autoencoders

Comparison between regularized autoencoder vs variational autoencoder vs VQGAN in terms of reconstruction loss (not for generation) Uses 102flowers dataset This means the latent space will preferrably be something not much smaller than the input dimensions

Building the docker image

docker build . -t b-lanc/autoencoders

Running the docker container

docker run -dit --name=autoencoders --runtime=nvidia --gpus=0 --shm-size=16gb -v .:/workspace -v /mnt/Data/datasets/102flowers/jpg:/dataset -v /mnt/Data2/DockerVolumes/autoencoders:/saves b-lanc/autoencoders

Going into the docker container

docker exec -it autoencoders /bin/bash

Current progress

Autoencoders when regularized (or without regularized) just does not train properly for some reason... all of the resulted pictures end up being the same picture... maybe a bug on the loss? or maybe L1 instead of L2?

Taking a break for now for... 2 weeks or so Next to do is to implement the dicsriminator and perceptual loss, and try the autoencoder with those