活性化関数

Kerasで学ぶIntroduction to Deep Learning

Miguel Esteban

Data Scientist & Founder

活性化関数の効果

どの活性化関数を使うべきか？

魔法の公式はない
性質はさまざま
問題設定に依存
各層での目的次第
ReLUはまず無難
深層ではSigmoidは非推奨
実験で調整

活性化関数の比較

# Set a random seed
np.random.seed(1)

# Return a new model with the given activation
def get_model(act_function):
    model = Sequential()
    model.add(Dense(4, input_shape=(2,), activation=act_function))
    model.add(Dense(1, activation='sigmoid'))
    return model

活性化関数の比較

# Activation functions to try out
activations = ['relu', 'sigmoid', 'tanh']

# Dictionary to store results
activation_results = {}

for funct in activations:
    model = get_model(act_function=funct)
    history = model.fit(X_train, y_train,
                        validation_data=(X_test, y_test),
                        epochs=100, verbose=0)
    activation_results[funct] = history

活性化関数の比較

import pandas as pd

# Extract val_loss history of each activation function
val_loss_per_funct = {k:v.history['val_loss'] for k,v in activation_results.items()}


# Turn the dictionary into a pandas dataframe
val_loss_curves = pd.DataFrame(val_loss_per_funct)


# Plot the curves
val_loss_curves.plot(title='Loss per Activation function')

¡Vamos a practicar!

Kerasで学ぶIntroduction to Deep Learning