PyTorch implementation of the inversion variational autoencoder (I-VAE).
git clone https://github.com/sisl/I-VAE
cd I-VAE/ivae
pip install .from ivae import IVAE, train
model = IVAE(state_channels, obs_channels, latent_dim, dropout_rate)
for epoch in range(epochs):
train(model, dataloader, val_dataloader, optimizer, device, epoch, epochs)Where dataloader holds tuples of (state, obs). In the provided implementation, the state is an 80×80 tensor and the observations fill out a 200×200 tensor.
Example applied to MNIST given partial pixel observations.
Other examples applied to geological inversion, used as a way to update a POMDP belief through posterior state samples.
In progress