-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmain.py
177 lines (168 loc) · 8.65 KB
/
main.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
import os
import time
import torch.optim as optim
from torch.nn.utils.rnn import PackedSequence, pad_packed_sequence
from tensorboardX import SummaryWriter
from build_model import build_model
from utils import *
from collections import defaultdict
import cv2
from configuration import get_config
import dataloader
args = get_config()
device = args.device
train_loader = dataloader.train_loader(args.dataset, args.data_directory, args.batch_size, args.data_config)
test_loader = dataloader.test_loader(args.dataset, args.data_directory, args.batch_size, args.data_config)
args.label_size = train_loader.dataset.a_size
args.q_size = train_loader.dataset.q_size
args.c_size = train_loader.dataset.c_size
models = build_model(args)
if args.load_model != '000000000000':
for model_name, model in models.items():
model.load_state_dict(torch.load(os.path.join(args.log_directory + args.project, args.load_model, model_name)))
args.time_stamp = args.load_model[:12]
print('Model {} loaded.'.format(args.load_model))
def epoch(epoch_idx, is_train):
epoch_start_time = time.time()
start_time = time.time()
mode = 'Train' if is_train else 'Test'
epoch_loss = 0
q_correct = defaultdict(lambda: 0)
q_num = defaultdict(lambda: 0)
if is_train:
for model in models.values():
model.train()
loader = train_loader
else:
for model in models.values():
model.eval()
loader = test_loader
for batch_idx, (image, question, answer) in enumerate(loader):
batch_size = image.size()[0]
optimizer.zero_grad()
image = image.to(device)
answer = answer.to(device)
if args.dataset == 'clevr':
question = PackedSequence(question.data.to(device), question.batch_sizes)
else:
question = question.to(device)
# answer = answer.squeeze(1)
code = models['text_encoder.pt'](question)
if args.model == 'baseline':
objects = models['conv.pt'](image * 2 - 1)
pairs = baseline_encode(objects, code)
relations = models['g_theta.pt'](pairs)
relations = relations.sum(1)
output = models['f_phi.pt'](relations)
elif args.model == 'rn':
objects = models['conv.pt'](image * 2 - 1)
pairs = rn_encode(objects, code)
relations = models['g_theta.pt'](pairs)
relations = lower_sum(relations)
relations = relations.sum(1)
output = models['f_phi.pt'](relations)
elif args.model == 'sarn':
objects = models['conv.pt'](image * 2 - 1)
coordinate_encoded, question_encoded = sarn_encode(objects, code)
logits = models['h_psi.pt'](question_encoded)
pairs = sarn_pair(coordinate_encoded, question_encoded, logits)
relations = models['g_theta.pt'](pairs)
relations = relations.sum(1)
output = models['f_phi.pt'](relations)
elif args.model == 'sarn_att':
objects = models['conv.pt'](image * 2 - 1)
coordinate_encoded, question_encoded = sarn_encode(objects, code)
logits = models['h_psi.pt'](question_encoded)
selected = sarn_select(coordinate_encoded, logits)
relations, att = models['attn.pt'](selected, coordinate_encoded, coordinate_encoded)
relations = models['g_theta.pt'](relations)
relations = relations.sum(1)
output = models['f_phi.pt'](relations)
elif args.model == 'new':
relations = models['conv.pt'](image * 2 - 1, code)
output = models['f_phi.pt'](relations)
elif args.model == 'film':
objects = models['conv.pt'](image * 2 - 1)
output = models['film.pt'](objects, code)
loss = F.cross_entropy(output, answer)
if is_train:
loss.backward()
optimizer.step()
epoch_loss += loss.item()
pred = torch.max(output.data, 1)[1]
correct = (pred == answer)
for i in range(args.q_size):
idx = question[:, 1] == i
q_correct[i] += (correct * idx).sum().item()
q_num[i] += idx.sum().item()
if is_train:
if batch_idx % args.log_interval == 0:
print('Train Batch: {} [{}/{} ({:.0f}%)] Loss: {:.4f} / Time: {:.4f} / Acc: {:.4f}'.format(
epoch_idx,
batch_idx * batch_size, len(loader.dataset),
100. * batch_idx / len(loader),
loss.item() / batch_size,
time.time() - start_time,
correct.sum().item() / batch_size))
idx = epoch_idx * len(loader) // args.log_interval + batch_idx // args.log_interval
writer.add_scalar('Batch loss', loss.item() / batch_size, idx)
writer.add_scalar('Batch accuracy', correct.sum().item() / batch_size, idx)
writer.add_scalar('Batch time', time.time() - start_time, idx)
start_time = time.time()
else:
if batch_idx == 0:
n = min(batch_size, 4)
if args.dataset == 'clevr':
pad_question, lengths = pad_packed_sequence(question)
pad_question = pad_question.transpose(0, 1)
question_text = [' '.join([loader.dataset.idx_to_word[i] for i in q]) for q in
pad_question.cpu().numpy()[:n]]
answer_text = [loader.dataset.answer_idx_to_word[a] for a in answer.cpu().numpy()[:n]]
text = []
for j, (q, a) in enumerate(zip(question_text, answer_text)):
text.append('Quesetion {}: '.format(j) + question_text[j] + '/ Answer: ' + answer_text[j])
writer.add_image('Image', torch.cat([image[:n]]), epoch_idx)
writer.add_text('QA', '\n'.join(text), epoch_idx)
else:
image = F.pad(image[:n], (0, 0, 0, args.input_h // 3), mode='constant', value=1).transpose(1,
2).transpose(
2, 3)
image = image.cpu().numpy()
for i in range(n):
cv2.line(image[i], (args.input_w // 2, 0), (args.input_w // 2, args.input_h), (0, 0, 0), 1)
cv2.line(image[i], (0, args.input_h // 2), (args.input_w, args.input_h // 2), (0, 0, 0), 1)
cv2.line(image[i], (0, args.input_h), (args.input_w, args.input_h), (0, 0, 0), 1)
cv2.putText(image[i], '{} {} {} {}'.format(
loader.dataset.idx_to_color[question[i, 0].item()],
loader.dataset.idx_to_question[question[i, 1].item()],
loader.dataset.idx_to_answer[answer[i].item()],
loader.dataset.idx_to_answer[pred[i].item()]),
(2, args.input_h + args.input_h // 6), cv2.FONT_HERSHEY_SIMPLEX, 0.5, (0, 0, 0))
image = torch.from_numpy(image).transpose(2, 3).transpose(1, 2)
writer.add_image('Image', torch.cat([image]), epoch_idx)
print('====> {}: {} Average loss: {:.4f} / Time: {:.4f} / Accuracy: {:.4f}'.format(
mode,
epoch_idx,
epoch_loss / len(loader.dataset),
time.time() - epoch_start_time,
sum(q_correct.values()) / len(loader.dataset)))
writer.add_scalar('{} loss'.format(mode), epoch_loss / len(loader.dataset), epoch_idx)
q_acc = {}
for i in range(args.q_size):
q_acc['question {}'.format(str(i))] = q_correct[i] / q_num[i]
q_corrects = list(q_correct.values())
q_nums = list(q_num.values())
writer.add_scalars('{} accuracy per question'.format(mode), q_acc, epoch_idx)
writer.add_scalar('{} non-rel accuracy'.format(mode), sum(q_corrects[:3]) / sum(q_nums[:3]), epoch_idx)
writer.add_scalar('{} rel accuracy'.format(mode), sum(q_corrects[3:]) / sum(q_nums[3:]), epoch_idx)
writer.add_scalar('{} total accuracy'.format(mode), sum(q_correct.values()) / len(loader.dataset), epoch_idx)
if __name__ == '__main__':
optimizer = optim.Adam([param for model in models.values() for param in list(model.parameters())], lr=args.lr)
writer = SummaryWriter(args.log)
for epoch_idx in range(args.start_epoch, args.start_epoch + args.epochs):
epoch(epoch_idx, True)
epoch(epoch_idx, False)
for model_name, model in models.items():
torch.save(model.state_dict(), args.log + model_name)
print('Model saved in ', args.log)
writer.close()