inference-suboptimality

Code regarding evaluation for paper Inference Suboptimality in Variational Autoencoders. [arxiv]

Dependencies

• python3
• pytorch==0.2.0
• tqdm

Training

To train on MNIST and Fashion, unzip the compressed files in folder datasets/.

python run.py --train --dataset <dataset> (--lr-schedule --warmup --early-stopping)

To train on CIFAR, set the argument for the dataset flag to cifar. The dataset should be downloaded automatically, if not already downloaded.

Evaluation

• IWAE: python run.py --eval-iwae --dataset <dataset> --eval-path <ckpt path>
• AIS: python run.py --eval-ais --dataset <dataset> --eval-path <ckpt path>
• Local FFG: python local_ffg.py --dataset <dataset> --eval-path <ckpt path>
• Local Flow: python local_flow.py --dataset <dataset> --eval-path <ckpt path>
• BDMC: python bdmc.py --eval-path <ckpt path> --n-ais-iwae <num samples> --n-ais-dist <num dist>

Other Experiments

For decoder size, flow affect amortization, test set gap and other experiments, refer to this.

Citation

If you use our code, please consider cite the following: Chris Cremer, Xuechen Li, David Duvenaud. Inference Suboptimality in Variational Autoencoders.

@article{cremer2018inference,
  title={Inference Suboptimality in Variational Autoencoders},
  author={Cremer, Chris and Li, Xuechen and Duvenaud, David},
  journal={ICML},
  year={2018}
}