-
Notifications
You must be signed in to change notification settings - Fork 2
/
train.py
288 lines (221 loc) · 11.5 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
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
import os
import numpy as np
import shutil
import time
import json
import torch.nn as nn
import torch.nn.functional as F
from torch.autograd import Variable
import torchvision as tv
import nntools as nt
import torch
import torch.utils.data as data
from models import *
from preprocess import *
images_dir = '/datasets/ee285f-public/VQA2017/'
q_dir = '/datasets/ee285f-public/VQA2017/v2_OpenEnded_mscoco_'
ans_dir = '/datasets/ee285f-public/VQA2017/v2_mscoco_'
train_set = MSCOCODataset(images_dir, q_dir,
ans_dir, mode='train',
image_size=(224, 224))
def collate_fn(batch):
# function to sort each batch from largest question sequence to smallest (needed for LSTM)
batch.sort(key=lambda x : x[2], reverse=True)
return data.dataloader.default_collate(batch)
class SANExperiment():
def __init__(self, train_set, output_dir, batch_size=200, num_epochs=10, early_stopping=False):
self.device = 'cuda' if torch.cuda.is_available() else 'cpu'
self.train_set = train_set
self.early_stopping = early_stopping
torch.backends.cudnn.benchmark = False
self.indices = np.random.permutation(len(self.train_set))
self.indices = self.indices[:int(len(self.indices)*0.5)]
# train and validation splits
train_ind = self.indices[:int(len(self.indices)*0.8)]
val_ind = self.indices[int(len(self.indices)*0.8):]
train_sampler = torch.utils.data.sampler.SubsetRandomSampler(train_ind)
val_sampler = torch.utils.data.sampler.SubsetRandomSampler(val_ind)
self.train_loader = torch.utils.data.DataLoader(self.train_set, batch_size=batch_size,
sampler=train_sampler,
collate_fn=collate_fn)
self.val_loader = torch.utils.data.DataLoader(self.train_set, batch_size=batch_size,
sampler=val_sampler,
collate_fn=collate_fn)
# init model
self.image_model = VGGNet(output_features=1024).to(self.device)
self.question_model = LSTM(vocab_size=len(self.train_set.vocab_q), embedding_dim=1000,
batch_size=batch_size, hidden_dim=1024).to(self.device)
self.attention = AttentionNet(num_classes=1000, batch_size=batch_size,
input_features=1024, output_features=512).to(self.device)
self.optimizer_parameter_group = [{'params': self.question_model.parameters()},
{'params': self.image_model.parameters()},
{'params': self.attention.parameters()}]
# loss function and optimizer
self.criterion = nn.CrossEntropyLoss()
self.optimizer = torch.optim.RMSprop(self.optimizer_parameter_group,
lr=4e-4, alpha=0.99, eps=1e-8, momentum=0.9)
self.total_ex = len(train_ind)
self.batch_size = batch_size
self.num_epochs = num_epochs
self.history = []
self.train_loss = []
self.train_acc = []
os.makedirs(output_dir, exist_ok=True)
self.checkpoint_path = os.path.join(output_dir,
"checkpoint.pth.tar")
self.config_path = os.path.join(output_dir, "config.txt")
# Transfer all local arguments/variables into attributes
locs = {k: v for k, v in locals().items() if k is not 'self'}
self.__dict__.update(locs)
if os.path.isfile(self.config_path):
with open(self.config_path, 'r') as f:
if f.read()[:-1] != repr(self):
raise ValueError(
"Cannot create this experiment: "
"I found a checkpoint conflicting with the current setting.")
self.load()
else:
self.save()
@property
def epoch(self):
return len(self.history)
def setting(self):
return {'ImageModel': self.image_model,
'QuestionModel' : self.question_model,
'AttentionModel' : self.attention,
'Train Set': self.train_set,
'Optimizer': self.optimizer,
'BatchSize': self.batch_size}
def __repr__(self):
"""Pretty printer showing the setting of the experiment. This is what
is displayed when doing ``print(experiment)``. This is also what is
saved in the ``config.txt`` file.
"""
string = ''
for key, val in self.setting().items():
string += '{}({})\n'.format(key, val)
return string
def state_dict(self):
"""Returns the current state of the experiment."""
return {'ImageModel': self.image_model.state_dict(),
'QuestionModel' : self.question_model.state_dict(),
'AttentionModel' : self.attention.state_dict(),
'Optimizer': self.optimizer.state_dict(),
'History': self.history,
'TrainLoss' : self.train_loss,
'TrainAcc' : self.train_acc,
'Indices' : self.train_ind}
def load_state_dict(self, checkpoint):
# load from pickled checkpoint
self.image_model.load_state_dict(checkpoint['ImageModel'])
self.question_model.load_state_dict(checkpoint['QuestionModel'])
self.attention.load_state_dict(checkpoint['AttentionModel'])
self.optimizer.load_state_dict(checkpoint['Optimizer'])
self.history = checkpoint['History']
self.train_loss = checkpoint['TrainLoss']
self.train_acc = checkpoint['TrainAcc']
self.train_ind = checkpoint['Indices']
for state in self.optimizer.state.values():
for k, v in state.items():
if isinstance(v, torch.Tensor):
state[k] = v.to(self.device)
def save(self):
"""Saves the experiment on disk, i.e, create/update the last checkpoint."""
torch.save(self.state_dict(), self.checkpoint_path)
with open(self.config_path, 'w') as f:
print(self, file=f)
def load(self):
"""Loads the experiment from the last checkpoint saved on disk."""
checkpoint = torch.load(self.checkpoint_path,
map_location=self.device)
self.load_state_dict(checkpoint)
del checkpoint
def evaluate(self):
self.image_model.eval()
self.question_model.eval()
self.attention.eval()
loader = self.val_loader
loss, acc = 0.0, 0.0
with torch.no_grad():
for i, q, s, a in loader:
if (self.device == 'cuda'):
i, q, s, a = i.cuda(), q.cuda(), s.cuda(), a.cuda()
i, q, s, a = Variable(i), Variable(q), Variable(s), Variable(a, required_grad=False)
# forward prop validation image/question through model
image_embed = self.image_model(i)
question_embed = self.question_model(q.long(), s.long())
output = self.attention(image_embed, question_embed)
_, y_pred = torch.max(output, 1)
# calculate loss
loss += self.criterion(output, a.long().squeeze(dim=1)).item()
acc += torch.sum((y_pred == a.long()).data)
loss = (float(loss) / float(len(self.val_ind)))
acc = (float(acc) / float(len(self.val_ind)))
print("Validation Loss:", loss)
print("Validation Accuracy:", acc)
return acc
def run(self):
self.image_model.train()
self.question_model.train()
self.attention.train()
loader = self.train_loader
start_epoch = self.epoch
prev_acc = 0
print("Start/Continue training from epoch {}".format(start_epoch))
for epoch in range(start_epoch, self.num_epochs):
running_loss, running_acc, num_updates = 0.0, 0.0, 0.0
counter = 0
for i, q, s, a in loader:
if (self.device == 'cuda'):
i, q, s, a = i.cuda(), q.cuda(), s.cuda(), a.cuda()
i, q, s, a = Variable(i), Variable(q), Variable(s), Variable(a)
self.optimizer.zero_grad()
# forward prop
image_embed = self.image_model(i)
question_embed = self.question_model(q.long(), s.long())
output = self.attention(image_embed, question_embed)
_, y_pred = torch.max(output, 1)
try:
# calculate loss
loss = self.criterion(output, a.long().squeeze(dim=1))
except RuntimeError as e:
if 'out of memory' in str(e):
if hasattr(torch.cuda, 'emtpy_cache'):
torch.cuda.empty_cache()
else:
raise e
# backprop
loss.backward()
self.optimizer.step()
with torch.no_grad():
running_loss += loss.item()
running_acc += torch.sum((y_pred == a.long()).data)
num_updates += 1
print("Epoch: {}, Batch: {}, Loss = {}, Acc = {}".format(epoch, counter,
(float(running_loss) / float(num_updates * self.batch_size)),
(float(running_acc) / float(num_updates * self.batch_size))))
torch.cuda.empty_cache()
if (counter % 50 == 0):
acc = float(running_acc) / float(num_updates * self.batch_size)
self.history.append(epoch)
self.train_loss.append(float(running_loss) / float(num_updates * self.batch_size))
self.train_acc.append(acc)
# early stopping code, stop when validation accuracy drops
if (self.early_stopping):
val_acc = self.evaluate()
if prev_acc > val_acc:
return 0
else:
prev_acc = val_acc
self.save()
counter += 1
loss = (float(running_loss) / float(self.total_ex))
acc = (float(running_acc) / float(self.total_ex))
print("Done with Epoch {}. Loss={}, Acc={}".format(epoch, loss, acc))
self.history.append(epoch)
self.train_loss.append(loss)
self.train_acc.append(acc)
self.save()
print("Finish training for {} epochs".format(self.num_epochs))
exp = SANExperiment(output_dir="exp_batch200", train_set=train_set)
exp.run()