Code for the paper:
Semi-Amortized Variational Autoencoders
Yoon Kim, Sam Wiseman, Andrew Miller, David Sontag, Alexander Rush

The code was tested in python 3.6 and pytorch 0.2. We also require the h5py package.

The raw datasets can be downloaded from here.

Text experiments use the Yahoo dataset from Yang et al. 2017, which is itself derived from Zhang et al. 2015.

Image experiments use the OMNIGLOT dataset Lake et al. 2015 with preprocessing from Burda et al. 2015.

Please cite the original papers when using the data.

After downloading the data, run

python preprocess_text.py --trainfile data/yahoo/train.txt --valfile data/yahoo/val.txt
--testfile data/yahoo/test.txt --outputfile data/yahoo/yahoo

This will create the *.hdf5 files (data tensors) to be used by the model, as well as the *.dict file which contains the word-to-integer mapping for each word.

The basic model command is

python train_text.py --train_file data/yahoo/yahoo-train.hdf5 --val_file data/yahoo/yahoo-val.hdf5
--gpu 1 --checkpoint_path model-path

where model-path is the path to save the best model and the *.hdf5 files are obtained from running preprocess_text.py. You can specify which GPU to use by changing the input to the --gpu command.

To train the various models, add the following:

• Autoregressive (i.e. language model): --model autoreg
• VAE: --model vae
• SVI: --model svi --svi_steps 20 --train_n2n 0
• VAE+SVI: --model savae --svi_steps 20 --train_n2n 0 --train_kl 0
• VAE+SVI+KL: --model savae --svi_steps 20 --train_n2n 0 --train_kl 1
• SA-VAE: --model savae --svi_steps 20 --train_n2n 1

Number of SVI steps can be changed with the --svi_steps command.

To evaluate, run

python train_text.py --train_from model-path --test_file data/yahoo/yahoo-test.hdf5 --test 1 --gpu 1

Make sure the append the relevant model configuration at test time too.

After downloading the data, run

python preprocess_img.py --raw_file data/omniglot/chardata.mat --output data/omniglot/omniglot.pt

To train, the basic command is

python train_img.py --data_file data/omniglot/omniglot.pt --gpu 1 --checkpoint_path model-path

To train the various models, add the following:

• Autoregressive (i.e. Gated PixelCNN): --model autoreg
• VAE: --model vae
• SVI: --model svi --svi_steps 20
• VAE+SVI: --model savae --svi_steps 20 --train_n2n 0 --train_kl 0
• VAE+SVI+KL: --model savae --svi_steps 20 --train_n2n 0 --train_kl 1
• SA-VAE: --model savae --svi_steps 20 --train_n2n 1

To evaluate, run

python train_img.py --train_from model-path --test 1 --gpu 1

Make sure the append the relevant model configuration at test time too.

Some of our code is based on VAE with a VampPrior.

MIT