A Res-Net Style VAE with an adjustable perceptual loss using a pre-trained vgg19.
Based off of
Added training script with loss logging etc. Dataset uses Pytorch "ImageFolder" dataset, code assumes there is no pre-defined train/test split and creates
one if w fixed random seed so it will be the same every time the code is run.
Basic train command:
python3 train_vae.py -mn test_run --dataset_root #path to dataset root#
Latent space interpolation
Results on validation images of the STL10 dataset at 64x64 with a latent vector size of 512 (images on top are the reconstruction)
NOTE: RES_VAE_64_old.py was used to generate the results below
With Perception loss
Without Perception loss
The images in the STL10 have a lot of variation meaning more "features" need to be encoded in the latent space to achieve a good reconstruction. Using a data-set with less variation (and the same latent vector size) should results in a higher quality reconstructed image.
New Model Test images from VAE trained on CelebA at 128x128 resolution (latent space is therefore 512x2x2) using all layers of the VGG model for the perception loss