コード

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

多層パーセプトロン [TensorFlowでDeep Learning 3] (mnist_softmax_fc.py)を書き換えると、

v2/mnist_softmax_fc.py

v2/mnist_softmax_fc.py

from helper import *

IMAGE_SIZE = 28 * 28
CATEGORY_NUM = 10
LEARNING_RATE = 0.1
FEATURE_DIM = 100
EPOCHS = 30
BATCH_SIZE = 100
LOG_DIR = 'log_softmax_fc'
EPS = 1e-10


def loss_fn(y_true, y):
    y = tf.clip_by_value(y, EPS, 1.0)
    return -tf.reduce_sum(y_true * tf.math.log(y), axis=1)

class Dense(tf.keras.layers.Layer):
    def __init__(self, units, activation, *args, **kwargs):
        super().__init__(*args, **kwargs)
        self.units = units
        self.activation = tf.keras.activations.get(activation)

    def build(self, input_shape):
        input_dim = int(input_shape[-1])
        self.W = self.add_weight(
            name='weight',
            shape=(input_dim, self.units),
            initializer=tf.keras.initializers.GlorotUniform()
        )
        self.b = self.add_weight(
            name='bias',
            shape=(self.units,),
            initializer=tf.keras.initializers.Zeros()
        )
        self.built = True

    def call(self, x):
        if self.activation is None:
            raise Exception('Activation function is None')
        return self.activation(tf.matmul(x, self.W) + self.b)

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

    model = tf.keras.models.Sequential()
    model.add(Dense(FEATURE_DIM, input_shape=(IMAGE_SIZE,), activation='relu'))
    model.add(Dense(CATEGORY_NUM, input_shape=(FEATURE_DIM,), activation='softmax'))
    model.compile(loss=loss_fn, 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。結局2つの層は活性化関数が異なるだけなのでそれを引数として渡せるように変更。
で、この、Denseクラスは、tf.keras.layers.Dense(tensorflow/python/keras/layers/core.py)とほぼ同じものなので、
わざわざ定義する必要は全くなく、シンプルに書ける。

v2/mnist_softmax_fc_simple.py

v2/mnist_softmax_fc_simple.py

from helper import *

IMAGE_SIZE = 28 * 28
CATEGORY_NUM = 10
LEARNING_RATE = 0.1
FEATURE_DIM = 100
EPOCHS = 30
BATCH_SIZE = 100
LOG_DIR = 'log_softmax_fc'
EPS = 1e-10


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

    model = tf.keras.models.Sequential()
    model.add(tf.keras.layers.Dense(FEATURE_DIM, input_shape=(IMAGE_SIZE,), activation='relu'))
    model.add(tf.keras.layers.Dense(CATEGORY_NUM, input_shape=(FEATURE_DIM,), 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))

めでたしめでたし