Test out if it is possible to make VAE's each learn to recognize a different object.
For example, let us consider the MNIST dataset restricted only to the digits 0 and 1. We would like to have 2 VAE's, and have each one of them learn to represent the digit 0, and the other one learn to represent the digit 1.
sample.py -- once a model has been trained, use this to sample images.
train.py -- run this in order to train the model.
vae.py -- this is the base variational autoencoder model.
supervae.py -- this is the model which contains multiple vae's inside, each of
which is supposed to learn a different concept.
How to run:
All options can either be given on the command line, or they can be provided via a config file.
In order to see a list of the possible options, take a look at cfg_sample.yaml:
# How many blocks to concatenate per width.
# expand_per_width: 2
# How many blocks to concatenate per height.
# expand_per_height: 1
# Size of the latent dimension.
# latent_dim: 128
# KL loss weight.
# beta: 2.0
# Entropy loss weight.
# gamma: 0.05
# How many CNN layers the model should have.
# nlayers: 4
# How many VAEs the module should include.
# nvaes: 2
# How many parameter updates an epoch should contain.
# epoch_length: 500
# How many epochs a single stage should last for.
# stage_length: 20
# How many stages to execute in total.
# nstages: 100
# Use the clevr dataset, and find it at this path.
# clevr: NoneOnce the model has started training, it will save all configuration parameters to cfg_all.yaml.
In order to train on clevr, you need to download the dataset from Stanford, and extract it.
Training example:
python3 train.py --name leonhard --config cfg.yaml --clevr <clevr_root_path> --nvaes 1First, you should checkout the mnist branch:
git checkout mnistThe model can download the mnist and fashion mnist datasets by themselves, so you only need to pass the training options:
python3 train.py --name leonhard --config cfg.yaml --epochs 10 --nvaes 2