-
Notifications
You must be signed in to change notification settings - Fork 3
/
Copy pathtrain.py
287 lines (281 loc) · 12.6 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
import os
import numpy as np
import random
import torch
import torch.nn as nn
import torch.optim as optim
# import torch.backends.cudnn as cudnn
# torch.backends.cudnn.deterministic = True
# torch.backends.cudnn.benchmark = True
from model.stirer import STIRER
from model.crnn import CRNN
from torch.utils.data import ConcatDataset
from utils.loss import ImageLoss
from utils.label_converter import get_charset, strLabelConverter, str_filt
from utils.ssim_psnr import calculate_psnr, SSIM
from dataset import LRSTRDataset, LRSTR_collect_fn
def eval(model, eval_dataset):
print('----------------<Evaluation>----------------')
model.eval()
global_accs = []
global_crnn_accs = []
# for CE decoder
decode_mapper = {}
for i,c in enumerate(charset):
decode_mapper[i+1] = c
decode_mapper[0] = ''
print(decode_mapper)
# Evaluation
for eval_dataset in eval_datasets:
crnn_accs = []
print("Evaluating", eval_dataset.dataset_name)
eval_data_loader = torch.utils.data.DataLoader(
eval_dataset, batch_size=64,
shuffle=False, num_workers=2,
drop_last=False,collate_fn=collect_fn_eval)
metrics_recorder = {}
for it, batch in enumerate(eval_data_loader):
img_hr, img_lr, label_tensors, length_tensors, label_strs, _ = batch
img_hr = img_hr.cuda()
img_lr = img_lr.cuda()
label_tensors = label_tensors.cuda()
length_tensors = length_tensors.cuda()
with torch.no_grad():
sr, logit, srs, logits = model(img_lr)
logits.append(logit)
srs.append(sr)
for step in range(len(srs)):
if not 'acc_'+str(step) in metrics_recorder.keys():
metrics_recorder['acc_'+str(step)] = []
metrics_recorder['psnr_'+str(step)] = []
metrics_recorder['ssim_'+str(step)] = []
BS = img_hr.size(0)
if args['sr']:
img_sr = srs[step]
if args['rec']:
if step == len(srs)-1:
pred_strs = []
for i in range(logits[step].size(0)):
ids = logits[step][i].tolist()
pred_str = ''
for j in range(len(ids)):
if ids[j] == len(charset)+1:
break
# print('HANDLING',ids[j],decode_mapper[ids[j]])
pred_str += decode_mapper[ids[j]]
# print("PP",pred_str)
pred_strs.append(pred_str)
# print('gt=',label_strs[i],'pred=',pred_str,'data=',logits[step][i])
else:
logit = logits[step].softmax(-1)
pred_strs = label_converter.decode(torch.argmax(logit,2).view(-1),torch.IntTensor([logit.size(1)]*BS))
for i in range(BS):
if args['sr']:
sr = img_sr[i,...]
hr = img_hr[i,...]
psnr = calculate_psnr(sr.unsqueeze(0),hr.unsqueeze(0)).cpu()
ssim = calculate_ssim(sr.unsqueeze(0),hr.unsqueeze(0)).cpu()
metrics_recorder['psnr_'+str(step)].append(psnr)
metrics_recorder['ssim_'+str(step)].append(ssim)
if args['rec']:
pred_str = pred_strs[i]
# pred_str = str_filt(pred_str,'lower')
gt_str = label_strs[i]
# gt_str = str_filt(gt_str,'lower')
if pred_str == gt_str:
metrics_recorder['acc_'+str(step)].append(True)
else:
metrics_recorder['acc_'+str(step)].append(False)
logits_last = crnn(srs[-1]).softmax(-1).transpose(0,1)
pred_strs_crnn = label_converter.decode(torch.argmax(logits_last,2).view(-1),torch.IntTensor([logits_last.size(1)]*BS))
for i in range(BS):
pred_str = pred_strs_crnn[i]
pred_str = str_filt(pred_str,'lower')
gt_str = label_strs[i]
gt_str = str_filt(gt_str,'lower')
if pred_str == gt_str:
crnn_accs.append(True)
else:
crnn_accs.append(False)
global_crnn_accs.append(crnn_accs[-1])
# os._exit(-1)
for k in metrics_recorder.keys():
if len(metrics_recorder[k]) == 0:
metrics_recorder[k].append(-1)
for step in range(len(srs)):
print("STEP %d Acc %.2f PSNR %.2f SSIM %.2f"%(step,
100.0*sum(metrics_recorder['acc_'+str(step)])/len(metrics_recorder['acc_'+str(step)]),
sum(metrics_recorder['psnr_'+str(step)])/len(metrics_recorder['psnr_'+str(step)]),
100.0*sum(metrics_recorder['ssim_'+str(step)])/len(metrics_recorder['ssim_'+str(step)])),flush=True)
# add the last
step = len(srs)-1
global_accs.append(sum(metrics_recorder['acc_'+str(step)])/len(metrics_recorder['acc_'+str(step)]))
print("CRNN Acc %.2f"%(100.0*sum(crnn_accs)/len(crnn_accs)))
final_acc = sum(global_accs)/len(global_accs)
print("Avg CRNN Acc %.2f"%(100.0*sum(global_crnn_accs)/len(global_crnn_accs)))
model.train()
return final_acc
args = {
'exp_name': 'STIRER_S',
'batch_size': 256,#128
'multi_card': True,
'train_dataset': [
'dataset/OCR_Syn_Train/ST/',
'dataset/OCR_Syn_Train/MJ/MJ_train/',
],
'eval_dataset': [
'dataset/textzoom/test/easy/',
'dataset/textzoom/test/medium/',
'dataset/textzoom/test/hard/'
],
'Epoch': 8,
'alpha': 0.5,
'print_iter': 100,
'num_gpu': 2 ,
'eval_iter': 500,
'resume':'',
'seed':3407,
'sr':True,
'rec':True,
'multi_stage': False,
'charset': 37,
'ar': True
}
torch.manual_seed(args['seed']) #cpu
torch.cuda.manual_seed(args['seed']) #gpu
np.random.seed(args['seed']) #numpy
random.seed(args['seed']) # random and transforms
torch.backends.cudnn.deterministic=True #cudnn
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = STIRER(upscale=2, img_size=(16, 64),
window_size=(16,2), img_range=1., depths=[4,5,6],
embed_dim=48, num_heads=[6, 6, 6], mlp_ratio=2, upsampler='pixelshuffledirect',is_eval=False).to(device)
# model = Coding(args).to(device)
crnn = CRNN(32, 1, 37, 256).cuda()
crnn.load_state_dict(torch.load('dataset/mydata/crnn.pth',map_location='cpu'),strict=True)
crnn.eval()
sr_param_list = []
for n,p in model.named_parameters():
if 'upsample' in n or 'conv_after_body' in n:
sr_param_list.append(id(p))
if args['resume'] != '':
print('Loading parameters from',args['resume'])
params = torch.load(args['resume'],map_location='cpu')
model.load_state_dict(params,strict=False)
if args['multi_stage']:
params_sr = torch.load('ckpt/coging_base_sronly.pth',map_location='cpu')
params_rec = torch.load('ckpt/coging_base_S_reconly_v2.pth',map_location='cpu')
params = {}
for k in params_sr.keys():
params[k] = params_sr[k]
for k in params_rec.keys():
params[k] = params_rec[k]
for q in params_rec.keys():
# copy the vit parameters
if 'rec_backbone' in q:
for i in range(args['unshared_encoder_layers']):
params['unshared_block_'+str(i)+'.'+'.'.join(q.split('.')[1:])] = params_rec[q]
print("Loading the following params:", params.keys())
model.load_state_dict(params,strict=False)
# for name, param in model.named_parameters():
# param.requires_grad=False
# print(name,type(param))
# for param in model.parameters():
# print(param.requires_grad)
# os._exit(233)
if args['multi_card']:
model = torch.nn.DataParallel(model, device_ids=range(args['num_gpu']))
# x = torch.randn(4,4,16,64).to(device)
# model(x)
# os._exit(2333)
train_dataset = ConcatDataset([LRSTRDataset(root,syn=True,max_len=20,train=True, args=args) for root in args['train_dataset']])
collect_fn = LRSTR_collect_fn(args=args)
collect_fn_eval = LRSTR_collect_fn(train=False, args=args)
eval_datasets = [LRSTRDataset(root,syn = False, args=args) for root in args['eval_dataset']]
train_loader = torch.utils.data.DataLoader(
train_dataset, batch_size=args['batch_size'],
shuffle=True, num_workers=16,
drop_last=True,collate_fn=collect_fn)
loss_pixel = ImageLoss()
loss_ctc = nn.CTCLoss(blank=0, reduction='mean',zero_infinity=True)
# log_probs = torch.randn(50, 16, 20).log_softmax(2).detach().requires_grad_().cuda()
# targets = torch.randint(1, 20, (16, 30), dtype=torch.long)
# input_lengths = torch.full((16,), 50, dtype=torch.long)
# target_lengths = torch.randint(10,30,(16,), dtype=torch.long)
# print(log_probs.shape, targets.shape, input_lengths.shape, target_lengths.shape)
# loss = loss_ctc(log_probs, targets, input_lengths, target_lengths)
# loss.backward()
# print("???")
# os._exit(233)
# optimizer = optim.AdamW(model.parameters(), lr=5e-5)
base_params = filter(lambda p: id(p) not in sr_param_list,model.parameters())
sr_params = filter(lambda p: id(p) in sr_param_list,model.parameters())
optimizer = optim.AdamW(
[
{"params": base_params},
{"params": sr_params, "lr": 5e-4},
],
lr=1e-4
)
charset = get_charset(args['charset'])
label_converter = strLabelConverter(get_charset(args['charset']))
calculate_ssim = SSIM()
max_acc = 0.0
for epoch in range(args['Epoch']):
for it, batch in enumerate(train_loader):
img_hr, img_lr, label_tensors, length_tensors, label_strs, label_ce = batch
assert label_tensors.size(0) == length_tensors.sum(), 'length not equal!'
img_hr = img_hr.to(device)
img_lr = img_lr.to(device)
label_tensors = label_tensors.to(device)
length_tensors = length_tensors.to(device)
# print(img_hr.shape, img_lr.shape, label_tensors.shape, length_tensors.shape)
sr, ar_logit, srs, logits = model(img_lr, label_ce)
# print(type(ar_logit))
# logits.append(logit)
srs.append(sr)
lengths_input = torch.zeros_like(length_tensors).fill_(logits[0].size(1)).long()
loss_sr, loss_rec = 0.0, (len(logits)+1) * ar_logit.loss.mean()
# print(loss_rec)
# # TODO
# srs = [srs[-1]]
# logits = []
if args['sr']:
for step, sr in enumerate(srs):
loss_sr += (step+1) * loss_pixel(sr, img_hr).mean()
if args['rec']:
for step, logit in enumerate(logits):
# print(logits.shape, label_tensors.shape, lengths_input.shape, length_tensors.shape)
# print(label_tensors[:50])
logit = torch.nn.functional.log_softmax(logit, -1)
# with torch.backends.cudnn.flags(enabled=False):
# print(max(label_tensors),logits.shape,"F",length_tensors)
# with torch.backends.cudnn.flags(enabled=False):
loss_rec += (step+1) * loss_ctc(
log_probs=logit.transpose(0,1),
targets=label_tensors.long(),
input_lengths=lengths_input.long(),
target_lengths=length_tensors.long())
loss = args['alpha'] * loss_sr + (1-args['alpha']) * loss_rec
# print("loss=",loss)
optimizer.zero_grad()
loss.backward()
optimizer.step()
if it % args['print_iter'] == 0:
print("EPOCH[%d] ITER[%d]/[%d] loss_tot[%.5f] loss_sr[%.5f] loss_rec[%.5f]"%(epoch,it,len(train_loader), loss, loss_sr, loss_rec),flush=True)
if it % args['eval_iter'] == 1 and it!=1:
acc = eval(model, eval_datasets)
# os._exit(233)
if acc > max_acc:
max_acc = acc
print("Saving Best")
save_dict = model.module.state_dict()
torch.save(save_dict, os.path.join('ckpt', args['exp_name']+'.pth'))
# os._exit(233)
# if it == 300:
# os._exit(233)
if epoch in [4]:
print("Reduce LR")
lr = optimizer.param_groups[0]['lr']
for param_group in optimizer.param_groups:
param_group['lr'] = lr/2