An unofficial toy implementation for NVAE 《A Deep Hierarchical Variational Autoencoder》
This implement is tiny, just for learning and validation, the model structure:
As shown in the figure above, the model architecture is very simple, the image is scaled to 64x64 size, in order to reduce the amount of calculation.
Hair and other details are still blurry, but this is mostly because the image used is a low-quality 64x64 size.
Just specify the image directory, see more with python train.py -h.
python train.py --dataset_path <img_directory> --batch_size 128During training, the dataloader will capture the central area of the image and
scale it to 64 x 64, and the checkpoints will be saved into checkpoints directory.
Replacing the default checkpoint path in the file random_sample.py and run example as follow:
python random_sample.pyor generate a 768 x 768 image contains 144 sub-imgs:
python gen_imgs.py