z_mean does not change in CVAE CNN about advanced-deep-learning-with-keras HOT 4 CLOSED

Hi. For the test set, when data is encoded, the z_mean is generated by the encoder.

For CVAE, this is expected to be almost evenly distributed within a circle (2d z_mean). z_mean changes depending on the input test data.

from advanced-deep-learning-with-keras.

encoder = Model([inputs, y_labels],
[z_mean, z_log_var, z],
name='encoder')

The output of the encoder is z_mean, z_log_var, z. I understand the z_mean is generated by the encoder, the problem I'm having is the values for z_mean are all the same for all the input test data. in CVAE. This is not the case when working with regular VAE.

from advanced-deep-learning-with-keras.

CVAE (and VAE) encoder z_mean distribution has an objective function to approximate a Gaussian with zero mean and std=1. On the test set, this can be verified by inserting this code:

z, _, _ = encoder.predict([x_test, to_categorical(y_test)],
                          batch_size=batch_size)
print("shape:", z.shape)
print("mean:", np.mean(z, axis=0))
print("std:", np.std(z, axis=0))
return

Then, it prints (verifying it is not constant but a distribution):

shape: (10000, 2)
mean: [ 0.07939055 0.02497174]
std: [ 0.91762519 0.92036796]

from advanced-deep-learning-with-keras.

When I plot the 10000 values, it does not resemble a normal distribution. All 10000 values are the same. This is for CVAE model only, the VAE model does produce a latent vector z with a normal distribution.

from advanced-deep-learning-with-keras.