Clean slate to begin research on the thermodynamic variational objective. The heart of the code is src/models/base.py, which is where most of the training logic lives. All new models should extend it to get access to the various losses. The outer training loop which calls all the code is main.py.
It uses sacred for its command line interface and wandb for experiment tracking. All hyperparameters are defined in my_config() and are available on the commandline via:
python main.py with model_name='continuous_vae' loss='tvo' lr=0.001 --unobserved
To save to your wandb database, drop the --unobserved:
python main.py with model_name='continuous_vae' loss='tvo' lr=0.001
Only strings, floats, ints, lists, and dicts should be defined in my_config(), everything else (numpy arrays, pickles, etc) should be instantiated in init().
- tvo
- tvo_reparam
- reinforce
- elbo
- iwae
- iwae dreg
- vimco
- wake-wake (ww)
- wake-sleep (ws)
- mnist
- fashion mnist
- kuzushiji mnist
- omniglot
- binarized mnist
- binarized omniglot
- pcfg
- log
- linear
- moments
- gp_bandits
- continuous vae
- discrete vaes
- bayesian neural networks (bnn)
- probabilistic context free grammar (pcfg)