-
Notifications
You must be signed in to change notification settings - Fork 1
/
train.py
248 lines (212 loc) · 10.3 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
import os
import time
import sys
import glob
from datetime import timedelta
from itertools import chain
from torch import load
import torch.nn as nn
import torch
from transformers import BertTokenizer, AdamW, get_cosine_schedule_with_warmup
from util import get_args, get_pytorch_device
from k_shot_testing import k_shot_testing
from tasks import *
from torch.utils.tensorboard import SummaryWriter
from models import MultiTaskLearner
from datetime import datetime
import torch.optim as optim
def train(tasks, model, args, device):
# Define logging
os.makedirs(args.save_path, exist_ok=True)
writer = SummaryWriter(
os.path.join(args.save_path, 'runs', '{}'.format(datetime.now()).replace(":", "_")))
header = ' Time Task Iteration Progress %Epoch ' + \
'Loss Dev/Loss Accuracy Dev/Acc'
log_template = '{:>10} {:>25} {:10.0f} {:5.0f}/{:<5.0f} {:5.0f}% ' + \
'{:10.6f} {:10.6f}'
dev_log_template = '{:>10} {:>25} {:10.0f} {:5.0f}/{:<5.0f} {:5.0f}%' + \
' {:10.6f} {:12.6f}'
test_template = 'Test mean: {}, Test std: {}'
print(header)
start = time.time()
# Define optimizers and loss function
optimizer_bert = AdamW(params=model.encoder.bert.parameters(), lr=args.bert_lr)
# TODO: don't access model internals, export function to get desired parameters
task_classifiers_params = [model._modules[m_name].parameters() for m_name in model._modules if 'task' in m_name]
optimizer = optim.Adam(params=chain(model.encoder.mlp.parameters(),
*task_classifiers_params),
lr=args.lr)
scheduler_bert = get_cosine_schedule_with_warmup(optimizer_bert, 200, args.num_iterations)
scheduler = get_cosine_schedule_with_warmup(optimizer, 0, args.num_iterations)
# TODO maybe find nicer solution for passing(handling) the tokenizer
print('Loading Tokenizer..')
tokenizer = BertTokenizer.from_pretrained('bert-base-uncased', do_lower_case=True)
# initialize task sampler
sampler = TaskSampler(tasks, method='random')
# Iterate over the data
train_iter = sampler.get_iter('train', tokenizer, batch_size=args.batch_size, shuffle=True)
train_iter_len = len(train_iter)
model.train()
# setup test model, task and episodes for evaluation
test_task = SentimentAnalysis(cls_dim=args.mlp_dims[-1])
test_model = type(model)(args)
test_model.add_task_classifier(test_task.get_name(), test_task.get_classifier().to(device))
output_layer_name = 'task_{}'.format(test_task.get_name())
output_layer_init = test_model._modules[output_layer_name].state_dict()
episodes = torch.load(args.episodes)
best_dev_acc = -1
iterations, running_loss = 0, 0.0
best_test_mean = -1
best_test_last = -1
convergence_tolerance_cnt = 0
for i in range(args.num_iterations):
batch = next(train_iter)
# Reset .grad attributes for weights
optimizer_bert.zero_grad()
optimizer.zero_grad()
# Extract the sentence_ids and target vector, send sentences to GPU
sentences = batch[0].to(device)
labels = batch[1]
attention_masks = batch[2].to(device)
# Feed sentences into BERT instance, compute loss, perform backward
# pass, update weights.
predictions = model(sentences, sampler.get_name(), attention_mask=attention_masks)
loss = sampler.get_loss(predictions, labels.to(device))
loss.backward()
optimizer.step()
optimizer_bert.step()
scheduler.step()
scheduler_bert.step()
running_loss += loss.item()
iterations += 1
if iterations % args.log_every == 0:
acc = sampler.calculate_accuracy(predictions, labels.to(device))
iter_loss = running_loss / args.log_every
writer.add_scalar('{}/Accuracy/train'.format(sampler.get_name()), acc, iterations)
writer.add_scalar('{}/Loss/train'.format(sampler.get_name()), iter_loss, iterations)
print(log_template.format(
str(timedelta(seconds=int(time.time() - start))),
sampler.get_name(),
iterations,
i+1, train_iter_len,
(i+1) / train_iter_len * 100,
iter_loss, acc))
running_loss = 0.0
# saving redundant parameters
# Save model checkpoints.
if iterations % args.save_every == 0:
acc = sampler.calculate_accuracy(predictions, labels.to(device))
snapshot_prefix = os.path.join(args.save_path, 'snapshot')
snapshot_path = (
snapshot_prefix +
'_acc_{:.4f}_loss_{:.6f}_iter_{}_model.pt'
).format(acc, loss.item(), iterations)
model.save_model(snapshot_path)
# Keep only the last snapshot
for f in glob.glob(snapshot_prefix + '*'):
if f != snapshot_path:
os.remove(f)
if iterations % args.eval_every == 0:
# ============================ EVALUATION ============================
dev_task_accs = []
for task in sampler.tasks:
dev_iter = task.get_iter('dev', tokenizer, batch_size=args.batch_size)
dev_iter_len = len(dev_iter)
model.eval()
# calculate accuracy on validation set
sum_dev_loss, sum_dev_acc = 0, 0
with torch.no_grad():
for dev_batch in dev_iter:
sentences = dev_batch[0].to(device)
labels = dev_batch[1]
attention_masks = dev_batch[2].to(device)
outputs = model(sentences, task.get_name(), attention_mask=attention_masks)
# Loss
batch_dev_loss = task.get_loss(outputs, labels.to(device))
sum_dev_loss += batch_dev_loss.item()
# Accuracy
acc = task.calculate_accuracy(outputs, labels.to(device))
sum_dev_acc += acc
dev_acc = sum_dev_acc / dev_iter_len
dev_loss = sum_dev_loss / dev_iter_len
print(dev_log_template.format(
str(timedelta(seconds=int(time.time() - start))),
task.get_name(),
iterations,
i+1, train_iter_len,
(i+1) / train_iter_len * 100,
dev_loss, dev_acc))
writer.add_scalar('{}/Accuracy/dev'.format(task.get_name()), dev_acc, iterations)
writer.add_scalar('{}/Loss/dev'.format(task.get_name()), dev_loss, iterations)
dev_task_accs.append(dev_acc)
mean_dev_acc = sum(dev_task_accs) / len(dev_task_accs)
if best_dev_acc < mean_dev_acc:
best_dev_acc = mean_dev_acc
snapshot_prefix = os.path.join(args.save_path, 'best_snapshot')
snapshot_path = (
snapshot_prefix +
'_acc_{:.4f}_iter_{}_model.pt'
).format(mean_dev_acc, iterations)
model.save_model(snapshot_path)
# Keep only the best snapshot
for f in glob.glob(snapshot_prefix + '*'):
if f != snapshot_path:
os.remove(f)
# evaluate in k shot fashion
test_model.encoder.load_state_dict(model.encoder.state_dict())
# ensure same output layer init for comparability
test_model._modules[output_layer_name].load_state_dict(output_layer_init)
test_mean, test_std = k_shot_testing(test_model, episodes, test_task, device,
num_test_batches=args.num_test_batches)
writer.add_scalar('{}/Acc'.format(test_task.get_name()), test_mean, iterations)
writer.add_scalar('{}/STD'.format(test_task.get_name()), test_std, iterations)
print(test_template.format(test_mean, test_std), flush=True)
if test_mean > best_test_mean:
best_test_mean = test_mean
snapshot_prefix = os.path.join(args.save_path, 'best_test_{}'.format(test_task.get_name()))
snapshot_path = (
snapshot_prefix +
'_acc_{:.5f}_iter_{}_model.pt'
).format(best_test_mean, iterations)
model.save_model(snapshot_path)
# Keep only the best snapshot
for f in glob.glob(snapshot_prefix + '*'):
if f != snapshot_path:
os.remove(f)
if test_mean > best_test_last:
best_test_last = best_test_mean
convergence_tolerance_cnt = 0
else:
convergence_tolerance_cnt += 1
if convergence_tolerance_cnt == args.convergence_tolerance:
break
writer.close()
if __name__ == '__main__':
args = get_args()
for key, value in vars(args).items():
print(key + ' : ' + str(value))
device = get_pytorch_device(args)
if args.resume_snapshot:
print("Loading models from snapshot")
model = MultiTaskLearner(args)
print("Tasks")
tasks = []
for emotion in SemEval18SingleEmotionTask.EMOTIONS:
tasks.append(SemEval18SingleEmotionTask(emotion, cls_dim=args.mlp_dims[-1]))
tasks.append(SarcasmDetection(cls_dim=args.mlp_dims[-1]))
tasks.append(OffensevalTask(cls_dim=args.mlp_dims[-1]))
for task in tasks:
model.add_task_classifier(task.get_name(), task.get_classifier().to(device))
model.load_model(args.resume_snapshot, device)
else:
model = MultiTaskLearner(args)
model.to(device)
print("Tasks")
tasks = []
for emotion in SemEval18SingleEmotionTask.EMOTIONS:
tasks.append(SemEval18SingleEmotionTask(emotion, cls_dim=args.mlp_dims[-1]))
tasks.append(SarcasmDetection(cls_dim=args.mlp_dims[-1]))
tasks.append(OffensevalTask(cls_dim=args.mlp_dims[-1]))
for task in tasks:
model.add_task_classifier(task.get_name(), task.get_classifier().to(device))
results = train(tasks, model, args, device)