LoginSignup
0
0

More than 3 years have passed since last update.

畳み込みニューラルネット Part1 [TensorFlow2.0でDeep Learning 4]

Last updated at Posted at 2019-05-30

(目次はこちら)

はじめに

畳み込みニューラルネット Part1 [TensorFlowでDeep Learning 4]をtensorflow2.0で実現するためにはどうしたらいいのかを書く(tf.keras)。

コード

Python: 3.6.8, Tensorflow: 2.0.0a0で動作確認済み

畳み込みニューラルネット Part1 [TensorFlowでDeep Learning 4]
(mnist_fixed_cnn_simple.py)を書き換えると、

v2/mnist_fixed_cnn_simple.py

v2/mnist_fixed_cnn_simple.py
from helper import *

IMAGE_WIDTH, IMAGE_HEIGHT = 28, 28
CATEGORY_NUM = 10
LEARNING_RATE = 0.1
FILTER_NUM = 2
EPOCHS = 15
BATCH_SIZE = 100
LOG_DIR = 'log_fixed_cnn'


class Prewitt(tf.keras.layers.Layer):
    def build(self, input_shape):
        v = np.array([[ 1, 0, -1]] * 3)
        h = v.swapaxes(0, 1)
        self.kernel = tf.constant(np.dstack([v, h]).reshape((3, 3, 1, 2)), dtype = tf.float32, name='prewitt')
        self.built = True

    def call(self, x):
        x_ = tf.reshape(x, [-1, x.shape[1], x.shape[2], 1])
        return tf.abs(tf.nn.conv2d(x_, self.kernel, strides=[1, 1, 1, 1], padding='SAME'))

if __name__ == '__main__':
    (X_train, y_train), (X_test, y_test) = mnist_samples()

    model = tf.keras.models.Sequential()
    model.add(Prewitt((IMAGE_HEIGHT * IMAGE_WIDTH, FILTER_NUM), input_shape=(IMAGE_HEIGHT, IMAGE_WIDTH)))
    model.add(tf.keras.layers.Flatten())
    model.add(tf.keras.layers.Dense(CATEGORY_NUM, activation='softmax'))
    model.compile(
            loss='categorical_crossentropy',
            optimizer=tf.keras.optimizers.SGD(LEARNING_RATE), metrics=['accuracy'])

    cb = [tf.keras.callbacks.TensorBoard(log_dir=LOG_DIR)]
    model.fit(X_train, y_train, batch_size=BATCH_SIZE, epochs=EPOCHS, callbacks=cb, validation_data=(X_test, y_test))
    print(model.evaluate(X_test, y_test))

と書ける。ただ、Prewittフィルタの表現が適切かどうかが不安が残るが、動作は適切。

めでたしめでたし。(1epochで、すでに精度高い)
image.png

0
0
0

Register as a new user and use Qiita more conveniently

  1. You get articles that match your needs
  2. You can efficiently read back useful information
  3. You can use dark theme
What you can do with signing up
0
0