wwxfromtju / versa

Code to reproduce experiments in "Meta-learning probabilistic inference for prediction"

Meta-Learning Probabilistic Inference For Prediction

This repository implements the models and algorithms necessary to reproduce the experiments carried out in the paper Meta-Learning Probabilistic Inference For Prediction, Gordon and Bronskill et al. It includes code for running few-shot classification experiments with Omniglot and miniImageNet, as well as for reproducing the ShapeNet view reconstruction experiments.

The code has been authored by: John Bronskill, Jonathan Gordon, and Matthias Bauer.

The main components of the repository are:

• run_classifier.py: script to run classification experiments on Omniglot and miniImageNet
• train_view_reconstruction.py: script to train view recovery models using ShapeNet objects
• evaluate_view_reconstruction.py: script to test view recovery models using ShapeNet objects
• features.py: deep neural networks for feature extraction and image generation
• inference.py: amortized inference networks for various versions of Versa
• utilities.py: assorted functions to support the repository

Dependencies

This code requires the following:

• python 2 or python 3
• TensorFlow v1.0+

Data

For Omniglot, miniImagenet, and ShapeNet data, see the usage instructions in data/save_omniglot_data.py, data/save_mini_imagenet_data.py, and data/save_shapenet_data.py, respectively.

Usage

• To run few-shot classification, see the usage instructions at the top of run_classifier.py.
• To run view reconstruction, see the usage instructions at the top of train_view_reconstruction.py and evaluate_view_reconstruction.py.

Contact

To ask questions or report issues, please open an issue on the issues tracker.

Extending the Model

There are a number of ways the repository can be extended:

• Data: to use alternative datasets, a class must be implemented to handle the new dataset. The necessary methods for the class are: __init__, get_batch, get_image_height, get_image_width, and get_image_channels. For example signatures see omniglot.py, mini_imagenet.py or omniglot.py. Note that the code currently handles only image data. Finally, add the initialization of the class to the file data.py.
• Feature extractors: to use alternative feature extractors, simply implement a desired feature extractor in features.py and change the function call in run_classifier.py. For the required signature of a feature extractor see the function extract_features in features.py.

Citation

If you use this code for your research, please cite our paper: :

@inproceedings{gordon2018metalearning,
  title={Meta-Learning Probabilistic Inference for Prediction},
  author={Jonathan Gordon and John Bronskill and Matthias Bauer and Sebastian Nowozin and Richard Turner},
  booktitle={International Conference on Learning Representations},
  year={2019},
  url={https://openreview.net/forum?id=HkxStoC5F7},
}