config.experiment_dir, 'last_checkpoint.pt')
config.config_save_path = os.path.join(config.experiment_dir, 'config.yaml')
config.train.dataset.params.num_points = args.num_dataset_items
config.train.dataset.params.invert_rotations = args.invert_rotations_train
config.train.dataset.params.invert_collisions = args.invert_collisions
config.train.dataset.params.not_use_turn_left = args.not_use_turn_left
config.train.dataset.params.not_use_turn_right = args.not_use_turn_right
config.train.dataset.params.not_use_move_forward = args.not_use_move_forward
config.train.dataset.params.not_use_rgb = args.not_use_rgb
config.val.dataset.params.num_points = args.num_dataset_items
config.val.dataset.params.invert_rotations = args.invert_rotations_val
config.val.dataset.params.invert_collisions = args.invert_collisions
config.val.dataset.params.not_use_turn_left = args.not_use_turn_left
config.val.dataset.params.not_use_turn_right = args.not_use_turn_right
config.val.dataset.params.not_use_move_forward = args.not_use_move_forward
config.val.dataset.params.not_use_rgb = args.not_use_rgb
if hasattr(config, 'train_val'):
config.train_val.dataset.params.num_points = args.num_dataset_items
config.train_val.dataset.params.invert_rotations = args.invert_rotations_val
config.train_val.dataset.params.invert_collisions = args.invert_collisions
config.train_val.dataset.params.not_use_turn_left = args.not_use_turn_left
config.train_val.dataset.params.not_use_turn_right = args.not_use_turn_right
config.train_val.dataset.params.not_use_move_forward = args.not_use_move_forward
config.train_val.dataset.params.not_use_rgb = args.not_use_rgb
config.freeze()
# init distributed if run with torch.distributed.launch
is_distributed = get_distrib_size()[2] > 1
if is_distributed:
local_rank, tcp_store = init_distrib_slurm(config.distrib_backend)
if rank0_only():
print("Initialized VO with {} workers".format(torch.distributed.get_world_size()))
config.defrost()
config.device = local_rank
config.train.loader.is_distributed = True
config.val.loader.is_distributed = True
config.freeze()
init_experiment(config)
set_random_seed(config.seed)
train_dataset = make_dataset(config.train.dataset)
train_loader = make_data_loader(config.train.loader, train_dataset)
train_metric_fns = make_metrics(config.train.metrics) if config.train.metrics else []
if hasattr(config, 'train_val'):
train_val_dataset = make_dataset(config.train_val.dataset)
train_val_loader = make_data_loader(config.train_val.loader, train_val_dataset)
train_val_metric_fns = make_metrics(config.train_val.metrics) if config.train_val.metrics else []
val_dataset = make_dataset(config.val.dataset)
val_loader = make_data_loader(config.val.loader, val_dataset)
val_metric_fns = make_metrics(config.val.metrics) if config.val.metrics else []
device = torch.device(config.device)
model = make_model(config.model).to(device)
if hasattr(config.model, 'pretrained_checkpoint') and config.model.pretrained_checkpoint is not None:
model.load_state_dict(torch.load(config.model.pretrained_checkpoint, map_location=device))
if is_distributed:
model = init_distributed(model, device, find_unused_params=True)
optimizer = make_optimizer(config.optim, model.parameters())
scheduler = None
loss_f = make_loss(config.loss)
early_stopping = EarlyStopping(
**config.stopper.params
)
# TODO: fix tensorboard logging as in PPOTrainer ??
train_writer = SummaryWriter(log_dir=os.path.join(config.tb_dir, 'train'))
val_writer = SummaryWriter(log_dir=os.path.join(config.tb_dir, 'val'))
if hasattr(config, 'train_val'):
train_val_writer = SummaryWriter(log_dir=os.path.join(config.tb_dir, 'train_val'))
for epoch in range(1, config.epochs + 1):
if rank0_only():
print(f'{datetime.now()} Epoch {epoch}')
train_metrics = train(model, optimizer, train_loader, loss_f, train_metric_fns, device, is_distributed, config.compute_metrics_per_action)
if is_distributed:
train_metrics = coalesce_post_step(train_metrics, device)
write_metrics(epoch, train_metrics, train_writer)
print_metrics('Train', train_metrics)
val_metrics = val(model, val_loader, loss_f, val_metric_fns, device, is_distributed, config.compute_metrics_per_action)
if is_distributed:
val_metrics = coalesce_post_step(val_metrics, device)
write_metrics(epoch, val_metrics, val_writer)
print_metrics('Val', val_metrics)
if hasattr(config, 'train_val'):
train_val_metrics = val(model, train_val_loader, loss_f, train_val_metric_fns, device, is_distributed, config.compute_metrics_per_action)
if is_distributed:
val_metrics = coalesce_post_step(val_metrics, device)
write_metrics(epoch, train_val_metrics, train_val_writer)
print_metrics('Train-val', train_val_metrics)
early_stopping(val_metrics['loss'])
if rank0_only() and config.model.save: # and early_stopping.counter == 0:
best_checkpoint_path = config.model.best_checkpoint_path.replace('.pt', f'_{str(epoch).zfill(3)}e.pt')
torch.save(model.state_dict(), best_checkpoint_path)
print('Saved best model checkpoint to disk.')
if early_stopping.early_stop:
print(f'Early stopping after {epoch} epochs.')
break
if scheduler:
scheduler.step()
train_writer.close()
val_writer.close()
if hasattr(config, 'train_val'):
train_val_writer.close()
if rank0_only() and config.model.save:
torch.save(model.state_dict(), config.model.last_checkpoint_path)
print('Saved last model checkpoint to disk.')
if __name__ == '__main__':
main()