Keras: arbitrary batch size input (LSTM - VAE)

Question

I\'m trying to build Variational autoencoder with LSTM as a input layer.
However, there is an issue that batch size is arbitrarily changed in training process.

G