I started to use Variational Autoencoder in a project, (and I have a hard time determining the weight for the reconstruction loss and KL-loss). I have an idea of training the VAE in two steps.:
- Train the network as is with a bigger weight for the KL-loss
- Freeze the encoder and the latent layer and train the decoder using the reconstruction loss only.
Please tell me your opinions on this method and link if you have found any similar approach