train.py
6.18 KB
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
# -*- coding: utf-8 -*-
# @Author : lk
# @Email : 9428.al@gmail.com
# @Create Date : 2022-04-19 21:15:37
# @Last Modified : 2022-04-23 15:06:00
# @Description :
import os
# os.environ['CUDA_VISIBLE_DEVICES'] = '-1'
import cv2
import time
import json
import math
import pyclipper
import numpy as np
import tensorflow as tf
import matplotlib.pyplot as plt
from shapely.geometry import Polygon
from shapely.geometry.polygon import LinearRing
from utils import MobileUNet, resize_with_padding, DBNetLoss
def gen_dbmap(image_path, label_path):
prob_map = np.zeros((1024, 1024), dtype=np.float32)
thre_map = np.zeros((1024, 1024), dtype=np.float32)
binary_map = np.zeros((1024, 1024), dtype=np.float32)
image_path = image_path.decode('utf-8')
image = cv2.imread(image_path)
if np.random.uniform(0, 1) > 0.5:
image = image[...,::-1]
image, ratio = resize_with_padding(image, 1024) # 1024 * 1024 dtype=float32
label_path = label_path.decode('utf-8')
label_dict = json.load(open(label_path, 'r', encoding="utf-8"))
for shape in label_dict['shapes']:
points = np.array(shape['points']) * ratio
points = points[::-1, :] if LinearRing(points).is_ccw else points # 使得都是顺时针
points = np.array(points, dtype=np.int32)
polygon = Polygon(points)
if polygon.is_valid:
distance = int(polygon.area * (1 - np.power(0.4, 2)) / polygon.length)
cv2.fillPoly(prob_map, [points], 1)
cv2.polylines(thre_map, [points], isClosed=True, color=1, thickness=distance)
binary_map = prob_map
prob_map = np.clip(prob_map - thre_map, 0, 1)
# plt.figure(figsize=(10, 20))
# plt.subplot(2, 2, 1)
# plt.imshow(image.astype(np.uint8))
# plt.subplot(2, 2, 2)
# plt.imshow(prob_map)
# plt.subplot(2, 2, 3)
# plt.imshow(thre_map)
# plt.subplot(2, 2, 4)
# plt.imshow(binary_map)
# plt.show()
label = np.stack([prob_map, thre_map, binary_map], axis=-1)
return image, label
def data_augment(image, label):
if tf.random.uniform(()) < 0.25:
image = tf.image.flip_left_right(image)
label = tf.image.flip_left_right(label)
if tf.random.uniform(()) < 0.25:
image = tf.image.flip_up_down(image)
label = tf.image.flip_up_down(label)
return image, label
if __name__ == '__main__':
dataDir = './dataset/wild_200/'
train_images = [os.path.join(dataDir, 'train/image', fn) for fn in os.listdir(dataDir+'train/image')]
train_labels = [os.path.join(dataDir, 'train/json', fn.replace('.jpg', '.json')) for fn in os.listdir(dataDir+'train/image')]
for fn in os.listdir('/home/lk/MyProject/文字检测相关/DBNet2022/dataset/inter_ocr/image'):
train_images.append(os.path.join('/home/lk/MyProject/文字检测相关/DBNet2022/dataset/inter_ocr/image', fn))
train_labels.append(os.path.join('/home/lk/MyProject/文字检测相关/DBNet2022/dataset/inter_ocr/json', fn.replace('.jpg', '.json')))
valid_images = [os.path.join(dataDir, 'test/image', fn) for fn in os.listdir(dataDir+'test/image')]
valid_labels = [os.path.join(dataDir, 'test/json', fn.replace('.jpg', '.json')) for fn in os.listdir(dataDir+'test/image')]
# for i, j in zip(train_images, train_labels):
# image, label = gen_dbmap(image_path=i.encode('utf-8'),
# label_path=j.encode('utf-8'))
# print(label.shape)
train_ds = tf.data.Dataset.from_tensor_slices((train_images, train_labels)).shuffle(10086)
train_ds = train_ds.map(lambda item1, item2: tf.numpy_function(gen_dbmap, [item1, item2], [tf.float32, tf.float32]), num_parallel_calls=tf.data.experimental.AUTOTUNE)
train_ds = train_ds.map(data_augment, num_parallel_calls=tf.data.experimental.AUTOTUNE)
train_ds = train_ds.batch(2).prefetch(buffer_size=32)
valid_ds = tf.data.Dataset.from_tensor_slices((valid_images, valid_labels)).shuffle(10086)
valid_ds = valid_ds.map(lambda item1, item2: tf.numpy_function(gen_dbmap, [item1, item2], [tf.float32, tf.float32]), num_parallel_calls=tf.data.experimental.AUTOTUNE)
valid_ds = valid_ds.batch(2).prefetch(buffer_size=32)
strategy = tf.distribute.MirroredStrategy()
with strategy.scope():
model = MobileUNet()
model.summary()
model.load_weights('./model/ckpt.h5', by_name=True, skip_mismatch=True)
loss_fn = DBNetLoss()
optimizer = tf.keras.optimizers.Adam(learning_rate=3e-4)
# for batch_data in train_ds:
# batch_images, batch_label = batch_data
# pred = model(batch_images, training=True)
# prob_map_pred, thre_map_pred = tf.split(pred, num_or_size_splits=2, axis=3)
# binary_map_pred = 1 / (1 + tf.math.exp(-50 * (prob_map_pred - thre_map_pred)))
# plt.figure(figsize=(10, 20))
# plt.subplot(1, 4, 1)
# plt.imshow(batch_images[0]/255.)
# plt.subplot(1, 4, 2)
# plt.imshow(prob_map_pred[0])
# plt.subplot(1, 4, 3)
# plt.imshow(thre_map_pred[0])
# plt.subplot(1, 4, 4)
# plt.imshow(binary_map_pred[0])
# plt.show()
# loss = loss_fn(batch_label, pred)
# print(loss)
model.compile(loss=loss_fn, optimizer=optimizer)
cp_callback = tf.keras.callbacks.ModelCheckpoint('./model/ckpt.h5', monitor='val_loss', save_best_only=True)
history = model.fit(train_ds,
validation_data=valid_ds,
epochs=200,
callbacks=[cp_callback],
)
model = tf.keras.models.load_model('./model/ckpt.h5', compile=False)
model.compile(loss=loss_fn, optimizer=optimizer)
model.evaluate(valid_ds)
fig, ax = plt.subplots(1, 3, figsize=(30, 3))
ax = ax.ravel()
for i, metric in enumerate(["loss"]):
ax[i].plot(history.history[metric])
ax[i].plot(history.history["val_" + metric])
ax[i].set_title("Model {}".format(metric))
ax[i].set_xlabel("epochs")
ax[i].set_ylabel(metric)
ax[i].legend(["train", "val"])
plt.savefig('./loss.png')
# plt.show()