Neural-Processes

This repository contains PyTorch implementations of the following Neural Process variants

• Recurrent Attentive Neural Process (ANPRNN)
• Neural Processes (NPs)
• Attentive Neural Processes (ANPs)

The model architectures follow the ones proposed in the papers.

• Marta Garnelo, et al. "Neural Process" ICML 2018 Workshop.
• Hyunjik Kim, et al. "Attentive Neural Processes" ICLR 2019.

This is the author implementation of NeurIPS 2019 Workshop paper

• Shenghao Qin, Jiacheng Zhu, et al. "Recurrent Attentive Neural Process for Sequential Data"

Installation & Requirements

It is recommended to use Python3. Using virtual environment is recommended.

PyTorch 
Numpy
Matplotlib

Descriptions

• The Neural Process models are under /neural_process_models folder
• In /neural_process_models/modules there are all the modules for building NP networks, including linear MLP, attention network and so on.
• The data generation functions are under \misc

Results

1d function regression

Usage

Simple example of training an Attentive Neural Process of 1d function regression

python3 main_ANP_1d_regression.py

For digital number inpainting trained on MNIST data

python3 main_ANP_mnist.py

See Neural Process models in /neural_process_models for detailed examples of construction NP models for specific tasks.

Acknowledgements

For any question, please refer to

• Eric Ma (yingchem@umich.edu)
• Jiacheng Zhu (jzhu4@andrew.cmu.edu)

License

MIT