-
Notifications
You must be signed in to change notification settings - Fork 3
/
train.py
218 lines (176 loc) · 7.97 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
import torch
from torch.utils.data import DataLoader
from torch.utils.tensorboard import SummaryWriter
import torchvision
from torchvision import transforms
import torchmetrics
import os
from tqdm import tqdm
from pathlib import Path
from copy import deepcopy
from mlcpl.loss import AsymmetricLoss
from mlcpl.helper import ExcelLogger
from mlcpl.label_strategy import *
from mlcpl.metric import PartialMultilabelMetric
from models import Model
import config
# from https://github.com/Alibaba-MIIL/PartialLabelingCSL/blob/main/src/helper_functions/helper_functions.py
def add_weight_decay(model, weight_decay=1e-4, skip_list=()):
decay = []
no_decay = []
for name, param in model.named_parameters():
if not param.requires_grad:
continue # frozen weights
if len(param.shape) == 1 or name.endswith(".bias") or name in skip_list:
no_decay.append(param)
else:
decay.append(param)
return [
{'params': no_decay, 'weight_decay': 0.},
{'params': decay, 'weight_decay': weight_decay}]
# from https://github.com/Alibaba-MIIL/PartialLabelingCSL/blob/main/src/helper_functions/helper_functions.py
class ModelEma(torch.nn.Module):
def __init__(self, model, decay=0.9997, device=None):
super(ModelEma, self).__init__()
# make a copy of the model for accumulating moving average of weights
self.module = deepcopy(model)
self.module.eval()
self.decay = decay
self.device = device # perform ema on different device from model if set
if self.device is not None:
self.module.to(device=device)
def _update(self, model, update_fn):
with torch.no_grad():
for ema_v, model_v in zip(self.module.state_dict().values(), model.state_dict().values()):
if self.device is not None:
model_v = model_v.to(device=self.device)
ema_v.copy_(update_fn(ema_v, model_v))
def update(self, model):
self._update(model, update_fn=lambda e, m: self.decay * e + (1. - self.decay) * m)
def set(self, model):
self._update(model, update_fn=lambda e, m: m)
def main():
device = 'cuda'
output_dir = 'output/train'
train_transform = transforms.Compose([
transforms.RandAugment(interpolation=torchvision.transforms.functional.InterpolationMode.BICUBIC),
transforms.Resize((448, 448)),
transforms.ToTensor(),
])
valid_transform = transforms.Compose([
transforms.Resize((448, 448)),
transforms.ToTensor(),
])
train_dataset = config.train_dataset
train_dataset.transform = train_transform
train_dataset.df.to_csv('train.csv')
valid_dataset = config.valid_dataset
valid_dataset.transform = valid_transform
valid_dataset.df.to_csv('valid.csv')
num_categories = train_dataset.num_categories
model = Model(num_categories)
weight_decay = 1e-4
loss_fn = AsymmetricLoss(gamma_neg=4, gamma_pos=0, clip=0.05)
batch_size = 32
accum_step = 4
lr = 2e-4
epochs = 60
early_stopping = 10
training_metrics = {}
validation_metrics = {
'mAP': PartialMultilabelMetric(torchmetrics.functional.classification.binary_average_precision),
}
monitor_validation_metric_name = 'mAP'
monitor_mode = 'max'
### below normally fixed
log_dir = os.path.join(output_dir, 'log')
Path(output_dir).mkdir(parents=True, exist_ok=True)
Path(log_dir).mkdir(parents=True, exist_ok=True)
train_dataloader = DataLoader(train_dataset, batch_size=batch_size, num_workers=20, shuffle=True)
valid_dataloader = DataLoader(valid_dataset, batch_size=batch_size, num_workers=20, shuffle=False)
model = model.to(device)
parameters = add_weight_decay(model, weight_decay=weight_decay)
ema = ModelEma(model, 0.999)
optimizer = torch.optim.Adam(parameters, lr=lr, weight_decay=0)
steps_per_epoch = len(train_dataloader)
scheduler = torch.optim.lr_scheduler.OneCycleLR(optimizer, lr, steps_per_epoch=steps_per_epoch, epochs=epochs, pct_start=0.2)
tblog = SummaryWriter(log_dir=log_dir)
excellog = ExcelLogger(os.path.join(log_dir, 'excel_log.xlsx'))
train_dataset.df.to_csv(os.path.join(log_dir, 'training_dataset.csv'))
best_score = 0
best_at_epoch = -1
try:
for epoch in range(epochs):
print(f'Epoch: {epoch}/{epochs}')
# Train Loop
model.train()
losses, preds, targets = [], [], []
bar = tqdm(enumerate(train_dataloader), total=train_dataloader.__len__())
for batch, (x, y) in bar:
x, y = x.to(device), unknown_to_negative(y.to(device))
pred = model(x)
loss = loss_fn(pred, y)
loss = loss / accum_step
loss.backward()
if (batch + 1) % accum_step == 0 or (batch + 1) == len(train_dataloader):
optimizer.step()
scheduler.step()
ema.update(model)
model.zero_grad()
bar.set_postfix({'Loss': f'{loss.cpu().detach().numpy():.4f}'})
losses.append(loss.detach().cpu().clone())
preds.append(pred.detach().cpu().clone())
targets.append(y.detach().cpu().clone())
losses = torch.tensor(losses)
preds = torch.cat(preds)
targets = torch.cat(targets)
# Calculate metrics and logging
mean_loss = torch.mean(losses).detach().numpy()
tblog.add_scalar('loss', mean_loss, epoch)
excellog.add('loss', mean_loss)
for name, metric in training_metrics.items():
result = metric(preds, targets).detach().numpy()
if len(result.shape) == 0:
tblog.add_scalar('train_'+name, result, epoch)
excellog.add('train_'+name, result)
else:
tblog.add_scalars('train_'+name, {str(no): result[no] for no in range(num_categories)}, epoch)
excellog.add('train_'+name, {str(no): result[no] for no in range(num_categories)})
# Valid Loop
model.eval()
preds, ys = [], []
with torch.no_grad():
enumerater = tqdm(enumerate(valid_dataloader), total=valid_dataloader.__len__())
for batch, (x, y) in enumerater:
x, y = x.to(device), y.to(device)
preds.append(ema.module(x))
ys.append(y)
preds, ys = torch.cat(preds), torch.cat(ys)
# Calculate metrics and logging
for name, metric in validation_metrics.items():
result = metric(preds, ys).detach().numpy()
if len(result.shape) == 0:
tblog.add_scalar('valid_'+name, result, epoch)
excellog.add('valid_'+name, result)
print(f'{name}: {result:.4f}')
else:
tblog.add_scalars('valid_'+name, {str(no): result[no] for no in range(num_categories)}, epoch)
excellog.add('valid_'+name, {str(no): result[no] for no in range(num_categories)})
if name == monitor_validation_metric_name:
current_score = result
if (monitor_mode == 'max' and current_score > best_score) or ((monitor_mode == 'min' and current_score < best_score)):
best_score = current_score
best_at_epoch = epoch
print(f'New best {monitor_validation_metric_name}: {best_score:.4f}')
torch.save(ema.module.state_dict(), os.path.join(output_dir, 'best.pth'))
if early_stopping is not None:
if epoch - best_at_epoch >= early_stopping:
print('Early stopping.')
break
except KeyboardInterrupt:
tblog.flush()
excellog.flush()
tblog.flush()
excellog.flush()
if __name__=='__main__':
main()