Pytorch implementation of popular generative models (early work in progress)

Learn about generative models and apply them to music generation (in raw waveform domain)

See arguments.py for explanation of CLI arguments and default values.

Train a small plain autoencoder with latent vector size 256 on cifar-10 images for 100 epochs: python main.py --latent_size 256 --sigmoid --wd 0.01 --epochs 100 --train --no_upsample --no_pool

Train a small VAE on cifar-10 images, generates samples, computes metrics (FID, inception score, etc), logs data for tensorboard, etc) python main.py --latent_size 256 --sigmoid --wd 0.01 --epochs 100 --train --variational --beta 5e-6 --no_upsample --no_pool --beta_mult 1 --tag test_tag_

This assumes a plain autoencoder (used for FID computation) has already been trained (see above) and the corresponding checkpoint is available in checkpoints folder (saved automatically)