-
Notifications
You must be signed in to change notification settings - Fork 2
/
train.py
249 lines (217 loc) · 10.8 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
import torch
import argparse
import wandb
import os
from datetime import datetime
import time
from tqdm import tqdm
from utils import AverageMeter, UpdatableDict, convert_from_string, data_to_device, load_yaml, \
fix_seed_for_reproducability, save_config
from build import get_model_and_optimizer
from dataloaders import get_train_val_dataloaders
def parse_arguments():
parser = argparse.ArgumentParser(description='Training configuration')
parser.add_argument('--config', default=None, help='Specify a config file path')
parser.add_argument('--exp_config', default=None, help='Specify an experiment config file path')
parser.add_argument('--restore', action='store_true', help='Restore the run')
parser.add_argument('--overfit', action='store_true', help='Overfit on small batches for debugging')
args, unknown_raw = parser.parse_known_args()
unknown = []
for ur in unknown_raw:
unknown.extend(ur.split("="))
return args, unknown
def train(cfgs, logger, train_dataloader, val_dataloader, model, optimizer, checkpoint_name=None, scheduler=None):
patience = cfgs.get("patience", cfgs["max_epochs"])
curr_patience = 0
metric_names = cfgs.get("early_stopping", None)
metric_modes = cfgs.get("early_stopping_mode", ['max'])
early_stopping = False
if metric_names is not None:
early_stopping = True
best_metric_init = {}
for metric_name, metric_mode in zip(metric_names, metric_modes):
if metric_mode == "max":
best_metric_init[metric_name] = -999
else:
best_metric_init[metric_name] = 999
if cfgs["restore"]:
chkpt_name = checkpoint_name if checkpoint_name is not None else 'checkpoint_last.pth.tar'
chkpt_file = os.path.join(cfgs["experiment_dir"], chkpt_name)
if os.path.exists(chkpt_file):
chkpt = torch.load(chkpt_file)
global_step = chkpt["step"]
start_epoch = chkpt["epoch"]
if (cfgs["model"] == "dsmnet") & (cfgs.get("include_autoencoder")==True):
model.load_state_dict(chkpt["state_dict"], strict=False)
else:
model.load_state_dict(chkpt["state_dict"])
optimizer.load_state_dict(chkpt["optimizer"])
if scheduler:
scheduler.load_state_dict(chkpt["scheduler"])
if cfgs["early_stopping"] is not None:
curr_patience = chkpt.get("patience", 0)
best_metric = chkpt.get("best_metric", best_metric_init)
else:
global_step = 0
start_epoch = 0
if cfgs["early_stopping"] is not None:
best_metric = best_metric_init
curr_patience = 0
else:
global_step = 0
start_epoch = 0
if cfgs["early_stopping"] is not None:
best_metric = best_metric_init
curr_patience = 0
if curr_patience == patience:
return
for epoch in range(start_epoch, cfgs["max_epochs"]):
model.train()
loss_train = AverageMeter()
for _, image, gt in tqdm(train_dataloader, desc=f"Epoch {epoch+1}: training ..."):
global_step += 1
log_flag = (global_step % cfgs["log_interval"] == 0)
image = data_to_device(image, device=cfgs["device"])
gt = data_to_device(gt, device=cfgs["device"])
losses, pred = model(image, gt)
loss_total = losses["loss_total"]
loss_train.update(loss_total.item(), len(image))
optimizer.zero_grad()
loss_total.backward()
#torch.nn.utils.clip_grad_norm_([layer.parameters() for layer in model.model.adaptive_bins_layer.patch_transformer.transformer_encoder.layers], 0.02)
#torch.nn.utils.clip_grad_norm_(model.model.adaptive_bins_layer.patch_transformer.transformer_encoder.parameters(), 0.002)
optimizer.step()
'''
for name, grad in zip(
['conv_out.weight', 'conv_out.bias', 'regressor.last.weight', 'regressor.last.bias'],
[model.model.conv_out[0].weight.grad, model.model.conv_out[0].bias.grad, model.model.adaptive_bins_layer.regressor[4].weight.grad, model.model.adaptive_bins_layer.regressor[4].bias]):
print(name, grad.min(), grad.max())
'''
if (cfgs.get("lr_policy", "constant") != "reduceonplateau") & (scheduler is not None):
scheduler.step()
if log_flag:
log_dict = {
'step': global_step,
'epoch': epoch
}
log_dict.update({'train/'+key: loss for key, loss in losses.items()})
if scheduler:
log_dict.update({'lr': float(optimizer.param_groups[0]['lr'])})
log_dict.update(model.vis(image, pred, gt))
logger.log(log_dict)
logger.log({
'epoch': epoch,
'train/loss_avg': loss_train.avg
})
model.eval()
loss_val = AverageMeter()
eval_dict = UpdatableDict()
skip_flag = False
with torch.no_grad():
for _, image, gt in tqdm(val_dataloader, desc=f"Epoch {epoch+1}: validating ..."):
tic = time.perf_counter()
image = data_to_device(image, device=cfgs["device"])
gt = data_to_device(gt, device=cfgs["device"])
losses, pred, eval_params = model(image, gt)
if not cfgs.get("rcnn", False):
loss_val.update(losses["loss_total"].item(), len(image))
eval_dict.update(eval_params)
toc = time.perf_counter()
if toc - tic > 10:
print(f"Epoch {epoch+1}: validation takes too long, skipping ...")
skip_flag = True
break
save_dict = {
'step': global_step,
'epoch': epoch + 1,
'state_dict': model.state_dict(),
'optimizer': optimizer.state_dict(),
}
if scheduler:
save_dict.update({'scheduler': scheduler.state_dict()})
log_dict = {'epoch': epoch}
if not cfgs.get("rcnn", False):
log_dict.update({'val/loss_total': loss_val.avg})
log_dict.update(model.vis(image, pred, gt, True))
if (not skip_flag) & early_stopping:
eval_res = model.evaluate(eval_dict())
log_dict.update(eval_res)
no_stop_flag = []
for metric_name, metric_mode in zip(metric_names, metric_modes):
if ((metric_mode == 'max') & (log_dict[metric_name] > best_metric[metric_name])) | \
((metric_mode == 'min') & (log_dict[metric_name] < best_metric[metric_name])):
best_metric[metric_name] = log_dict[metric_name]
no_stop_flag.append(True)
curr_patience = 0
else:
no_stop_flag.append(False)
if (cfgs.get("lr_policy", "constant") == "reduceonplateau") & (scheduler is not None):
scheduler.step(eval_res["rmse"])
if not any(no_stop_flag):
curr_patience += 1
else:
save_dict.update({
'best_metric': best_metric,
})
for metric_name, no_stop in zip(metric_names, no_stop_flag):
if no_stop:
if '/' in metric_name:
metric_name = metric_name.split('/')[-1]
torch.save(save_dict, os.path.join(cfgs["experiment_dir"], 'checkpoint_best_{:s}.pth.tar'.format(metric_name)))
save_dict.update({
"patience": curr_patience
})
logger.log(log_dict)
if (epoch % cfgs["checkpoint_interval"]) == 0:
torch.save(save_dict, os.path.join(cfgs["experiment_dir"], 'checkpoint_{:03d}.pth.tar'.format(epoch)))
torch.save(save_dict, os.path.join(cfgs["experiment_dir"], 'checkpoint_last.pth.tar'))
if (not skip_flag) & early_stopping & (curr_patience == patience):
print(f"Epoch {epoch+1}: maximum patience reached, early stopping ...")
break
def main():
args, unknown = parse_arguments()
cfgs = {}
if not args.restore:
assert (args.config is not None) \
& (args.exp_config is not None) \
& os.path.isfile(args.config) \
& os.path.isfile(args.exp_config), "Config files should be specified and exist"
cfgs = load_yaml(args.config)
cfgs.update(load_yaml(args.exp_config))
cfgs["restore"] = False
cfgs["overfit"] = args.overfit
cfgs["checkpoint_dir"] = os.path.join(cfgs["checkpoint_dir"], cfgs["model"])
cfgs["experiment_dir"] = os.path.join(cfgs["checkpoint_dir"], datetime.now().strftime('%y%m%d_%H%M%S'))
if cfgs["overfit"]:
cfgs["log_interval"] = 1
#cfgs["early_stopping"] = None
cfgs["checkpoint_interval"] = cfgs["max_epochs"]
cfgs["batch_size"] = 1#2 if cfgs["model"] != "adabins_global" else 1
cfgs["patience"] = cfgs["max_epochs"]
print(f"Starting from {args.exp_config}...")
else:
assert (args.exp_config is not None) \
& os.path.isfile(args.exp_config), "Experiment config file should be specified and exist"
cfgs = load_yaml(args.exp_config)
print(f"Restoring from {args.exp_config}...")
cfgs['restore'] = True
if unknown:
assert (len(unknown)%2==0), "Misc variables should be in pairs, key and value"
for key, value in zip(unknown[0::2], unknown[1::2]):
cfgs[key] = convert_from_string(value)
project = cfgs.get("project", 'GBH')
runname = cfgs.get("name", None)
logger = wandb.init(project=project, entity='chen_sn', id=cfgs["wandb_run_id"], resume='must') if cfgs["restore"] else wandb.init(project=project, entity='chen_sn', name=runname)
cfgs["wandb_run_id"] = cfgs["wandb_run_id"] if cfgs["restore"] else logger.id
print(cfgs)
save_config(cfgs, os.path.join(cfgs["experiment_dir"], 'config.yaml'))
logger.config.update(cfgs, allow_val_change=True)
seed = cfgs.get("seed", 42)
fix_seed_for_reproducability(seed)
train_loader, val_loader = get_train_val_dataloaders(cfgs)
model, optimizer = get_model_and_optimizer(cfgs)
model.to(cfgs["device"])
logger.watch(model)
train(cfgs, logger, train_loader, val_loader, model, optimizer)
if __name__ == "__main__":
main()