loss increasing sharply when training goes further

[xiaolinAndy]

Hi, I have been trying to follow your work on SLU data augmentation. However, when I increase the training set size(4000) and the training epochs(50), the KL loss and reconstruction loss increase sharply. I wonder is it a bug of code or the reason of VAE model? Besides, I didn't change the parameters of optimizer or KL annealing weight. The loss curve looks as below