Bug with built-in Keras loss in WGAN
When tyring to use Keras' built-in losses for the reconstruction loss in the WGAN model (e.g.. setting recon_loss = "mae"
in config_sha_wgan.json
), the training breaks when checkpointing the model. In particular, the monitored reconstruction loss val_recon_loss
looks as expected, but when checkpointing, a tensor is retrieved instead of a loss-value.
An example logfile is attached here.
Edited by Michael Langguth