-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
213 lines (164 loc) · 5.49 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
'''
Script for training the auto-encoder model.
Hyperparameters and settings are chosen in settings-file
'''
from settings import (
# DECODER
EMBEDDING_DIM,
ATTENTION_DIM,
DECODER_DIM,
DECODER_DROP_PROB,
DECODER_PAD_INDEX,
# GENERAL
BATCH_SIZE,
LEARNING_RATE,
TEATHER_FORCING_PROB,
OPTIMIZER,
WEIGHT_DECAY,
LR_STEP,
LOSS_PAD_INDEX,
NUMBER_OF_ITERATIONS,
DEVICE,
)
from DataPreparator import ImageDataset
from EncoderDecoder import DecoderWithAttention
from vocabulary import max_len, vocab_size
from CaptionCoder import deTokenizeCaptions
from evaluation import evaluate
from HelperFunctions import saveLoss, saveAccuracy
import os
import torch
import torch.nn as nn
from torch.optim.lr_scheduler import StepLR
def train_batch(
input_features, # (batch_size, feature_len)
input_sentences, # (batch_size, sentence_len)
decoder,
decoder_optimizer,
criterion,
):
global max_len
decoder_optimizer.zero_grad()
loss = 0
outputs, _ = decoder(input_sentences, input_features)
targets = input_sentences[:, 1:]
loss = criterion(outputs.permute(0, 2, 1), targets)
loss.backward()
decoder_optimizer.step()
return loss.item(), outputs[0]
def getPredictions(outputs):
predictions = outputs.argmax(-1)
return predictions
def train_loop(
trainloader,
decoder,
optimizer,
n_iters,
lr, # Probability of using target as input
setting_filename,
print_every=20,
):
decoder_file = f'{setting_filename}_decoder.pt'
decoder_model_path = os.path.join('results', 'models', decoder_file)
losses = []
train_accuracy = []
validation_accuracy = []
best_avg_update_loss = 0
decoder_optimizer = optimizer(decoder.parameters(), lr=lr, weight_decay = WEIGHT_DECAY)
scheduler = StepLR(decoder_optimizer, step_size=LR_STEP, gamma=0.1)
criterion = nn.CrossEntropyLoss(ignore_index=LOSS_PAD_INDEX)
def train_iter(iter):
nonlocal trainloader, decoder, optimizer, criterion
nonlocal decoder_model_path, print_every
nonlocal losses, train_accuracy, validation_accuracy
nonlocal best_avg_update_loss
decoder.train()
print_loss_total = 0
batch_loss_total = 0
for i, (_,_, sentences, features) in enumerate(trainloader):
# Send batch to device
sentences = sentences.to(DEVICE)
features = features.to(DEVICE)
loss, output = train_batch(
features,
sentences,
decoder,
decoder_optimizer,
criterion,
)
losses.append((iter, i, loss))
batch_loss_total += loss
print_loss_total += loss
if (i+1) % print_every == 0:
print_loss_avg = print_loss_total / print_every
print_loss_total = 0
prediction = getPredictions(output)
print(f'{print_loss_avg} : {iter}[{i}]')
print(f'Prediction: {deTokenizeCaptions(prediction)}')
print(f'Target: {deTokenizeCaptions(sentences[0, 1:])}')
print()
# Save model if average batch loss over an iteration has decreased
isFirstIteration = iter == 0
avg_update_loss = batch_loss_total / i
isBetterLoss = best_avg_update_loss > avg_update_loss
if isFirstIteration or isBetterLoss:
best_avg_update_loss = avg_update_loss
torch.save(decoder, decoder_model_path)
train_acc = evaluate(decoder, 'train')
val_acc = evaluate(decoder, 'dev')
train_accuracy.append(train_acc)
validation_accuracy.append(val_acc)
print(f'Train evaluation: {train_acc}')
print(f'Validation evaluation: {val_acc}\n')
scheduler.step()
for iter in range(n_iters):
try:
train_iter(iter)
except KeyboardInterrupt:
raise SaveFiles(losses, train_accuracy, validation_accuracy)
return losses, train_accuracy, validation_accuracy
class SaveFiles(Exception):
pass
def train_main(settings):
# Initialize
trainloader = initializeDataLoader(type='train')
decoder = initializeDecoder()
# Run training loop
try:
losses, train_acc, val_acc = train_loop(
trainloader=trainloader,
decoder=decoder,
optimizer=OPTIMIZER,
n_iters=NUMBER_OF_ITERATIONS,
lr=LEARNING_RATE,
setting_filename=settings,
)
except SaveFiles as sf:
losses, train_acc, val_acc = sf.args
finally:
saveLoss(losses, settings)
saveAccuracy(train_acc, 'train', settings)
saveAccuracy(val_acc, 'dev', settings)
pass
def initializeDecoder():
model = DecoderWithAttention(
vocab_size=vocab_size,
embedding_dim=EMBEDDING_DIM,
attention_dim=ATTENTION_DIM,
encoder_dim=512,
decoder_dim=DECODER_DIM,
device=DEVICE,
drop_prob=DECODER_DROP_PROB,
pad_token=DECODER_PAD_INDEX,
).to(DEVICE)
return model
def initializeDataLoader(type: str = 'train'):
train_dataset = ImageDataset(f'{type}_labels.txt')
trainloader = torch.utils.data.DataLoader(
train_dataset,
batch_size=BATCH_SIZE,
shuffle=True
)
return trainloader
if __name__ == '__main__':
train_main('test')