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 | # This script trains a neural network model using MNIST with image augmentation
import gc, keras, os, re, datetime, math
import numpy as np
from keras.models import Sequential
from keras.layers import Dense, Activation, Reshape
from keras.layers.convolutional import Conv2D
from keras.layers.advanced_activations import LeakyReLU
from keras.layers import Flatten, Dropout
from keras.datasets import mnist
from keras.preprocessing.image import ImageDataGenerator, NumpyArrayIterator
TRAIN_EPOCH = 100
BATCH_SIZE = 32
MASK_WIDTH = 8 # Width of the region to be erased
MASK_HEIGHT = 8 # Height of the region to be erased
USE_MASKED_IMAGE_GENERATOR = True # If False, use ImageDataGenerator
# Image size of MNIST
IMAGE_WIDTH = 28
IMAGE_HEIGHT = 28
# Define a discriminator model for numbers
def num_discriminator_model():
model = Sequential()
model.add(Conv2D(64, (5, 5), strides=(2, 2), padding='same',
input_shape=(IMAGE_WIDTH, IMAGE_HEIGHT, 1),
data_format='channels_last'))
model.add(LeakyReLU(0.2))
model.add(Conv2D(128, (5, 5), strides=(2, 2), data_format='channels_last'))
model.add(LeakyReLU(0.2))
model.add(Flatten())
model.add(Dense(256))
model.add(LeakyReLU(0.2))
model.add(Dropout(0.5))
model.add(Dense(10))
model.add(Activation('softmax'))
print(model.summary())
return model
class MaskIterator(NumpyArrayIterator):
def __init__(self, x, y, masked_image_generator,
batch_size=32, shuffle=False, seed=None,
data_format='channels_last',
save_to_dir=None, save_prefix='', save_format='png'):
super(MaskIterator, self).__init__(x, y, masked_image_generator,
batch_size, shuffle, seed,
data_format, save_to_dir,
save_prefix, save_format)
# Erase the part of an image randomly
class MaskedImageGenerator(ImageDataGenerator):
def __init__(self, mask_width=MASK_WIDTH, mask_height=MASK_HEIGHT):
self.mask_width = mask_width
self.mask_height = mask_height
# channels_last
self.channel_axis = 3
self.row_axis = 1
self.col_axis = 2
def flow(self, x, y=None, batch_size=32, shuffle=True, seed=None,
save_to_dir=None, save_prefix='', save_format='png'):
return MaskIterator(
x, y, self,
batch_size=batch_size,
shuffle=shuffle,
seed=seed,
data_format='channels_last',
save_to_dir=save_to_dir,
save_prefix=save_prefix,
save_format=save_format)
def random_transform(self, x, seed=None):
img_row_axis = self.row_axis - 1
img_col_axis = self.col_axis - 1
img_channel_axis = self.channel_axis - 1
if seed is not None:
np.random.seed(seed)
# Select the upper left corner of the box to be erased
row = np.random.randint(x.shape[img_row_axis] - self.mask_height)
col = np.random.randint(x.shape[img_col_axis] - self.mask_width)
for r in range(row, row + self.mask_height):
for c in range(col, col + self.mask_width):
x[r][c] = -1 # Fill a pixel by background color
return x
def standardize(self, x):
return x
def format_x(x):
x = (x.astype(np.float32) - 127.5)/127.5 # [0,255] --> [-1,1]
x = x.reshape((x.shape[0], x.shape[1], x.shape[2], 1))
return x
def train_by(model, opt, datagen, fname):
model.compile(optimizer=opt,
loss='categorical_crossentropy',
metrics=['accuracy'])
mc = keras.callbacks.ModelCheckpoint(filepath="models/ndisc_"+fname+"-{epoch:02d}.hdf5")
hist = model.fit_generator(datagen.flow(x_train, y_train, batch_size=BATCH_SIZE),
steps_per_epoch=len(x_train)/BATCH_SIZE, epochs=TRAIN_EPOCH,
validation_data=(x_val, y_val),
callbacks=[mc]).history
# Evaluate the last model by train and validation data
score_train = model.evaluate(x_train, y_train, batch_size=32)
score_val = model.evaluate(x_val, y_val, batch_size=32)
# Write training history into a file
with open("history.input-mask.log", mode="a") as f:
d = datetime.datetime.today()
f.write("#"+d.strftime("%Y-%m-%d %H:%M:%S")+"\n")
f.write("#Data gen type : "+str(type(datagen))+"\n")
f.write("#Data gen params : "+str(vars(datagen))+"\n")
f.write("#Optimizer :"+str(type(opt))+str(opt.get_config())+"\n")
f.write("#Data in the last line are calculated by the last model"
" and not calculated in model.fit()\n")
f.write("#epoch train-loss train-acc val-loss val-acc\n")
for v in range(0, TRAIN_EPOCH):
f.write("{0} {1:10.6f} {2:10.6f} {3:10.6f} {4:10.6f}\n"
.format(v, hist["loss"][v], hist["acc"][v],
hist["val_loss"][v], hist["val_acc"][v]))
f.write("{0} {1:10.6f} {2:10.6f} {3:10.6f} {4:10.6f}\n"
.format(v+1, score_train[0], score_train[1],
score_val[0], score_val[1]))
f.write("\n\n")
model.save("models/ndisc_"+fname+"-final.hdf5")
if __name__ == '__main__':
if not os.path.exists("models"):
os.mkdir("models")
# Load MNIST data
(x_train, y_train), (x_val, y_val) = mnist.load_data()
x_train = format_x(x_train)
x_val = format_x(x_val)
# Encode labels into 1-hot vectors
y_train = keras.utils.to_categorical(y_train, num_classes=10)
y_val = keras.utils.to_categorical(y_val, num_classes=10)
# Select MaskedImageGenerator or ImageDataGenerator
if USE_MASKED_IMAGE_GENERATOR:
fname = "input-mask-w{0}-h{1}".format(MASK_WIDTH, MASK_HEIGHT)
datagen = MaskedImageGenerator()
else:
fname = "image-gen"
datagen = ImageDataGenerator(
rotation_range=20,
width_shift_range=0.2,
height_shift_range=0.2,
shear_range=math.pi/4, # 45 degree
zoom_range=0.3,
fill_mode="constant",
cval=-1, # constant value for fill_mode
)
# Make and train a model
model = num_discriminator_model()
o = keras.optimizers.Adamax(lr=0.002, beta_1=0.9, beta_2=0.999,
epsilon=1e-08, decay=1e-4)
train_by(model, o, datagen, fname)
gc.collect() # To suppress error messages of TensorFlow
|
