This is an implementation of the VAE (Variational Autoencoder) for Cifar10
You can read about dataset here -- CIFAR10
All images are taken from the test set. Left row is the original image. Right row is the reconstruction.
Original | Reconstruction |
---|---|
frog | |
bird |
conda env create
python setup.py develop
To initialize training, simply go ahead
./main.py --train
[--dataset {mnist,cifar10,cifar100}]
[--kernel-num KERNEL_NUM] [--z-size Z_SIZE]
[--epochs EPOCHS] [--batch-size BATCH_SIZE]
[--sample-size SAMPLE_SIZE] [--lr LR]
[--weight-decay WEIGHT_DECAY]
[--loss-log-interval LOSS_LOG_INTERVAL]
[--image-log-interval IMAGE_LOG_INTERVAL]
[--resume] [--checkpoint-dir CHECKPOINT_DIR]
[--sample-dir SAMPLE_DIR] [--no-gpus]
dataset == 'mnist', 'cifar10', 'cifar100'
It will convert imgs from срщщыут dataset
./main.py --test
[--dataset {mnist,cifar10,cifar100}]
[--kernel-num KERNEL_NUM] [--z-size Z_SIZE]
[--epochs EPOCHS] [--batch-size BATCH_SIZE]
[--sample-size SAMPLE_SIZE] [--lr LR]
[--weight-decay WEIGHT_DECAY]
[--loss-log-interval LOSS_LOG_INTERVAL]
[--image-log-interval IMAGE_LOG_INTERVAL]
[--resume] [--checkpoint-dir CHECKPOINT_DIR]
[--sample-dir SAMPLE_DIR] [--no-gpus]
PyTorch implementation of Auto-Encoding Variational Bayes, arxiv:1312.6114