Variational autoencoder - comparative analysis of code - keras

Question

Could you comment two version of variational autoencoder loss and show me why they give me different results?
Dataset:
data1 = np.array([0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0,
1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0], dtype='int32')
data2 = np.array([1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0,
1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0], dtype='int32')
data3 = np.array([0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0,
1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0], dtype='int32')
100 samples each, so I have 300 samples.
Code 1:
def vae_loss(x, x_decoded_mean):
xent_loss = objectives.binary_crossentropy(x, x_decoded_mean)
kl_loss = -0.5 * K.mean(1 + z_log_var - K.square(z_mean) - K.exp(z_log_var))
loss = xent_loss + kl_loss
return loss
vae.compile(optimizer='rmsprop', loss=vae_loss)
Code 2:
def zero_loss(y_true, y_pred):
return K.zeros_like(y_pred)
class CustomVariationalLayer(Layer):
def __init__(self, **kwargs):
self.is_placeholder = True
super(CustomVariationalLayer, self).__init__(**kwargs)
def vae_loss(self, x, x_decoded_mean):
xent_loss = original_dim * metrics.binary_crossentropy(x, x_decoded_mean)
K.exp(z_log_var), axis=-1)
return K.mean(xent_loss + kl_loss)
def call(self, inputs):
x = inputs[0]
x_decoded_mean = inputs[1]
loss = self.vae_loss(x, x_decoded_mean)
self.add_loss(loss, inputs=inputs)
return K.ones_like(x)
loss_layer = CustomVariationalLayer()([x, x_decoded_mean])
vae = Model(x, [loss_layer])
vae.compile(optimizer='rmsprop', loss=[zero_loss])
Results are so different and I don't see where? Latent dimension are different. Code 2 shows the separation between groups and code 1 not.
code 1, vae.predict... is not accurate and code 2 give me 1 on all features.
Code 2 gives me accurate feedback of the code:
sent_encoded = encoder.predict(np.array(test), batch_size = batch_size)
sent_decoded = generator.predict(sent_encoded)
and code 1 is not accurate at all.
Both experiments have the same layers. So, once again, where is the different and what is the best solution for dataset like described above?