Amortized Adaptive Importance Sampling (AAIS) for variational autoencoder

This repository provides an implementation for reproducing the results in the paper, "Amortized Adaptive Importance Sampling for variational autoencoder"

Requirement

• python==3.11.9
• torch
• pandas
• scipy
• matplotlib
• umap

Their versions are unimportant because we used these packages' fundamental functions.

How to start

There are only two arguments: seed or datatype. All datasets and the associated hyperparameters had already been stored in the config and data folders, respectively. If you want to explore the data generation process, please check data_generation.ipynb, data_generation.R, and GLMM.R for VerbAgg.

Example

$ python main.py -s 1 -d exp

• --seed (or -s): Seed. Default = 42.
• --dtype (or -d): Datatype. Possible arguments are exp, mix, copula, MNIST, cortex, VerbAgg. Default = exp.

Output

After training, the evaluation results are automatically saved. model_states.pt and model_kwargs.json are always saved. Qualitative and quantitative results are different for each experiment. For example, if you run with -d exp, histogram.png visualizes latent variable modeling on exp data, and distances.json has evaluation metrics used for the histogram.

Customization

For customization, you can prepare your data and create a config file with your desired values. The data should be loaded through the dataloader.py file, and the config file must be placed in the config folder. If the likelihood for your data is not included in nll.py, you need to add it yourself. If you want to perform an evaluation, add your custom function to eval.py. Furthermore, if you want to use another architecture, define a new model class in model.py and integrate it appropriately with the LatentModel class.