2016-03-19 16:07:15 +00:00
|
|
|
'''Trains a simple convnet on the MNIST dataset.
|
2015-12-09 02:49:14 +00:00
|
|
|
|
2016-03-19 16:07:15 +00:00
|
|
|
Gets to 99.25% test accuracy after 12 epochs
|
|
|
|
(there is still a lot of margin for parameter tuning).
|
2015-12-09 02:49:14 +00:00
|
|
|
16 seconds per epoch on a GRID K520 GPU.
|
|
|
|
'''
|
|
|
|
|
2015-04-30 00:17:22 +00:00
|
|
|
from __future__ import print_function
|
2015-06-27 00:41:13 +00:00
|
|
|
import numpy as np
|
2015-07-26 08:00:18 +00:00
|
|
|
np.random.seed(1337) # for reproducibility
|
2015-06-27 00:41:13 +00:00
|
|
|
|
2015-04-30 00:17:22 +00:00
|
|
|
from keras.datasets import mnist
|
|
|
|
from keras.models import Sequential
|
2016-05-12 01:45:37 +00:00
|
|
|
from keras.layers import Dense, Dropout, Activation, Flatten
|
|
|
|
from keras.layers import Convolution2D, MaxPooling2D
|
2015-05-04 17:31:03 +00:00
|
|
|
from keras.utils import np_utils
|
2016-09-06 22:53:56 +00:00
|
|
|
from keras import backend as K
|
2015-04-30 00:17:22 +00:00
|
|
|
|
2015-06-19 19:52:43 +00:00
|
|
|
batch_size = 128
|
2015-04-30 00:17:22 +00:00
|
|
|
nb_classes = 10
|
2015-06-19 22:38:33 +00:00
|
|
|
nb_epoch = 12
|
2015-05-04 17:31:03 +00:00
|
|
|
|
2015-10-05 01:44:49 +00:00
|
|
|
# input image dimensions
|
|
|
|
img_rows, img_cols = 28, 28
|
2015-09-04 23:24:47 +00:00
|
|
|
# number of convolutional filters to use
|
|
|
|
nb_filters = 32
|
2015-10-05 01:44:49 +00:00
|
|
|
# size of pooling area for max pooling
|
2016-09-06 22:53:56 +00:00
|
|
|
pool_size = (2, 2)
|
2015-10-05 01:44:49 +00:00
|
|
|
# convolution kernel size
|
2016-08-18 03:00:49 +00:00
|
|
|
kernel_size = (3, 3)
|
2015-09-04 23:24:47 +00:00
|
|
|
|
2016-04-04 18:30:24 +00:00
|
|
|
# the data, shuffled and split between train and test sets
|
2015-05-04 03:14:51 +00:00
|
|
|
(X_train, y_train), (X_test, y_test) = mnist.load_data()
|
2015-05-04 17:31:03 +00:00
|
|
|
|
2016-09-06 22:53:56 +00:00
|
|
|
if K.image_dim_ordering() == 'th':
|
|
|
|
X_train = X_train.reshape(X_train.shape[0], 1, img_rows, img_cols)
|
|
|
|
X_test = X_test.reshape(X_test.shape[0], 1, img_rows, img_cols)
|
|
|
|
input_shape = (1, img_rows, img_cols)
|
|
|
|
else:
|
|
|
|
X_train = X_train.reshape(X_train.shape[0], img_rows, img_cols, 1)
|
|
|
|
X_test = X_test.reshape(X_test.shape[0], img_rows, img_cols, 1)
|
|
|
|
input_shape = (img_rows, img_cols, 1)
|
|
|
|
|
2015-12-09 02:49:14 +00:00
|
|
|
X_train = X_train.astype('float32')
|
|
|
|
X_test = X_test.astype('float32')
|
2015-05-04 03:14:51 +00:00
|
|
|
X_train /= 255
|
|
|
|
X_test /= 255
|
2015-06-19 22:38:33 +00:00
|
|
|
print('X_train shape:', X_train.shape)
|
2015-04-30 00:17:22 +00:00
|
|
|
print(X_train.shape[0], 'train samples')
|
|
|
|
print(X_test.shape[0], 'test samples')
|
|
|
|
|
|
|
|
# convert class vectors to binary class matrices
|
|
|
|
Y_train = np_utils.to_categorical(y_train, nb_classes)
|
|
|
|
Y_test = np_utils.to_categorical(y_test, nb_classes)
|
|
|
|
|
|
|
|
model = Sequential()
|
2015-06-19 22:38:33 +00:00
|
|
|
|
2016-08-18 03:00:49 +00:00
|
|
|
model.add(Convolution2D(nb_filters, kernel_size[0], kernel_size[1],
|
2015-12-04 06:03:28 +00:00
|
|
|
border_mode='valid',
|
2016-09-06 22:53:56 +00:00
|
|
|
input_shape=input_shape))
|
2015-04-30 00:17:22 +00:00
|
|
|
model.add(Activation('relu'))
|
2016-08-18 03:00:49 +00:00
|
|
|
model.add(Convolution2D(nb_filters, kernel_size[0], kernel_size[1]))
|
2015-04-30 00:17:22 +00:00
|
|
|
model.add(Activation('relu'))
|
2016-09-06 22:53:56 +00:00
|
|
|
model.add(MaxPooling2D(pool_size=pool_size))
|
2015-06-19 22:38:33 +00:00
|
|
|
model.add(Dropout(0.25))
|
|
|
|
|
|
|
|
model.add(Flatten())
|
2015-10-05 01:44:49 +00:00
|
|
|
model.add(Dense(128))
|
2015-06-19 22:38:33 +00:00
|
|
|
model.add(Activation('relu'))
|
|
|
|
model.add(Dropout(0.5))
|
2015-10-05 01:44:49 +00:00
|
|
|
model.add(Dense(nb_classes))
|
2015-04-30 00:17:22 +00:00
|
|
|
model.add(Activation('softmax'))
|
|
|
|
|
2016-03-19 16:07:15 +00:00
|
|
|
model.compile(loss='categorical_crossentropy',
|
|
|
|
optimizer='adadelta',
|
|
|
|
metrics=['accuracy'])
|
2015-04-30 00:17:22 +00:00
|
|
|
|
2015-12-09 02:49:14 +00:00
|
|
|
model.fit(X_train, Y_train, batch_size=batch_size, nb_epoch=nb_epoch,
|
2016-03-19 16:07:15 +00:00
|
|
|
verbose=1, validation_data=(X_test, Y_test))
|
|
|
|
score = model.evaluate(X_test, Y_test, verbose=0)
|
2015-05-04 17:31:03 +00:00
|
|
|
print('Test score:', score[0])
|
|
|
|
print('Test accuracy:', score[1])
|