0

I'm trying to define the neural network model as a function and use it after inside a for-loop. But I'm having different results when defining the model inside the for loop as the following:

The first Case:

def model_1():
    model_1 = keras.Sequential()
    model_1.add(Dense(4, input_dim=5, activation='tanh'))
    model_1.add(Dropout(0.3))
    model_1.add(Dense(2, activation='sigmoid'))
    model_1.compile(loss= 'binary_crossentropy', optimizer='adam', metrics=['accuracy'])
    return model_1

and next

for train_index_1, test_index_1 in kfold_1.split(x1, y1):
    [X_train_1, X_test_1] = x1[train_index_1], x1[test_index_1]
    [y_train_1, y_test_1] = y1[train_index_1], y1[test_index_1]
    history_1 = model_1().fit(X_train_1, y_train_1, validation_data=(X_test_1, y_test_1), batch_size=64, verbose=0, epochs=500)

The second Case:

for train_index_1, test_index_1 in kfold_1.split(x1, y1):
    [X_train_1, X_test_1] = x1[train_index_1], x1[test_index_1]
    [y_train_1, y_test_1] = y1[train_index_1], y1[test_index_1]
    model_1 = keras.Sequential()
    model_1.add(Dense(4, input_dim=5, activation='tanh'))
    model_1.add(Dropout(0.3))
    model_1.add(Dense(2, activation='sigmoid'))
    model_1.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy'])
    history_1 = model_1.fit(X_train_1, y_train_1, validation_data=(X_test_1, y_test_1), batch_size=64, verbose=0, epochs=500)

What is the problem with those? What is the difference?

Ynjxsjmh
  • 28,441
  • 6
  • 34
  • 52
Dua'a Hamdan
  • 1
  • 1
  • What do you mean by "having different results"? Consider there is a lot of randomness involved in the training process, which leaves to different results even for the same model run twice. You need to use seeds to make it deterministic, and even in those cases it can be hard to achieve. – Rodrigo Laguna Aug 16 '22 at 11:25
  • That said, both alternatives look the same implementation. This section from keras FAQ may help: https://keras.io/getting_started/faq/#how-can-i-obtain-reproducible-results-using-keras-during-development – Rodrigo Laguna Aug 16 '22 at 11:30

0 Answers0