-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
287 lines (233 loc) · 11.2 KB
/
train.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
from tqdm import tqdm
import numpy as np
import os, shutil
from options.train_options import TrainOptions
from dataloader.data_loader import dataloader_full
from model.models import create_model
from utils.evaluate import get_dict_motion_category, train_evaluate
from utils.util import print_current_errors
import torch
from utils.util import RunningAverageDict
from torch.utils.tensorboard import SummaryWriter
import math
def prepare_summary(opt, clear_summary=False, purge_step=None):
summary_dir = os.path.join(opt.log_dir, opt.experiment_name, 'summary')
if clear_summary:
if os.path.exists(summary_dir) and os.path.isdir(summary_dir):
test_result = os.path.join(opt.log_dir, opt.experiment_name, 'test_result.txt')
if os.path.exists(test_result):
old_summary_idx = 0
old_summary_dir = summary_dir + '_' + str(old_summary_idx)
while os.path.exists(old_summary_dir) and os.path.isdir(old_summary_dir):
old_summary_idx += 1
old_summary_dir = summary_dir + '_' + str(old_summary_idx)
shutil.move(summary_dir, old_summary_dir)
test_result = os.path.join(opt.log_dir, opt.experiment_name, 'test_result.txt')
old_test_result = test_result[:-4] + '_' + str(old_summary_idx) + ".txt"
shutil.move(test_result, old_test_result)
else:
shutil.rmtree(summary_dir)
writer = SummaryWriter(log_dir=summary_dir, purge_step=purge_step)
return writer
def record_dataset_information():
dataset_log_dir = os.path.join(opt.log_dir, opt.experiment_name, 'dataset')
if os.path.exists(dataset_log_dir) and os.path.isdir(dataset_log_dir):
shutil.rmtree(dataset_log_dir)
os.makedirs(dataset_log_dir, exist_ok=True)
mod_dataset_path = os.path.join(opt.data_dir, "modify_dataset_log.txt")
if os.path.exists(mod_dataset_path):
shutil.copy(mod_dataset_path, os.path.join(dataset_log_dir, "modify_dataset_log.txt"))
script_path = os.path.join(opt.data_dir, "script.py")
if os.path.exists(script_path):
shutil.copy(script_path, os.path.join(dataset_log_dir, "script.py"))
def test_model(opt, model):
test_dataset = dataloader_full(opt, mode='test')
print('test images = {}'.format(len(test_dataset) * opt.batch_size))
print("\n")
print("load best model ...")
metrics_test = train_evaluate(opt, model, test_dataset, "best")
print("best test metrics:")
for k, v in metrics_test.items():
print("{}: {:.4e}".format(k, v.item()))
return metrics_test
def train_main(opt, checkpoint_dir=None, ray_config=None):
opt.use_ray = ray_config is not None
if ray_config is not None:
from ray import tune
for k, v in ray_config.items():
opt.__setattr__(k, v)
print("preparing dataset ... ")
train_dataset = dataloader_full(opt, mode='train')
val_dataset = dataloader_full(opt, mode='validation')
opt.epoch_iter_cnt = len(train_dataset)
print('train images = {}'.format(len(train_dataset) * opt.batch_size))
print('validation images = {}'.format(len(val_dataset) * opt.batch_size))
model = create_model(opt)
# model = torch.compile(model)
total_steps=0
current_best_metrics = np.inf
best_metrics = None
if not opt.use_ray:
writer = prepare_summary(opt, clear_summary=(opt.epoch_count==1))
record_dataset_information()
print('---------------------Start Training-----------------------')
model.train()
# Continue training from epoch
if checkpoint_dir is not None:
model.load_networks(checkpoint_path=checkpoint_dir)
if opt.epoch_count > 1:
model.load_networks(which_epoch=opt.epoch_count-1)
loss_records = {}
if opt.use_ray:
opt.use_slurm = True
total_itr = 0
epoch = opt.epoch_count
while epoch <= opt.niter+opt.niter_decay:
n_trained_sample = 0
print('-----------------Train Epoch: {}-----------------'.format(str(epoch)))
curr_loss = {}
if not opt.use_slurm:
bar_train = tqdm(enumerate(train_dataset), total=len(train_dataset), desc=f"Epoch: {epoch}", position=0, leave=True, dynamic_ncols=True)
else:
bar_train = enumerate(train_dataset)
total_loss = RunningAverageDict()
restart_epoch = False
# training
for i, data in bar_train:
total_steps += 1
n_trained_sample += opt.batch_size
model.set_input(data)
model.optimize_parameters()
if 'cos_anneal' in opt.lr_policy:
model.update_learning_rate()
if not opt.use_ray:
writer.add_scalar(f"Batch/lr", model.optimizers[0].param_groups[0]['lr'], i + len(train_dataset) * (epoch - 1))
curr_itr = total_itr + i
total_loss.update(model.get_current_errors())
for k, v in model.get_current_errors().items():
if math.isnan(v):
if opt.use_ray:
tune.report(loss=float('nan'), should_terminate=True)
else:
print("{} loss is NaN!".format(k))
model.save_networks('nan')
if epoch > 1:
model.load_networks(which_epoch=epoch-1)
restart_epoch = True
break
else:
if opt.auto_terminate:
return True
return False
if math.isinf(v):
if opt.use_ray:
tune.report(loss=float('inf'), should_terminate=True)
else:
print("{} loss is Inf!".format(k))
model.save_networks('inf')
if epoch > 1:
model.load_networks(which_epoch=epoch-1)
restart_epoch = True
break
else:
if opt.auto_terminate:
return True
return False
check_itr = 3000 if "Heatmap" in model.name() else 8000
if opt.auto_restart and curr_itr < check_itr:
if k not in loss_records:
loss_records[k] = (curr_itr, v)
else:
if v < loss_records[k][1]:
loss_records[k] = (curr_itr, v)
else:
threshold = 200 if "Heatmap" in model.name() else 400
if curr_itr - loss_records[k][0] > threshold:
print("Early convergence detected at: {} at {} for {}!".format(i, v, k))
if opt.auto_restart:
return False
if not opt.use_ray:
writer.add_scalar(f"Batch/{k}", v, i + len(train_dataset) * (epoch - 1))
if restart_epoch:
break
curr_loss = list(model.get_current_errors().values())
curr_loss = ''.join(['%.3E ' % v for v in curr_loss])
if not opt.use_slurm:
bar_train.set_description(f"Epoch: {epoch}, Error: {curr_loss}")
data = None
if restart_epoch:
continue
if (epoch % opt.val_epoch_freq == 0):
print('-----------------Validation Epoch: {}-----------------'.format(str(epoch)))
metrics = train_evaluate(opt, model, val_dataset, epoch)
for k, v in metrics.items():
if not opt.use_ray:
writer.add_scalar(f"Validation/{k}", v, epoch)
metric_string = ' '.join(['%s: %.4E' % (k, v) for k, v in metrics.items()])
print(metric_string)
if metrics['{}'.format(model.eval_key)] < current_best_metrics:
current_best_metrics = metrics['{}'.format(model.eval_key)]
if not opt.use_ray:
model.save_networks('best')
best_metrics = metrics
if opt.use_ray:
with tune.checkpoint_dir(epoch) as checkpoint_dir:
model.save_networks(None, checkpoint_path=checkpoint_dir)
tune.report(loss=metrics[opt.tune_criteria].cpu().numpy().item())
if epoch % opt.print_epoch_freq == 0:
print_current_errors(epoch, n_trained_sample, total_loss.get_value(), epoch)
for k, v in total_loss.get_value().items():
if not opt.use_ray:
writer.add_scalar(f"Train/{k}", v, epoch)
if not opt.use_ray:
if epoch % opt.save_epoch_freq == 0:
model.save_networks(epoch)
# Note that this part has dependency with the get_scheduler
if 'cos_anneal' not in opt.lr_policy:
model.update_learning_rate()
total_itr += len(train_dataset.dataset)
print('dir name: {}'.format(opt.experiment_name))
epoch += 1
print("\n")
print("train finished !!!")
if not opt.use_ray:
writer.close()
print("\n")
print("best validation metrics: {}".format(best_metrics))
print("\n")
print('-----------------Test Best Model-----------------')
model.load_networks("best")
metrics_test = test_model(opt, model)
print("\n")
print("test finished !!!")
print("\n")
test_result_path = os.path.join(opt.log_dir, opt.experiment_name, "test_result.txt")
test_result_file = open(test_result_path, "w")
for k, v in metrics_test.items():
test_result_file.write("{}: {:.4e}".format(k, v.item()))
print("\n")
print('-----------------Start Category-Specific Evaluation-----------------')
print("\n")
def print_and_write(string):
test_result_file.write(string + "\n")
dict_motion_category = get_dict_motion_category()
for key, value in dict_motion_category.items():
key_test_dataset = dataloader_full(opt, mode="test", id=key)
if len(key_test_dataset) == 0:
print("{}:{} Test Dataset is Empty!".format(key, value))
continue
key_metrics_test = train_evaluate(opt, model, key_test_dataset, "best_" + key)
print_and_write("category: {}".format(key + "_" + value))
print_and_write("number of batches: {}".format(len(key_test_dataset)))
for k, v in key_metrics_test.items():
print_and_write("{}: {}".format(k, v))
print("\n")
print('-----------------All Process Finished-----------------')
print("\n")
return True
if __name__ == '__main__':
opt = TrainOptions().parse()
while True:
result = train_main(opt)
if result:
break