-
Notifications
You must be signed in to change notification settings - Fork 15
/
train.py
302 lines (237 loc) · 10.6 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
289
290
291
292
293
294
295
296
297
298
299
300
301
302
# TODO: change iteration to epoch
import os
import argparse
import time
import torch
import matplotlib
import torchvision
matplotlib.use('Agg')
from dataloader import *
from utils.utils import *
from PIL import Image, ImageFile
from torch.autograd import Variable
from tensorboardX import SummaryWriter
from model.model import GlyphAdaptor
import _pickle as cp
import numpy as np
import os.path as osp
import torch.nn as nn
import torch.optim as optim
import torch.nn.functional as F
import torch.utils.data as data
import torch.backends.cudnn as cudnn
ImageFile.LOAD_TRUNCATED_IMAGES = True
import cv2
cv2.setNumThreads(0)
def worker_init_fn(worker_id):
np.random.seed(np.random.get_state()[1][0] + worker_id)
def main():
parser = argparse.ArgumentParser(description='template_model')
parser.add_argument('--name', default='init', type=str)
## data setting
parser.add_argument('--root', default='./data',type=str)
parser.add_argument('--trainset', default='attribute4', type=str,help = 'number of training attribute used')
parser.add_argument('--evalset', default='FontSynth', type=str,help = 'FontSynth,100')
parser.add_argument('--load_height', default=32, type=int)
parser.add_argument('--imgW', type=int, default=100, help='the width of the input image')
parser.add_argument('--lang', default="EN", type=str)
parser.add_argument("--data",default = 'data',type=str,help="mixed,syn")
parser.add_argument("--cross", dest="cross", action = 'store_true')
parser.add_argument("--weight",default=1.0, type =float)
# training params
parser.add_argument("--gpus", dest="gpu", default="0", type=str)
parser.add_argument("--char_aug", dest="char_aug", action = 'store_true')
parser.add_argument("--downsample", dest="downsample", action = 'store_true')
parser.add_argument("--init_ims",default=6, type =int)
parser.add_argument('--d_model', default=360, type=int)
parser.add_argument('--num_workers', default=12, type=int)
## model setting
parser.add_argument('--alphabet', default='/ abcdefghijklmnopqrstuvwxyz-', type=str)
parser.add_argument("--model",default='transformer',type=str)
parser.add_argument('--resume_i', default=0, type=int)
parser.add_argument("--TPS", dest="TPS", action = 'store_true')
## optim setting
parser.add_argument('--batch_size', default=24, type=int)
parser.add_argument('--lr', default=0.001, type=float)
parser.add_argument('--beta1', type=float, default=0.5, help='beta1 for adam. default=0.5')
parser.add_argument('--momentum', default=0.9, type=float)
parser.add_argument('--weight_decay', default=1e-5, type=float)
parser.add_argument('--gamma', default=0.1, type=float)
parser.add_argument('--optim', default='adam', type=str, help='sgd, adam, adadelta')
parser.add_argument('--max_norm', default=400, type=int, help='Norm cutoff to prevent explosion of gradients')
parser.add_argument('--max_epoches', default=1000000, type=int)
## output setting
parser.add_argument('--log_iter', default=10, type=int)
parser.add_argument('--eval_iter', default=2000,type=int )
parser.add_argument('--save_iter', default=2000, type=int)
parser.add_argument('--save_folder', default='./ckpts', type=str)
parser.add_argument('--tbx_folder', default='./tbx', type=str)
parser.add_argument('--max_iter', default=200000, type=int)
args = parser.parse_args()
make_dirs(args)
writer = SummaryWriter(args.tbx_dir)
args.nClasses = len(args.alphabet)
os.environ["CUDA_VISIBLE_DEVICES"]=args.gpu
device = torch.device("cuda:0")
torch.set_default_tensor_type('torch.FloatTensor')
## setup glyph converter
converter = strLabelConverter(args.alphabet,)
converter2 =strLabelConverter(args.alphabet[1:])
args.alphabet_gt = args.alphabet
## read list of training and validation fonts
fonts_list,val_fonts = load_fonts(args)
args.baseline = False
## setup model
net = GlyphAdaptor(args)
net = torch.nn.DataParallel(net).to(device)
## load checkpoint
if args.resume_i !=0:
resume_file = osp.join(args.save_folder,str(args.resume_i)+'.pth')
checkpoint = torch.load(resume_file)
net.load_state_dict(checkpoint['model_state_dict'],strict=True)
## setup dataset
if args.data == 'omniglot':
trainset = Omniglot(args)
elif args.data == 'eng':
trainset = LineLoader(args,converter,aug=True)
elif args.data == 'omnieng':
trainset1 = LineLoader(args,converter,aug=True)
trainset2 = Omniglot(args,args.root,size=len(trainset1)//45)
trainset = torch.utils.data.ConcatDataset([trainset1,trainset2])
print('english/synthetic dataset size ratio:',len(trainset1)/len(trainset2))
else :
print('no training dataset specified')
train_loader = data.DataLoader(trainset, args.batch_size,num_workers=args.num_workers,
shuffle=True, collate_fn=text_collate, pin_memory=True, worker_init_fn=worker_init_fn)
## setup optimizer
parameters = net.parameters()
if args.optim == 'sgd':
optimizer = optim.SGD(parameters, lr=args.lr,
momentum=args.momentum, weight_decay=args.weight_decay)
elif args.optim == 'adam':
optimizer = optim.Adam(parameters, lr=args.lr, weight_decay=args.weight_decay)
elif args.optim == 'adadelta':
optimizer = optim.Adadelta(parameters,lr=1.0, rho=0.95, eps=1e-8)
optimizer = optim.Adam(parameters, lr=args.lr, weight_decay=args.weight_decay)
## train model
cudnn.benchmark = True
net.train()
iter_counter = args.resume_i+1
running_loss,running_loss1,running_loss2 = 0.,0.,0.
for i in range(args.max_epoches):
i = args.resume_i +i
t0 = time.time()
for j, batch_samples in enumerate(train_loader):
imgs, labels, char_seg_labels,img_lengths,seg_labels,char_seg_labels2= batch_samples
imgs = torch.transpose(imgs.unsqueeze(1),2,3) #[B,C,H,W]
char_seg_labels = char_seg_labels.to(device)
preds, sims,sims_conv= net(imgs,char_seg_labels,img_lengths,char_seg_labels2)
loss1,loss2 = compute_loss(args,preds,img_lengths,sims,labels,\
char_seg_labels,seg_labels)
loss = loss1+args.weight*loss2
optimizer.zero_grad()
running_loss += loss.item()
running_loss1+= loss1.item()
running_loss2 += loss2.item()
loss.backward()
torch.nn.utils.clip_grad_norm_(net.parameters(), args.max_norm)
optimizer.step()
if iter_counter % (args.log_iter) == 0:
t1 = time.time()
writer.add_scalar('train/train_loss',running_loss/args.log_iter,iter_counter)
writer.add_scalar('train/train_loss_ctc',running_loss2/args.log_iter,iter_counter)
writer.add_scalar('train/train_loss_ce',running_loss1/args.log_iter,iter_counter)
sen_acc,cer,wer,pred_samples,label_samples= lex_free_acc(preds,labels,converter,'ctc')
print('iter:%6d loss:%4.4f cer:%4.1f %4.1fs/batch'\
%(iter_counter,running_loss/args.log_iter, cer*100, (t1-t0)/args.log_iter))
writer.add_scalar('train/train_cer',cer,iter_counter)
visual_txt1 = 'gt: '+ str(label_samples[-1])+ '----- logits_pred: '+str(pred_samples[-1])
print(visual_txt1)
t0 = time.time()
running_loss,running_loss1,running_loss2 = 0.,0.,0.
torch.cuda.empty_cache()
if iter_counter % args.save_iter == 0:
print('Saving state, epoch: %d iter:%d'%(i, j))
torch.save({'model_state_dict': net.state_dict(),'optimizer_state_dict': optimizer.state_dict(),},\
args.save_folder + '/' + repr(iter_counter) + '.pth')
if iter_counter % args.eval_iter == 0:
## eval model
net.eval()
n_correct,cers,wers,counter = 0,0,0,0
for val_font in val_fonts:
args.fontname = val_font
args.ref_font = val_font
testset = TestLoader(args,aug= False)
seq_sampler = data.SequentialSampler(testset)
test_loader = data.DataLoader(testset, 1, num_workers=args.num_workers,
sampler = seq_sampler,pin_memory=True,
collate_fn=text_collate_eval,drop_last=False,
worker_init_fn=worker_init_fn)
cer_font,acc_font,wer_font = [],[],[]
for index, sample in enumerate(test_loader):
imgs, char_seg_label, labels, length, alphabet= sample
imgs = torch.transpose(imgs.unsqueeze(1),2,3)
preds,sims,final_conv= net(imgs,char_seg_label,length.long(),char_seg_label) #[batch,len,classes]
correct,cer,wer,pred_str,gt_str = lex_free_acc(preds,labels,converter,'ctc',converter_gt = converter)
counter += 1
cer_font.append(cer)
acc_font.append(correct)
wer_font.append(wer)
n_correct += correct
cers += cer
wers += wer
print('font:%s cer = %.1f wer = %.1f' %(args.fontname,np.mean(cer_font)*100,np.mean(wer_font)*100))
acc = n_correct*1.0/counter
char_er = cers*1.0 / counter
word_er = wers*1.0 / counter
print('accuracy=%.1f cer = %.1f wer = %.1f total_samples= %d'\
%(acc,char_er*100,word_er*100,counter))
print('----------')
if args.eval_iter != 1:
writer.add_scalar('val/val_sentence_accuracy',acc,iter_counter)
writer.add_scalar('val/val_cer',char_er,iter_counter)
writer.add_scalar('val/val_wer',word_er,iter_counter)
converter = strLabelConverter(args.alphabet)
iter_counter += 1
net.train()
def load_fonts(args):
fonts_list = open(osp.join(args.root,'gt','train_att4.txt'),'r').readlines()
val_fonts = []
if args.evalset == 'FontSynth':
lines = open(osp.join(args.root,'gt','test_FontSynth.txt'),'r').readlines()
elif args.evalset == '100':
lines = open(osp.join(args.root,'gt','val_100.txt'),'r').readlines()
for line in lines:
val_fonts.append(line.split('/')[-1].replace('.ttf','').replace('\n',''))
return fonts_list,val_fonts
def make_dirs(args):
if osp.exists(args.save_folder) == False:
os.mkdir(args.save_folder)
args.save_folder = osp.join(args.save_folder ,args.name)
if osp.exists(args.save_folder) == False:
os.mkdir(args.save_folder)
args.tbx_dir =osp.join(args.tbx_folder,args.name)
if osp.exists(args.tbx_folder) == False:
os.mkdir(args.tbx_folder)
if osp.exists(args.tbx_dir) == False:
os.mkdir(args.tbx_dir)
result_dir = osp.join(args.tbx_dir,'results')
if osp.exists(result_dir) == False:
os.mkdir(result_dir)
def compute_loss(args,preds,img_lengths,sims,labels,char_seg_labels,seg_labels):
## setup criterion
ce_loss = nn.CrossEntropyLoss()
ctc_loss = nn.CTCLoss(zero_infinity =True)
# loss on similarity map
max_len = max(img_lengths[0::2]).item()
logits = compute_logits(sims,char_seg_labels, args.nClasses)
seg_labels = seg_labels[:,:max_len].reshape(-1).long().cuda() + 1
loss1 = ce_loss(logits, seg_labels)
# ctc loss on sequence prediction
labels = labels.view(preds.shape[0],-1)[:,:preds.shape[1]]
preds_ctc,labels_ctc,input_lens,label_lens = \
convert_ctc_format(preds,labels,img_lengths[::2],args.alphabet)
loss2 = ctc_loss(preds_ctc,labels_ctc,input_lens,label_lens)
return loss1, loss2
if __name__ == '__main__':
main()