forked from zhenming33/RAN_torch
-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
136 lines (102 loc) · 4.61 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
import os
os.environ["CUDA_VISIBLE_DEVICES"] = "3"
import torch
from torch import nn
import torch.optim as optim
from torch.nn.utils.rnn import pack_padded_sequence
from model import Model
from tensorboardX import SummaryWriter
import numpy as np
from tqdm import tqdm
from data_loader import dataset
import Levenshtein
from utils import *
from torch.utils.data import DataLoader
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
train_steps = 50
train_steps_size = 1000
batch_size = 256
lr = 1e-3
image_channel = 3
alpha_c = 1.
grad_clip = 5.
valid_interval = 50
valid_steps = 1
save_interval = 1000
label_path = 'labels.txt'
vocab_path = 'radical_alphabet.txt'
dict_path = 'char2seq_dict.pkl'
word_map = open(vocab_path, encoding='utf-8').readlines()[0]
word_map = word_map + 'sep'
vocab_size = len(word_map)
save_dir = 'weights'
log_dir = 'logs/lr1e-3+batch256+edropout0.5+xvaier+data_shuff+grad_clip+lrdeacy'
if not os.path.exists(save_dir):
os.mkdir(save_dir)
# train step
def train_step(model, criterion, optimizer,images, encoded_captions, caption_lengths):
model.train()
optimizer.zero_grad()
scores, caps_sorted, decode_lengths, alphas, sort_ind = model(images, encoded_captions, caption_lengths)
targets = caps_sorted
scores_pad, _ = pack_padded_sequence(scores, decode_lengths, batch_first=True)
targets_pad, _ = pack_padded_sequence(targets, decode_lengths, batch_first=True)
loss = criterion(scores_pad, targets_pad)
loss += alpha_c * ((1. - alphas.sum(dim=1)) ** 2).mean()
loss.backward()
clip_gradient(model, grad_clip)
# clip_gradient(optimizer, grad_clip)
optimizer.step()
return loss
def valid_step(model, criterion, images, encoded_captions, caption_lengths):
with torch.no_grad():
model.eval()
scores, caps_sorted, decode_lengths, alphas, sort_ind = model(images, encoded_captions, caption_lengths)
targets = caps_sorted
scores_pad, _ = pack_padded_sequence(scores, decode_lengths, batch_first=True)
targets_pad, _ = pack_padded_sequence(targets, decode_lengths, batch_first=True)
loss = criterion(scores_pad, targets_pad)
loss += alpha_c * ((1. - alphas.sum(dim=1)) ** 2).mean()
pred_index = torch.argmax(scores[0], 1).cpu().numpy()
preds = [word_map[c] for c in pred_index]
label_index = caps_sorted[0, :max(decode_lengths)].cpu().numpy()
labels = [word_map[c] for c in label_index]
preds = ''.join(preds)
labels = ''.join(labels)
return loss, alphas, preds, labels, sort_ind
if __name__ == "__main__":
# add records
writer = SummaryWriter(log_dir)
dataloader = data_gen(batch_size, dataset, label_path, vocab_path, dict_path, train_percent=0.7, num_workers=1)
model = Model(image_channel, vocab_size).to(device)
optimizer = optim.Adam(model.parameters(), lr=lr)
criterion = nn.CrossEntropyLoss().to(device)
global_step = 0
# record_graph(writer, model)
for i in range(train_steps):
print('train steps ' + str(i))
for k in tqdm(range(train_steps_size)):
if global_step == 3000:
adjust_learning_rate(optimizer, 0.1)
lr = lr * 0.1
record_scale(writer, lr, global_step, 'lr')
images, encoded_captions, caption_lengths = dataloader.train()
loss = train_step(model, criterion, optimizer, images, encoded_captions, caption_lengths)
record_scale(writer, loss, global_step, 'train/loss')
if global_step % valid_interval == 0:
images, encoded_captions, caption_lengths = dataloader.valid()
loss, alphas, preds, labels, sort_ind = valid_step(model, criterion, images, encoded_captions, caption_lengths)
images = images[sort_ind]
record_scale(writer, loss, global_step, 'valid/loss')
# record_images(writer, images, global_step)
# for t in range(max(caption_lengths).item()):
# record_attention(writer, alphas, t, global_step)
# record_text(writer, preds, global_step, 'valid/preds')
# record_text(writer, labels, global_step, 'valid/labels')
edit_distance = Levenshtein.distance(preds,labels)
normalized_edit_distance = edit_distance / max(len(preds), len(labels))
record_scale(writer, normalized_edit_distance, global_step, 'valid/N.E.D')
if global_step % save_interval == 0:
torch.save(model, save_dir + '/' + str(global_step) + '.pth')
global_step = global_step + 1
writer.close()