Model Architecture

[collinarnett]

Build GAN architecture to accept inputs and produce folded outputs based on the architecture diagram in the supplementary material.

[collinarnett]

Model architecture has been implemented as described in the paper however, the GAN does not yield similar results. Hyperparameters that are not mentioned in the paper include:

• batch_size = greatly affects results
• random_seed = randomly affects results

I built more code pushed in 06cf8a3a0861c9dfb573941785676d61817ff085 to make testing these hyperparameters easier. This also makes testing other hyperparameters easier as well.

[collinarnett]

I don't understand why the authors included morphing the output of the generator to be symmetric and clamped. The construction of the dataset could be the answer as to why but including the use of such morphing makes the generator's output look significantly different to the input maps.

This could also be because the authors train the generator and discriminator at different paces but there is no mention of this method in the supplements. Will have to do more experimentation.

[collinarnett]

For now the architecture is finished and more testing needs to be done to tune the model.