-
Notifications
You must be signed in to change notification settings - Fork 50
/
ocr_finetune.py
136 lines (121 loc) · 6.39 KB
/
ocr_finetune.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
# Copyright (c) 2019, RangerUFO
#
# This file is part of alpr_utils.
#
# alpr_utils is free software: you can redistribute it and/or modify
# it under the terms of the GNU General Public License as published by
# the Free Software Foundation, either version 3 of the License, or
# (at your option) any later version.
#
# alpr_utils is distributed in the hope that it will be useful,
# but WITHOUT ANY WARRANTY; without even the implied warranty of
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
# GNU General Public License for more details.
#
# You should have received a copy of the GNU General Public License
# along with alpr_utils. If not, see <https://www.gnu.org/licenses/>.
import os
import time
import random
import argparse
import mxnet as mx
from dataset import load_dataset, ocr_batches
from utils import Vocabulary
from ocr_net import OcrNet
def finetune(max_epochs, learning_rate, batch_size, dims, max_hw, max_len, sgd, context):
print("Loading dataset...", flush=True)
dataset = load_dataset("data/train", "finetune.json")
split = int(len(dataset) * 0.9)
training_set = dataset[:split]
print("Training set: ", len(training_set))
validation_set = dataset[split:]
print("Validation set: ", len(validation_set))
print("Loading model...", flush=True)
vocab = Vocabulary()
vocab.load("model/vocabulary.json")
print("Vocabulary size: ", vocab.size())
model = OcrNet(max_hw, vocab.size(), max_len)
loss = mx.gluon.loss.SoftmaxCrossEntropyLoss(axis=2)
model.load_parameters("model/ocr_net.params", ctx=context)
print("Learning rate: ", learning_rate)
if sgd:
print("Optimizer: SGD")
trainer = mx.gluon.Trainer(model.collect_params(), "SGD", {
"learning_rate": learning_rate,
"momentum": 0.5
})
else:
print("Optimizer: Nadam")
trainer = mx.gluon.Trainer(model.collect_params(), "Nadam", {
"learning_rate": learning_rate
})
if os.path.isfile("model/ocr_net.state"):
trainer.load_states("model/ocr_net.state")
print("Traning...", flush=True)
for epoch in range(max_epochs):
ts = time.time()
random.shuffle(training_set)
training_total_L = 0.0
training_batches = 0
training_num_correct = 0
training_num_inst = 0
for x, tgt, tgt_len, lbl in ocr_batches(training_set, batch_size, dims, max_hw, vocab, max_len, context):
training_batches += 1
with mx.autograd.record():
y, self_attn, context_attn = model(x, tgt, tgt_len)
L = loss(y, lbl, mx.nd.not_equal(lbl, vocab.char2idx("<PAD>")).expand_dims(-1))
L.backward()
trainer.step(x.shape[0])
training_batch_L = mx.nd.mean(L).asscalar()
if training_batch_L != training_batch_L:
raise ValueError()
training_total_L += training_batch_L
pred_lbl = mx.nd.argmax(y, axis=-1)
batch_num_correct = (pred_lbl == lbl).sum().asscalar()
batch_num_inst = len(pred_lbl.reshape((-1))) - (lbl == vocab.char2idx("<PAD>")).sum().asscalar()
print("[Epoch %d Batch %d] batch_loss %.10f average_loss %.10f accuracy %.10f elapsed %.2fs" % (
epoch, training_batches, training_batch_L, training_total_L / training_batches, batch_num_correct / batch_num_inst, time.time() - ts
), flush=True)
training_num_correct += batch_num_correct
training_num_inst += batch_num_inst
training_avg_L = training_total_L / training_batches
validation_total_L = 0.0
validation_batches = 0
validation_num_correct = 0
validation_num_inst = 0
for x, tgt, tgt_len, lbl in ocr_batches(validation_set, batch_size, dims, max_hw, vocab, max_len, context):
validation_batches += 1
y, self_attn, context_attn = model(x, tgt, tgt_len)
L = loss(y, lbl, mx.nd.not_equal(lbl, vocab.char2idx("<PAD>")).expand_dims(-1))
validation_batch_L = mx.nd.mean(L).asscalar()
if validation_batch_L != validation_batch_L:
raise ValueError()
validation_total_L += validation_batch_L
pred_lbl = mx.nd.argmax(y, axis=-1)
batch_num_correct = (pred_lbl == lbl).sum().asscalar()
validation_num_correct += (pred_lbl == lbl).sum().asscalar()
validation_num_inst += len(pred_lbl.reshape((-1))) - (lbl == vocab.char2idx("<PAD>")).sum().asscalar()
validation_avg_L = validation_total_L / validation_batches
print("[Epoch %d] training_loss %.10f validation_loss %.10f training_accuracy %.10f validation_accuracy %.10f duration %.2fs" % (
epoch + 1, training_avg_L, validation_avg_L, training_num_correct / training_num_inst, validation_num_correct / validation_num_inst, time.time() - ts
), flush=True)
model.save_parameters("model/ocr_net.params")
trainer.save_states("model/ocr_net.state")
if __name__ == "__main__":
parser = argparse.ArgumentParser(description="Fine-tune the OCR network.")
parser.add_argument("--max_epochs", help="set the max epochs (default: 10)", type=int, default=10)
parser.add_argument("--learning_rate", help="set the learning rate (default: 1e-6)", type=float, default=1e-6)
parser.add_argument("--batch_size", help="set the batch size (default: 32)", type=int, default=32)
parser.add_argument("--dims", help="set the sample dimentions (default: 208)", type=int, default=208)
parser.add_argument("--img_w", help="set the max width of input images (default: 144)", type=int, default=144)
parser.add_argument("--img_h", help="set the max height of input images (default: 48)", type=int, default=48)
parser.add_argument("--seq_len", help="set the max length of output sequences (default: 8)", type=int, default=8)
parser.add_argument("--sgd", help="using sgd optimizer", action="store_true")
parser.add_argument("--device_id", help="select device that the model using (default: 0)", type=int, default=0)
parser.add_argument("--gpu", help="using gpu acceleration", action="store_true")
args = parser.parse_args()
if args.gpu:
context = mx.gpu(args.device_id)
else:
context = mx.cpu(args.device_id)
finetune(args.max_epochs, args.learning_rate, args.batch_size, args.dims, (args.img_h, args.img_w), args.seq_len, args.sgd, context)