kristpapadopoulos / neural-processes

Keras, Tensorflow eager execution implementation of Neural Processes

Neural Processes

Keras, Tensorflow Probability and Eager Execution Implementation

Neural Processes for 1D regression implemented as per paper: Neural Processes

Code developed from:

1. Kasper Martens: https://kasparmartens.rbind.io/post/np/

2. Chris Orm https://chrisorm.github.io/NGP.html

File: np_v0.1.py - Oct 11, 2018

• Tensorflow 1.10.0
• Numpy 1.14.5
• Epochs = 10001
• Learning Rate = 0.001

Items for further development:

• Deeper and wider layers
• Different activation functions
• Different priors and latent space transformations
• Learning over mutiple related functions

Figures: 30 Neural Process Function Draws for Different Training Epochs

• training context x range -2 to 2 and output range sin(x)
• test x range -10 to 10 to estimate posterior function samples