/
train_cocr.py
140 lines (124 loc) · 5.89 KB
/
train_cocr.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
import os
import pickle
import json
import sys
import editdistance
import json
import torchvision.transforms as transforms
import torch
from torch.utils.data import DataLoader
from tqdm import tqdm
from orfac_dataset import ORFACDataset
from orfac_dataset import orfac_collate_batch
from network import OCROnly
from network import COCR
from converter import Converter
device = 'cuda:0'
batch = 32
patience = 30
model_path = os.path.join('models', 'cocr')
cs = json.load(open('charmap.json', 'rt'))
cs['/PAD/'] = len(cs)+1
converter = Converter(cs)
augs =[transforms.Grayscale()]
augs.append(transforms.ColorJitter(brightness=0.2, contrast=0.2))
augs.append(transforms.RandomAffine(degrees=0, shear=5))
augs.append(transforms.ToTensor())
trans = transforms.Compose(augs)
ds = ORFACDataset(folder=('data-aug/train/multiple', 'data-aug/train/single/antiqua', 'data-aug/train/single/bastarda', 'data-aug/train/single/fraktur', 'data-aug/train/single/gotico-antiqua', 'data-aug/train/single/greek', 'data-aug/train/single/hebrew', 'data-aug/train/single/italic', 'data-aug/train/single/rotunda', 'data-aug/train/single/schwabacher', 'data-aug/train/single/textura'), charset=cs, transform=trans)
train_dataloader = DataLoader(ds, batch_size=batch, shuffle=True, collate_fn=orfac_collate_batch, num_workers=7)
trans = transforms.Compose([
transforms.Grayscale(),
transforms.ToTensor()
])
ds2 = ORFACDataset(folder=('data-aug/valid/multiple', 'data-aug/valid/single/antiqua', 'data-aug/valid/single/bastarda', 'data-aug/valid/single/fraktur', 'data-aug/valid/single/gotico-antiqua', 'data-aug/valid/single/italic', 'data-aug/valid/single/rotunda', 'data-aug/valid/single/schwabacher', 'data-aug/valid/single/textura'), charset=None, transform=trans)
valid_dataloader = DataLoader(ds2, batch_size=1, shuffle=False, collate_fn=orfac_collate_batch, num_workers=7)
network = COCR(
OCROnly(nb_classes=13, feature_dim=32, lstm_layers=1).to(device),
{n: OCROnly(nb_classes=(len(cs)+1), feature_dim=128).to(device) for n in range(13)}
).to(device)
try:
network.load(model_path)
print('Previous weights reloaded')
except:
print('Could not load previous model, loading default ones')
network.classifier.load(os.path.join('models', 'sequence_classifier'))
for n in tqdm(sorted(network.models), desc='Loading OCR models'):
network.models[n].load(os.path.join('models', 'all_heavy_aug'))
print('Default weights loaded')
ctc_loss = torch.nn.CTCLoss(zero_infinity=True)
# ~ optimizer = torch.optim.SGD(network.parameters(), lr=0.0001, momentum=0.9)
# ~ params = [p for p in network.classifier.parameters()]
# ~ for n in network.models:
# ~ params += [p for p in network.models[n].parameters()]
# ~ optimizer = torch.optim.Adam(params, lr=0.0001)
optimizers = [torch.optim.Adam(network.models[n].parameters(), lr=0.0001) for n in sorted(network.models)]
for o in optimizers:
o.load_state_dict(torch.load(os.path.join('models', 'all_heavy_aug', 'optimizer.pth')))
optimizers.append(torch.optim.Adam(network.classifier.parameters(), lr=0.0001))
# ~ optimizers[-1].load_state_dict(torch.load(os.path.join('models', 'all_heavy_aug', 'optimizer.pth')))
schedulers = [torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, 'min', factor=0.5, patience=5) for optimizer in optimizers]
cer=100
os.makedirs(model_path, exist_ok=True)
with open(os.path.join(model_path, 'logs.txt'), 'wt') as lfile:
# ~ with torch.no_grad():
# ~ network.eval()
# ~ d_sum = 0
# ~ c_sum = 0
# ~ for tns, base_width, lbl, _, cf, pf in tqdm(valid_dataloader, desc='Validation', leave=False):
# ~ out = network(tns.to(device)).transpose(0,1)
# ~ am = torch.argmax(out[:, :, :], 2)
# ~ res = converter.decode(am, base_width)
# ~ for i in range(len(lbl)):
# ~ d_sum += editdistance.eval(res[i], lbl[i])
# ~ c_sum += len(lbl[i])
# ~ best_cer = (100*d_sum/c_sum)
# ~ print('Initial CER:', best_cer)
best_cer = 100
no_imp = 0
for epoch in range(1000):
loss_sum = 0
network.train()
batches = 0
for tns, base_width, lbl, base_length, cf, pf in tqdm(train_dataloader, desc='COCR'):
tns = tns.to(device)
lbl = lbl.to(device)
out = network(tns)
il = network.convert_widths(base_width, out.shape[0])
ol = torch.Tensor([l for l in base_length]).long()
loss = ctc_loss(out.log_softmax(2), lbl, input_lengths=il, target_lengths=ol)
for optimizer in optimizers:
optimizer.zero_grad()
loss.backward()
for optimizer in optimizers:
optimizer.step()
loss_sum += loss.item()
batches += 1
with torch.no_grad():
network.eval()
d_sum = 0
c_sum = 0
for tns, base_width, lbl, _, cf, pf in tqdm(valid_dataloader, desc='Validation', leave=False):
# ~ for tns, base_width, lbl, _, cf, pf in valid_dataloader:
out = network(tns.to(device)).transpose(0,1)
am = torch.argmax(out[:, :, :], 2)
res = converter.decode(am, base_width)
for i in range(len(lbl)):
d_sum += editdistance.eval(res[i], lbl[i])
c_sum += len(lbl[i])
cer = (100*d_sum/c_sum)
for scheduler in schedulers:
scheduler.step(cer)
tqdm.write('Loss sum: %.6f' % (loss_sum/batches))
tqdm.write(' CER: %.2f' % cer)
lfile.write('%d;%f;%f\n' % (epoch, loss_sum, cer))
lfile.flush()
if cer<=best_cer:
no_imp = 0
network.save(model_path)
best_cer = cer
else:
no_imp += 1
if no_imp>patience:
print('No improvement, lowest CER: %.2f' % best_cer)
break