Consistent Keras model subclassing
In this branch, the models U-Net and WGAN will be integrated using Keras model subclassing. This is due to the fact that the joint main-script (see comments in #44 (closed)) implementation has some clear limititations yet that will be tackled in this branch. These limitations are among others:
- hacky U-Net implementation without clear inheritance from
keras.Model
- d_steps must be parsed with hparams_dict as model is uninstantiated at this point and thus no default parameters are available
- flag named_targets must be set to False in hparams_dict for WGAN to work with U-Net
Edited by Michael Langguth