forked from ussamazahid96/ctc-ocr-pytorch
-
Notifications
You must be signed in to change notification settings - Fork 0
/
trainer.py
executable file
·193 lines (152 loc) · 7.66 KB
/
trainer.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
import os
import random
import numpy as np
import torch
import torch.nn as nn
import torch.optim as optim
import warpctc_pytorch as wp
from torch.autograd import Variable
from torch.utils.data import DataLoader
from model import BiLSTM
from decoder import seq_mnist_decoder
from data import seq_mnist_train, seq_mnist_val
class Seq_MNIST_Trainer():
def __init__(self, trainer_params, args):
self.args = args
self.trainer_params = trainer_params
random.seed(trainer_params.random_seed)
torch.manual_seed(trainer_params.random_seed)
if args.cuda:
torch.cuda.manual_seed_all(trainer_params.random_seed)
kwargs = {'num_workers': 1, 'pin_memory': True} if args.cuda else {}
self.train_data = seq_mnist_train(trainer_params)
self.val_data = seq_mnist_val(trainer_params)
self.train_loader = DataLoader(self.train_data, batch_size=trainer_params.batch_size, shuffle=True, **kwargs)
self.val_loader = DataLoader(self.val_data, batch_size=trainer_params.test_batch_size, shuffle=True, **kwargs)
self.starting_epoch = 1
self.prev_loss = 10000
self.model = BiLSTM(trainer_params)
self.criterion = wp.CTCLoss(size_average=True)
self.labels = [i for i in range(trainer_params.num_classes-1)]
self.decoder = seq_mnist_decoder(labels=self.labels)
if args.resume or args.eval or args.export:
print("Loading model from {}".format(args.save_path))
package = torch.load(args.save_path, map_location=lambda storage, loc: storage)
self.model.load_state_dict(package['state_dict'])
if args.cuda:
torch.cuda.set_device(args.gpus)
self.model = self.model.cuda()
self.optimizer = optim.Adam(self.model.parameters(), lr=trainer_params.lr)
if args.resume:
self.optimizer.load_state_dict(package['optim_dict'])
self.starting_epoch = package['starting_epoch']
self.prev_loss = package['prev_loss']
if args.cuda:
for state in self.optimizer.state.values():
for k, v in state.items():
if torch.is_tensor(v):
state[k] = v.cuda()
if args.init_bn_fc_fusion:
if not trainer_params.prefused_bn_fc:
self.model.batch_norm_fc.init_fusion()
self.trainer_params.prefused_bn_fc = True
else:
raise Exception("BN and FC are already fused.")
def serialize(self, model, trainer_params, optimizer, starting_epoch, prev_loss):
package = {'state_dict': model.state_dict(),
'trainer_params': trainer_params,
'optim_dict' : optimizer.state_dict(),
'starting_epoch' : starting_epoch,
'prev_loss': prev_loss
}
return package
def save_model(self, epoch, loss_value):
print("Model saved at: {}\n".format(self.args.save_path))
self.prev_loss = loss_value
torch.save(self.serialize(model=self.model, trainer_params=self.trainer_params,
optimizer=self.optimizer, starting_epoch=epoch + 1, prev_loss=self.prev_loss), self.args.save_path)
def train(self, epoch):
self.model.train()
for i, (item) in enumerate(self.train_loader):
data, labels, output_len, lab_len = item
data = Variable(data.transpose(1,0), requires_grad=False)
labels = Variable(labels.view(-1), requires_grad=False)
output_len = Variable(output_len.view(-1), requires_grad=False)
lab_len = Variable(lab_len.view(-1), requires_grad=False)
if self.args.cuda:
data = data.cuda()
output = self.model(data)
# print("Input = ", data.shape)
# print("model output (x) = ", output)
# print("GTs (y) = ", labels.type())
# print("model output len (xs) = ", output_len.type())
# print("GTs len (ys) = ", lab_len.type())
# exit(0)
loss = self.criterion(output, labels, output_len, lab_len)
loss_value = loss.data[0]
print("Loss value for epoch = {}/{} and batch {}/{} is = {:.4f}".format(epoch,
self.trainer_params.epochs, (i+1)*self.trainer_params.batch_size, len(self.train_data) , loss_value))
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
if self.args.cuda:
torch.cuda.synchronize()
def test(self, epoch=0, save_model_flag=False):
self.model.eval()
loss_value = 0
for i, (item) in enumerate(self.val_loader):
data, labels, output_len, lab_len = item
data = Variable(data.transpose(1,0), requires_grad=False)
labels = Variable(labels.view(-1), requires_grad=False)
output_len = Variable(output_len.view(-1), requires_grad=False)
lab_len = Variable(lab_len.view(-1), requires_grad=False)
if self.args.cuda:
data = data.cuda()
output = self.model(data)
# print("Input = ", data)
# print("model output (x) = ", output.shape)
# print("model output (x) = ", output)
# print("Label = ", labels)
# print("model output len (xs) = ", output_len)
# print("GTs len (ys) = ", lab_len)
index = random.randint(0,self.trainer_params.test_batch_size-1)
label = labels[index*self.trainer_params.word_size:(index+1)*self.trainer_params.word_size].data.numpy()
label = label-1
prediction = self.decoder.decode(output[:,index,:], output_len[index], lab_len[index])
accuracy = self.decoder.hit(prediction, label)
print("Sample Label = {}".format(self.decoder.to_string(label)))
print("Sample Prediction = {}".format(self.decoder.to_string(prediction)))
print("Accuracy on Sample = {:.2f}%\n\n".format(accuracy))
loss = self.criterion(output, labels, output_len, lab_len)
loss_value += loss.data.numpy()
loss_value /= (len(self.val_data)//self.trainer_params.test_batch_size)
print("Average Loss Value for Val Data is = {:.4f}\n".format(float(loss_value)))
if loss_value < self.prev_loss and save_model_flag:
self.save_model(epoch, loss_value)
def eval_model(self):
self.test()
def train_model(self):
for epoch in range(self.starting_epoch, self.trainer_params.epochs + 1):
self.train(epoch)
self.test(epoch=epoch, save_model_flag=True)
if epoch%20==0:
self.optimizer.param_groups[0]['lr'] = self.optimizer.param_groups[0]['lr']*0.98
def export_model(self, simd_factor, pe):
self.model.eval()
self.model.export('r_model_fw_bw.hpp', simd_factor, pe)
def export_image(self, idx=100):
img, label = self.val_data.images[:,idx,:], self.val_data.labels[0][idx]
img = img.transpose(1, 0)
label -= 1
label = self.decoder.to_string(label)
from PIL import Image
from matplotlib import cm
im = Image.fromarray(np.uint8(cm.gist_earth(img)*255))
im.save('test_image.png')
img = img.transpose(1, 0)
img = np.reshape(img, (-1, 1))
np.savetxt("test_image.txt", img, fmt='%.10f')
f = open('test_image_gt.txt','w')
f.write(label)
f.close()
print("Exported image with label = {}".format(label))