discgen
discgen copied to clipboard
Custom training routines
These series of commits adds two new experiments which act as general purpose tools as discussed in #2, which this pull request is meant to replace. It adds a more generic version of train_classifier and train_vae, which are intended to be compatible with the celeba versions but with many additional options.
- Command line options for things that were hard coded like
- --oldmodel to start training from a previously saved state
- --classifier to use a different classifier filename
- --model to use a different model filename
- --batch-size to use different batch-sizes (generally smaller for memory reasons)
- --z-dim to change network architecture allowing different sized latent space
- --monitor-every and --checkpoint-every to change frequency of those events
- Split out
discriminative_term
in the cost function so it could be monitored separately - Ability to scale relative cost of reconstruction, kl, and discriminative
- Classifier can train on a subset of labels given at runtime
- Can train against compatible fuel datasets other than CelebA
@dribnet I did a first review pass. Could you also flake8
the files and fix issues that might have been introduced by your changes?
Thanks @vdumoulin for the constructive feedback, glad to hear you think overall this would be a welcome addition. I'll be reviewing and updating this PR over the next week.
@vdumoulin - I've addressed most issues in the previous review including a general flake8 code cleanup.