Accompanying code to Sparse-Coding Variational Auto-Encoders (SVAE). This code base provides code to train and evaluate both the SVAE (presented in the paper above) and the Olshausen & Field (1996, 1997) original sparse coding model (Sparsenet here) on naturalistic images from the Berkeley Segmentation Dataset (BDSD300).
First, create and activate the svae conda environment using the environment.yml file, and install the local svae package with setup.py. Finally, make sure to change the paths.user_home_dir in the conf/config.yaml file to your working directory.
The SVAE model can be trained using the bin/train_svae.py file, with arguments handled through Hydra. For example:
python3 bin/train_svae.py \
models.svae.prior='CAUCHY' \
train.svae.learning_rate=1e-05 \
train.svae.num_epochs=32
and similarly for the Sparsenet model, using bin/train_sparsenet.py. Consult the conf/config.yaml file to see the possible training and modelling parameters. All models are saved in ./outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}.
We use Annealed Importance Sampling (AIS) to estimate the log-likelihood on held out data of all models under consideration. The appropriate step-size is determined in the script if eval.evaluate_ll.search_epsilon is True (default). To test a SVAE model saved in relative/path/to/pretrained_model/ run
python3 evaluate/evaluate_ll.py \
eval.evaluate_ll.mdl_path="relative/path/to/pretrained_model" \
eval.evaluate_ll.chain_length=1024