This repository is pytorch based implementation of Neural Process. paper is here: https://arxiv.org/abs/1807.01622
Neural Process is a class of neural latent variable models. This models combine the best of Neural Network and Gaussian Process.
This figure is a graphical model of neural process. Nerual Process has two model, encoder, aggregater, and conditional decoder. An encoder is parametarised as neural network, and it from input space into representation space that takes in pairs of (x, y) context values. An aggregater that summarises the encoded inputs. A conditional decoder g that takes as input the sam- pled global latent variable z as
In this repository, I implement the few-shot learning experiments using MNIST. This task is a image completation as a 2D regression task. inputs context data point and complete image by predicting luck data point
You can run experiments using Docker:
docker-compose -f docker/docker-compose-cpu.yml build
docker-compose -f docker/docker-compose-cpu.yml run experiment python3 experiment.py
[1] M. Garnelo et al., “Neural Processes.”