Inducing Point Neural Processes

This repository is used to implement the Neural Processes experiments from Semi-Parametric Inducing Point Networks and Neural Processes.

Acknowledgements

This repository is largely based on the implementation of Transformer Neural Processes, which in turn was based on Bootstrapping Neural Processes.

Install

Install the dependencies as listed in ipnp_env.yml and activate the environment:

conda env create -f ipnp_env.yml
conda activate ipnp

Usage

Use the run_regression_trainer_slurm.sh to launch regression jobs using a SLURM queue manager:

./run_regression_trainer_slurm.sh \
  --exp <gp |celeba> \
  --mode <train | eval> \
  --model <anp | banp | canp | ipanp | ipcanp> \
  --seed <seed> \
  --max_num_ctx <max_num_ctx> \
  --min_num_ctx <min_num_ctx> \
  --max_num_tar 64 \
  --min_num_tar 4

To run the trainer.py file directly:

cd ./regression
python trainer.py <gp | celeba> <train | eval> \
  --expid <unique_exp_id> \
  --train_seed=<seed> \
  --eval_seed=<seed> \
  --model=<anp | banp | canp | ipanp | ipcanp> \
  --max_num_ctx=<max_num_ctx> \
  --min_num_ctx=<min_num_ctx> \
  --max_num_tar=64 \
  --min_num_tar=4

Citation

If you find this repo useful in your research, please consider citing our paper:

@inproceedings{
    rastogi2023semiparametric,
    title={Semi-Parametric Inducing Point Networks and Neural Processes},
    author={Richa Rastogi and Yair Schiff and Alon Hacohen and Zhaozhi Li and Ian Lee and Yuntian Deng and Mert R. Sabuncu and Volodymyr Kuleshov},
    booktitle={The Eleventh International Conference on Learning Representations },
    year={2023},
    url={https://openreview.net/forum?id=FE99-fDrWd5}
}