PyTorch implementation of Neural Processes (NP) by Garnelo et al [https://arxiv.org/abs/1807.01622]
MNIST image completion
The task is to complete an image given some number [1;784] of context points (coordinates) at which we know the greyscale pixel intensity [0;1].
The first row shows the observed greyscale context points. Unobserved pixels are in blue.
The five rows below show realisations of different samples of the global latent variable
z given the context points above. Compare with Figure 4 in the paper.
10 context points
100 context points
300 context points
784 context points (full image)
How to run
python main.py produces the results above. The script saves examples of reconstructed images at the end of every epoch in
- Python 3
- PyTorch 0.4.1
Other NP implementations
R + TensorFlow - [https://github.com/kasparmartens/NeuralProcesses]