Variation Autoencoder with Missing values

A Pytorch implelmentation of VAE for handling missing data in test. Here is my medium post.

Requirements:

• pytorch
• numpy
• matplotlib
• tqdm

How to run

This is a sample code to working with a VAE in pytorch in with missing data. There are two experiments with two datasets: MNIST and synthetic timeseries data. In the following you can find how to run each experiment.

Running with MNIST

To train VAE:

python main.py --epochs 100 --train_mode mnist

To test with partial data:

python tests.py --test_partial_data_mnist

An example of sampling output:

An example of reconstruction output:

An example of reconstruction output with partial data:

Running with synthetic timeseries

First we need to generate synthetic data for both train and test data

python tests.py --generate_timeseries_synthetic_train
python tests.py --generate_timeseries_synthetic_test

And then train VAE model:

python main.py --epochs 100 --train_mode synthetic_timeserie

And finally test with partial data

python tessts.py --test_partial_data_synthetic_ts

An example of reconstruction:

An example of reconstruction with partial data: