-
Notifications
You must be signed in to change notification settings - Fork 49
/
train.py
225 lines (198 loc) · 8.35 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 os
import types
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.utils.data import Dataset, DataLoader, Subset
import numpy as np
from tqdm import tqdm
from utils import AverageMeter, accuracy
from loss import LossComputer
from pytorch_transformers import AdamW, WarmupLinearSchedule
def run_epoch(epoch, model, optimizer, loader, loss_computer, logger, csv_logger, args,
is_training, show_progress=False, log_every=50, scheduler=None):
"""
scheduler is only used inside this function if model is bert.
"""
if is_training:
model.train()
if args.model == 'bert':
model.zero_grad()
else:
model.eval()
if show_progress:
prog_bar_loader = tqdm(loader)
else:
prog_bar_loader = loader
with torch.set_grad_enabled(is_training):
for batch_idx, batch in enumerate(prog_bar_loader):
batch = tuple(t.cuda() for t in batch)
x = batch[0]
y = batch[1]
g = batch[2]
if args.model == 'bert':
input_ids = x[:, :, 0]
input_masks = x[:, :, 1]
segment_ids = x[:, :, 2]
outputs = model(
input_ids=input_ids,
attention_mask=input_masks,
token_type_ids=segment_ids,
labels=y
)[1] # [1] returns logits
else:
outputs = model(x)
loss_main = loss_computer.loss(outputs, y, g, is_training)
if is_training:
if args.model == 'bert':
loss_main.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
scheduler.step()
optimizer.step()
model.zero_grad()
else:
optimizer.zero_grad()
loss_main.backward()
optimizer.step()
if is_training and (batch_idx+1) % log_every==0:
csv_logger.log(epoch, batch_idx, loss_computer.get_stats(model, args))
csv_logger.flush()
loss_computer.log_stats(logger, is_training)
loss_computer.reset_stats()
if (not is_training) or loss_computer.batch_count > 0:
csv_logger.log(epoch, batch_idx, loss_computer.get_stats(model, args))
csv_logger.flush()
loss_computer.log_stats(logger, is_training)
if is_training:
loss_computer.reset_stats()
def train(model, criterion, dataset,
logger, train_csv_logger, val_csv_logger, test_csv_logger,
args, epoch_offset):
model = model.cuda()
# process generalization adjustment stuff
adjustments = [float(c) for c in args.generalization_adjustment.split(',')]
assert len(adjustments) in (1, dataset['train_data'].n_groups)
if len(adjustments)==1:
adjustments = np.array(adjustments* dataset['train_data'].n_groups)
else:
adjustments = np.array(adjustments)
train_loss_computer = LossComputer(
criterion,
is_robust=args.robust,
dataset=dataset['train_data'],
alpha=args.alpha,
gamma=args.gamma,
adj=adjustments,
step_size=args.robust_step_size,
normalize_loss=args.use_normalized_loss,
btl=args.btl,
min_var_weight=args.minimum_variational_weight)
# BERT uses its own scheduler and optimizer
if args.model == 'bert':
no_decay = ['bias', 'LayerNorm.weight']
optimizer_grouped_parameters = [
{'params': [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay)], 'weight_decay': args.weight_decay},
{'params': [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
]
optimizer = AdamW(
optimizer_grouped_parameters,
lr=args.lr,
eps=args.adam_epsilon)
t_total = len(dataset['train_loader']) * args.n_epochs
print(f'\nt_total is {t_total}\n')
scheduler = WarmupLinearSchedule(
optimizer,
warmup_steps=args.warmup_steps,
t_total=t_total)
else:
optimizer = torch.optim.SGD(
filter(lambda p: p.requires_grad, model.parameters()),
lr=args.lr,
momentum=0.9,
weight_decay=args.weight_decay)
if args.scheduler:
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(
optimizer,
'min',
factor=0.1,
patience=5,
threshold=0.0001,
min_lr=0,
eps=1e-08)
else:
scheduler = None
best_val_acc = 0
for epoch in range(epoch_offset, epoch_offset+args.n_epochs):
logger.write('\nEpoch [%d]:\n' % epoch)
logger.write(f'Training:\n')
run_epoch(
epoch, model, optimizer,
dataset['train_loader'],
train_loss_computer,
logger, train_csv_logger, args,
is_training=True,
show_progress=args.show_progress,
log_every=args.log_every,
scheduler=scheduler)
logger.write(f'\nValidation:\n')
val_loss_computer = LossComputer(
criterion,
is_robust=args.robust,
dataset=dataset['val_data'],
step_size=args.robust_step_size,
alpha=args.alpha)
run_epoch(
epoch, model, optimizer,
dataset['val_loader'],
val_loss_computer,
logger, val_csv_logger, args,
is_training=False)
# Test set; don't print to avoid peeking
if dataset['test_data'] is not None:
test_loss_computer = LossComputer(
criterion,
is_robust=args.robust,
dataset=dataset['test_data'],
step_size=args.robust_step_size,
alpha=args.alpha)
run_epoch(
epoch, model, optimizer,
dataset['test_loader'],
test_loss_computer,
None, test_csv_logger, args,
is_training=False)
# Inspect learning rates
if (epoch+1) % 1 == 0:
for param_group in optimizer.param_groups:
curr_lr = param_group['lr']
logger.write('Current lr: %f\n' % curr_lr)
if args.scheduler and args.model != 'bert':
if args.robust:
val_loss, _ = val_loss_computer.compute_robust_loss_greedy(val_loss_computer.avg_group_loss, val_loss_computer.avg_group_loss)
else:
val_loss = val_loss_computer.avg_actual_loss
scheduler.step(val_loss) #scheduler step to update lr at the end of epoch
if epoch % args.save_step == 0:
torch.save(model, os.path.join(args.log_dir, '%d_model.pth' % epoch))
if args.save_last:
torch.save(model, os.path.join(args.log_dir, 'last_model.pth'))
if args.save_best:
if args.robust or args.reweight_groups:
curr_val_acc = min(val_loss_computer.avg_group_acc)
else:
curr_val_acc = val_loss_computer.avg_acc
logger.write(f'Current validation accuracy: {curr_val_acc}\n')
if curr_val_acc > best_val_acc:
best_val_acc = curr_val_acc
torch.save(model, os.path.join(args.log_dir, 'best_model.pth'))
logger.write(f'Best model saved at epoch {epoch}\n')
if args.automatic_adjustment:
gen_gap = val_loss_computer.avg_group_loss - train_loss_computer.exp_avg_loss
adjustments = gen_gap * torch.sqrt(train_loss_computer.group_counts)
train_loss_computer.adj = adjustments
logger.write('Adjustments updated\n')
for group_idx in range(train_loss_computer.n_groups):
logger.write(
f' {train_loss_computer.get_group_name(group_idx)}:\t'
f'adj = {train_loss_computer.adj[group_idx]:.3f}\n')
logger.write('\n')