-
Notifications
You must be signed in to change notification settings - Fork 7
/
Copy pathtrain.py
225 lines (190 loc) · 10.9 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
import argparse
import torch
import datetime
import os
import yaml
from monai.utils import set_determinism
from random import randint
from models.model import define_model
from models.networks import init_weights
import time
from shutil import copyfile
from copy import deepcopy
from rich.live import Live
from rich.progress import Progress, TimeElapsedColumn
from rich.spinner import Spinner
from rich.console import Group
group = Group()
from data.image_dataset import get_dataset, get_post_transformation
from utils.metrics import MetricsManager
from utils.visualizer import Visualizer, DynamicDisplay
from models.base_model_abc import BaseModelABC
from utils.enums import Phase
def train(args: argparse.Namespace, config: dict[str,dict]):
global group
for phase in Phase:
if phase not in config:
continue
for k in config[phase]["data"].keys():
if not config[phase]["data"][k].get("split", ".txt").endswith(".txt"):
assert bool(args.split), "You have to specify a split!"
config[phase]["data"][k]["split"] = config[phase]["data"][k]["split"] + args.split + ".txt"
max_epochs = config[Phase.TRAIN]["epochs"]
val_interval = config[Phase.TRAIN].get("val_interval") or 1
save_interval = config[Phase.TRAIN].get("save_interval") or 100
# use amp to accelerate training
scaler = torch.cuda.amp.GradScaler(enabled=bool(config["General"].get("amp")))
device = torch.device(config["General"].get("device") or "cpu")
visualizer = Visualizer(config, args.start_epoch>0, epoch=args.epoch)
train_loader = get_dataset(config, Phase.TRAIN, num_workers=args.num_workers)
post_transformations_train = get_post_transformation(config, Phase.TRAIN)
if Phase.VALIDATION in config:
val_loader = get_dataset(config, Phase.VALIDATION, num_workers=args.num_workers)
post_transformations_val = get_post_transformation(config, Phase.VALIDATION)
else:
val_loader = None
print("No validation config. Skipping validation steps.")
with DynamicDisplay(group, Spinner("bouncingBall", text="Loading training data...")):
init_mini_batch = next(iter(train_loader))
input_key = [k for k in init_mini_batch.keys() if not k.endswith("_path")][0]
init_mini_batch["image"] = init_mini_batch[input_key]
with DynamicDisplay(group, Spinner("bouncingBall", text="Initializing model...")):
model: BaseModelABC = define_model(deepcopy(config), phase = Phase.TRAIN)
model.initialize_model_and_optimizer(init_mini_batch, init_weights, config, args, scaler, phase=Phase.TRAIN)
visualizer.save_model_architecture(model, init_mini_batch["image"].to(device, non_blocking=True) if init_mini_batch else None)
metrics = MetricsManager(phase=Phase.TRAIN)
if args.start_epoch>0:
best_metric, best_metric_epoch = visualizer.get_max_of_metric("metric", metrics.get_comp_metric(Phase.VALIDATION))
else:
best_metric = -1
best_metric_epoch = -1
total_start = time.time()
progress = Progress(*Progress.get_default_columns(), TimeElapsedColumn(), speed_estimate_period=300)
epochs = range(args.start_epoch, max_epochs)
with DynamicDisplay(group, progress):
progress.add_task("Epochs", total=len(epochs))
for epoch in epochs:
epoch_metrics: dict[str, dict[str, float]] = dict()
epoch_metrics["loss"] = dict()
model.train()
epoch_loss = 0
step = 0
save_best = False
# TRAINING LOOP
progress.add_task("Train Batch", total=len(train_loader))
for mini_batch in train_loader:
step += 1
outputs, losses = model.perform_training_step(mini_batch, scaler, post_transformations_train, device)
with torch.cuda.amp.autocast():
model.compute_metric(outputs, metrics)
for loss_name, loss in losses.items():
if f"train_{loss_name}" in epoch_metrics["loss"]:
epoch_metrics["loss"][f"train_{loss_name}"] += loss
else:
epoch_metrics["loss"][f"train_{loss_name}"] = loss
main_loss = list(losses.keys())[0]
epoch_loss += losses[main_loss]
progress.update(task_id=1, advance=1, description=f"train {main_loss}: {losses[main_loss]:.4f}")
for lr_scheduler in model.lr_schedulers:
lr_scheduler.step()
epoch_metrics["loss"] = {k: v/step for k,v in epoch_metrics["loss"].items()}
epoch_metrics["metric"] = metrics.aggregate_and_reset(prefix=Phase.TRAIN)
epoch_loss /= step
if args.save_latest or save_best or (epoch + 1) % save_interval == 0:
with DynamicDisplay(group, Spinner("bouncingBall", text="Saving training visuals...")):
train_sample_path = model.plot_sample(
visualizer,
mini_batch,
outputs,
suffix="train_latest"
)
progress.remove_task(1)
# VALIDATION
if val_loader is not None and (epoch + 1) % val_interval == 0:
model.eval()
val_loss = 0
progress.add_task("Validation Batch", total=len(val_loader))
with torch.no_grad():
step = 0
for val_mini_batch in val_loader:
step += 1
with torch.cuda.amp.autocast():
outputs, losses = model.inference(val_mini_batch, post_transformations_val, device=device, phase=Phase.VALIDATION)
model.compute_metric(outputs, metrics)
for loss_name, loss in losses.items():
if f"val_{loss_name}" in epoch_metrics["loss"]:
epoch_metrics["loss"][f"val_{loss_name}"] += loss.item()
else:
epoch_metrics["loss"][f"val_{loss_name}"] = loss.item()
main_loss = list(losses.keys())[0]
val_loss += losses[main_loss].item()
progress.update(task_id=2, advance=1, description=f"val {main_loss}: {losses[main_loss].item():.4f}")
epoch_metrics["loss"] = {k: v/step if k.startswith("val_") else v for k,v in epoch_metrics["loss"].items()}
epoch_metrics["metric"].update(metrics.aggregate_and_reset(prefix=Phase.VALIDATION))
val_loss /= step
metric_comp = epoch_metrics["metric"][metrics.get_comp_metric(Phase.VALIDATION)]
if metric_comp > best_metric:
best_metric = metric_comp
best_metric_epoch = epoch
save_best = True
if args.save_latest or save_best or (epoch + 1) % save_interval == 0:
with DynamicDisplay(group, Spinner("bouncingBall", text="Saving validation visuals...")):
val_sample_path = model.plot_sample(
visualizer,
val_mini_batch,
outputs,
suffix="val_latest"
)
progress.remove_task(2)
if (epoch + 1) % save_interval == 0:
copyfile(train_sample_path, train_sample_path.replace("latest", str(epoch+1)))
if val_loader is not None and (epoch + 1) % val_interval == 0:
copyfile(val_sample_path, val_sample_path.replace("latest", str(epoch+1)))
if save_best:
copyfile(train_sample_path, train_sample_path.replace("latest", "best"))
copyfile(val_sample_path, val_sample_path.replace("latest", "best"))
# Checkpoint saving
if args.save_latest or save_best or (epoch + 1) % save_interval == 0:
with DynamicDisplay(group, Spinner("bouncingBall", text="Saving checkpoints...")):
for optimizer_name in model.optimizer_mapping.keys():
checkpoint_path = visualizer.save_model(None, getattr(model,optimizer_name), epoch+1, config, f"latest_{optimizer_name}")
if (epoch + 1) % save_interval == 0:
copyfile(checkpoint_path, checkpoint_path.replace("latest", str(epoch+1)))
if save_best:
copyfile(checkpoint_path, checkpoint_path.replace("latest", "best"))
for model_names in model.optimizer_mapping.values():
for model_name in model_names:
checkpoint_path = visualizer.save_model(getattr(model,model_name), None, epoch+1, config, f"latest_{model_name}")
if (epoch + 1) % save_interval == 0:
copyfile(checkpoint_path, checkpoint_path.replace("latest", str(epoch+1)))
if save_best:
copyfile(checkpoint_path, checkpoint_path.replace("latest", "best"))
with DynamicDisplay(group, Spinner("bouncingBall", text="Saving metrics...")):
visualizer.plot_losses_and_metrics(epoch_metrics, epoch)
visualizer.log_model_params(model, epoch)
progress._task_index = 1
progress.advance(task_id=0, advance=1)
total_time = time.time() - total_start
print(f"Finished training after {str(datetime.timedelta(seconds=total_time))}.")
if best_metric_epoch > -1:
print(f'Best metric: {best_metric} at epoch: {best_metric_epoch}.')
if __name__ == "__main__":
# Parse input arguments
parser = argparse.ArgumentParser(description='')
parser.add_argument('--config_file', type=str, required=True)
parser.add_argument('--start_epoch', type=int, default=0)
parser.add_argument('--epoch', type=str, default='latest')
parser.add_argument('--split', type=str, default='')
parser.add_argument('--save_latest', type=bool, default=True, help="If true, save a checkpoint and visuals after each epoch under the tag 'latest'.")
parser.add_argument('--num_workers', type=int, default=None, help="Number of cpu cores for dataloading. If not set, use half of available cores.")
args = parser.parse_args()
# Read config file
path: str = os.path.abspath(args.config_file)
assert os.path.isfile(path), f"Your provided config path {args.config_file} does not exist!"
with open(path, "r") as stream:
config: dict[str,dict] = yaml.safe_load(stream)
if "seed" not in config["General"]:
config["General"]["seed"] = randint(0,1e6)
set_determinism(seed=config["General"]["seed"])
with Live(group, refresh_per_second=10):
train(args, config)