Do Keras dropout and GaussianNoise layers apply different instances of noise in every batch?

Question

I am building a denoising autoencoder in Keras. The model I'm using is

input_img = Input(shape=(10,))
encoded = GaussianNoise(0.01)(input_img)
encoded = Dropout(0.1)(encoded)
encoded = Dense(20,activation='relu')(encoded)
decoded = Dense(10, activation='sigmoid')(encoded)
ae = Model(input=input_img, output=decoded)

If I subsequently call

ae.fit(x_train, x_train,
                     nb_epoch=3,
                     batch_size=5,
                     shuffle=True,
                     validation_data=(x_test, x_test))

is there a new instance of the noise created for each batch? In other words, for each epoch above are there different instances of the noise for each of the batches? Or is the noise instance fixed to the same thing for all batches and only changes when the epoch changes? Or worse is there only one noise instance selected for the entire thing?

here the differences between GausDropout Dropout and GausNoise https://stackoverflow.com/a/65504785/10375049 — Marco Cerliani, Dec 30 '20 at 11:48

score 4 · Answer 1 · edited Jul 09 '21 at 15:50

4

A different instance of the noise is created for each batch during training.

edited Jul 09 '21 at 15:50

JakeBoggs

274
4
17

answered May 27 '17 at 15:02

Walter Hugo Lopez Pinaya

41
4

3

Welcome to SO. Please read this [how-to-answer](http://stackoverflow.com/help/how-to-answer) and follow the guideline there to provide quality answer. – thewaywewere May 27 '17 at 15:26

Do Keras dropout and GaussianNoise layers apply different instances of noise in every batch?

1 Answers1