-
Notifications
You must be signed in to change notification settings - Fork 4
/
one_file_end_to_end.py
255 lines (221 loc) · 10.4 KB
/
one_file_end_to_end.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
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
Created on Thu Jun 27 15:00:42 2019
@author: tuanna118
"""
from keras.preprocessing.image import *
from keras.layers.core import *
import tensorflow as tf
from keras.layers import *
from keras.models import *
import keras
import numpy as np
from keras import backend as K
import pandas as pd
from keras.callbacks import *
np.random.seed(0)
import matplotlib.pyplot as plt
from keras.utils.np_utils import to_categorical
from keras.regularizers import l2
import cv2
letters = " ABCDEFGHIJKLMNPQRSTUVWXYZ0123456789"
dic = {}
for i in range(len(letters)):
dic[i] = letters[i]
invert_dic = {}
for i in range(len(letters)):
invert_dic[letters[i]] = i
db = pd.read_csv("/media/tuanna118/DATA/trainVal.csv")
base_dir = "/media/tuanna118/DATA/data/"
n = len(db)
train_x = []
train_y = []
val_x = []
val_y = []
for i in range(4*n//5):
temp_y= np.zeros((8))
if db["train"][i]==1:
path = base_dir + db["image_path"][i]
temp_x = cv2.imread(path, cv2.IMREAD_GRAYSCALE)
temp_x = cv2.resize(temp_x,(256,64))
#temp_x = temp_x/255
train_x.append(temp_x)
for j,k in enumerate(db["lp"][i]):
temp_y[j] = invert_dic[k]
# temp_y = to_categorical(temp_y,36)
# db["y_col"][i] = temp_y
train_y.append(temp_y)
if db["train"][i]==0:
temp_x = cv2.imread(base_dir+db["image_path"][i], cv2.IMREAD_GRAYSCALE)
temp_x = cv2.resize(temp_x,(256,64))
#temp_x = temp_x/255
val_x.append(temp_x)
for j,k in enumerate(db["lp"][i]):
temp_y[j] = invert_dic[k]
# temp_y = to_categorical(temp_y,36)
# db["y_col"][i] = temp_y
val_y.append(temp_y)
train_x = np.array(train_x).reshape(-1,64,256,1)
train_y = np.array(train_y)
val_x = np.array(val_x).reshape(-1,64,256,1)
val_y = np.array(val_y)
val_y = to_categorical(val_y,36)
train_y = to_categorical(train_y,36)
train_x = train_x /255
val_x = val_x/255
def VGG(shape=(64, 256, 1),n_channels=64,weight_decay=0,batch_momentum=0.99):
bn_axis = 3
input_ = Input(shape=shape)
x = Conv2D(128, (3, 3), padding='same', name='block1_conv1', kernel_initializer='he_normal', kernel_regularizer=l2(weight_decay))(input_)
x = BatchNormalization(axis=bn_axis, name='bn00_x1', momentum=batch_momentum)(x)
x = Activation('relu')(x)
x = Conv2D(128, (3, 3), padding='same', name='block1_conv2', kernel_initializer='he_normal', kernel_regularizer=l2(weight_decay))(x)
x = BatchNormalization(axis=bn_axis, name='bn01_x2', momentum=batch_momentum)(x)
x = Activation('relu')(x)
x = MaxPooling2D((2, 2), strides=(2, 2), name='block1_pool')(x)
# Block 2
x = Conv2D(128, (3, 3), padding='same', name='block2_conv1', kernel_initializer='he_normal', kernel_regularizer=l2(weight_decay))(x)
x = BatchNormalization(axis=bn_axis, name='bn11_x1', momentum=batch_momentum)(x)
x = Activation('relu')(x)
x = Conv2D(128, (3, 3), padding='same', name='block2_conv2', kernel_initializer='he_normal', kernel_regularizer=l2(weight_decay))(x)
x = BatchNormalization(axis=bn_axis, name='bn12_x2', momentum=batch_momentum)(x)
x = Activation('relu')(x)
x = MaxPooling2D((2, 2), strides=(2, 2), name='block2_pool')(x)
# Block 3
x = Conv2D(256, (3, 3), padding='same', name='block3_conv1', kernel_initializer='he_normal', kernel_regularizer=l2(weight_decay))(x)
x = BatchNormalization(axis=bn_axis, name='bn21_x1', momentum=batch_momentum)(x)
x = Activation('relu')(x)
x = Conv2D(256, (3, 3), padding='same', name='block3_conv2', kernel_initializer='he_normal', kernel_regularizer=l2(weight_decay))(x)
x = BatchNormalization(axis=bn_axis, name='bn22_x2', momentum=batch_momentum)(x)
x = Activation('relu')(x)
# x = Conv2D(256, (3, 3), padding='same', name='block3_conv3', kernel_initializer='glorot_uniform', kernel_regularizer=l2(weight_decay))(x)
# x = BatchNormalization(axis=bn_axis, name='bn23_x3', momentum=batch_momentum)(x)
# x = Activation('relu')(x)
x = MaxPooling2D((2, 2), strides=(2, 2), name='block3_pool')(x)
# Block 4
x = Conv2D(512, (3, 3), padding='same', name='block4_conv1', kernel_initializer='he_normal', kernel_regularizer=l2(weight_decay))(x)
x = BatchNormalization(axis=bn_axis, name='bn31_x2', momentum=batch_momentum)(x)
x = Activation('relu')(x)
x = Conv2D(512, (3, 3), padding='same', name='block4_conv2', kernel_regularizer=l2(weight_decay))(x)
x = BatchNormalization(axis=bn_axis, name='bn32_x2', momentum=batch_momentum)(x)
x = Activation('relu')(x)
# x = Conv2D(512, (3, 3), padding='same', name='block4_conv3', kernel_initializer='glorot_uniform', kernel_regularizer=l2(weight_decay))(x)
# x = BatchNormalization(axis=bn_axis, name='bn33_x2', momentum=batch_momentum)(x)
# x = Activation('relu')(x)
x = MaxPooling2D((2, 2), strides=(2, 2), name='block4_pool')(x)
# Block 5
x = Conv2D(512, (3, 3), padding='same', name='block5_conv1', kernel_initializer='he_normal', kernel_regularizer=l2(weight_decay))(x)
x = BatchNormalization(axis=bn_axis, name='bn41_x2', momentum=batch_momentum)(x)
x = Activation('relu')(x)
# x = Conv2D(512, (3, 3), padding='same', name='block5_conv2', kernel_initializer='glorot_uniform', kernel_regularizer=l2(weight_decay))(x)
# x = BatchNormalization(axis=bn_axis, name='bn42_x2', momentum=batch_momentum)(x)
# x = Activation('relu')(x)
x = Conv2D(1024, (3, 3), padding='same', name='block5_conv3', kernel_initializer='he_normal', kernel_regularizer=l2(weight_decay))(x)
x = BatchNormalization(axis=bn_axis, name='bn43_x2', momentum=batch_momentum)(x)
x = Activation('relu')(x)
x = MaxPooling2D((2, 2), strides=(2, 2), name='block5_pool')(x)
x = Conv2D(1024, (3, 3), padding='same', name='block6_conv1', kernel_initializer='he_normal', kernel_regularizer=l2(weight_decay))(x)
x = BatchNormalization(axis=bn_axis, name='bn51_x2', momentum=batch_momentum)(x)
x = Activation('relu')(x)
x = Conv2D(1024*2, (3, 3), padding='same', name='block6_conv12', kernel_initializer='he_normal', kernel_regularizer=l2(weight_decay))(x)
x = BatchNormalization(axis=bn_axis, name='bn51_x22', momentum=batch_momentum)(x)
x = Activation('relu')(x)
x= Dropout(0.3, noise_shape=None, seed=None)(x)
#block5
X = AveragePooling2D((2, 2), strides = (2, 1), name='avg_pool1',padding ='same')(x)
X = Reshape((8,1024*2))(X)
X = Conv1D(512, 3, strides=1, padding='same',name = 'conv1y' ,activation=None, dilation_rate=1, use_bias=True, kernel_initializer="he_normal", kernel_regularizer=regularizers.l2(0.000))(X)
X = BatchNormalization(axis = 2, name = 'bn01y')(X)
X = Activation('relu')(X)
X= Dropout(0.3, noise_shape=None, seed=None)(X)
X = Conv1D(36, 1 , strides=1, padding='same',name = 'conv1x' ,activation=None, dilation_rate=1, use_bias=True, kernel_initializer="he_normal")(X)
X = BatchNormalization(axis = 2, name = 'bnhe')(X)
X = Activation('softmax')(X)
model = Model(inputs = [input_], outputs = [X])
return model
model = VGG(shape=(64, 256, 1))
def custom_loss(y_true, y_pred):
# softmax = K.softmax(y_pred)
s = K.shape(y_pred)
y_true = K.reshape(y_true,(-1,s[-1]))
y_pred = K.reshape(y_pred,(-1,s[-1]))
loss = K.sum(keras.losses.categorical_crossentropy(y_true, y_pred))
#K.mean(keras.losses.categorical_crossentropy(y_true, y_pred))
num = K.shape(y_true)[0]
num=tf.cast(num,tf.float32)
return K.mean(loss)/num
def pixel_acc(y_true, y_pred):
# y_true = y_true[:,:,:,0]
# y_true = K.one_hot(tf.to_int32(K.flatten(y_true)), K.int_shape(y_pred)[-1])
s = K.shape(y_pred)
# reshape such that w and h dim are multiplied together
y_true_reshaped = K.reshape( y_true, (-1,s[-1]) )
y_pred_reshaped = K.reshape( y_pred, (-1, s[-1]) )
# correctly classified
clf_pred = K.argmax(y_pred_reshaped,axis = -1)
y_true = K.argmax(y_true_reshaped,axis = -1)
correct_pixels_per_class = K.cast( K.equal(clf_pred,y_true), dtype='float32')
return K.sum(correct_pixels_per_class) / K.cast(K.prod(s[:-1]), dtype='float32')
def acc1(y_true, y_pred):
# y_true = y_true[:,:,:,0]
# y_true = K.one_hot(tf.to_int32(K.flatten(y_true)), K.int_shape(y_pred)[-1])
s = K.shape(y_pred)
# reshape such that w and h dim are multiplied together
#y_true_reshaped = K.reshape( y_true, (-1,s[-1]) )
#y_pred_reshaped = K.reshape( y_pred, (-1, s[-1]) )
# correctly classified
clf_pred = K.argmax(y_pred,axis = -1)
y_true = K.argmax(y_true,axis = -1)
correct_pixels_per_class = K.cast(K.all( K.equal(clf_pred,y_true),axis=-1), dtype='float32')
return K.sum(correct_pixels_per_class) / K.cast(K.prod(s[0]), dtype='float32')
model.compile(loss = custom_loss,
optimizer='adam',
metrics=[pixel_acc,acc1])
model.summary()
datagen = ImageDataGenerator(width_shift_range=0.14,
height_shift_range=0.08,
fill_mode='constant',
zoom_range = 0.1,
rotation_range = 10,
#rescale =1./255
)
mcp_save = ModelCheckpoint('plate2.h5', save_best_only=True, monitor='val_loss', mode='min',verbose=1)
def scheduler(epoch):
if epoch <3 :
return 0.001/5
elif epoch < 10:
return 0.001/10
elif epoch < 15:
return 0.00001
elif epoch <30:
return 0.00001/2
n = train_x.shape[0]
model.load_weights("plate.h5")
lr_reduce = LearningRateScheduler(scheduler,verbose = 1)
model.fit_generator(datagen.flow(train_x, train_y,batch_size=64),
epochs = 10,
steps_per_epoch=n//64,
callbacks=[lr_reduce,mcp_save],
validation_data=(val_x, val_y))
#model.load_weights("plate.h5")
#import matplotlib.pyplot as plt
#import matplotlib.image as mpimg
#import random
#n = val_x.shape[0]
#count = 0
#pred = model.predict(val_x)
#pred = np.argmax(pred,axis=-1)
#true = np.argmax(val_y,axis = -1)
#for i in range(n):
# if np.all(pred[i,:]==true[i]):
# count +=1
#print(count/n)
#k = random.randint(0,10000)
#imgplot = plt.imshow(val_x[k,:,:,:].reshape(64,256))
#arr = []
#for i in range(8):
# arr.append(dic[pred[k,i]])
#print(arr)
#0.9757237083180652 plate.h5
#0.9758153169659216