ECNP Code

This is the repository for the Code of Evidential Conditional Neural Processes

Appendix

The appendix and supplimentary works of the paper are provided in appendix.pdf file

This code requires

• Python3
• Pytorch

The datasets used are

• Synthetic Datasets: Sinusoidal regression and GP regression (generated in the code itself)
• Real world Image Completion Datasets: MNIST, CelebA and Cifar10

Download the datasets and organize the datasets with following structure

datasets/
├── celeba
│   ├── test
│   ├── train
│   └── val
├── cifar10
│   ├── test
│   └── train
└── mnist
    ├── test
    └── train

Celeba has just 1 folder in both train and test consisting of all the task imags
MNIST and Cifar10 should have 10 folders (numbered 0 - 9 corresponding to the 10 classes) in both train and test.

Running evidential models

enp_run.py is used to run ECNP-A and ECNP
For eg. to run 50-shot image-completion experiment with mnist for ECNP, use
python3 enp_run.py --dataset "mnist" -use_det "true" -use_lat "false" --max_context_points 50 --model_type "CNP"

Running baseline models (NP, CNP, ANP)

np_baseline.py script can be used to run the baseline models
For eg. to run 50 shot image completion experiment with mnist for CNP, use
python3 np_baseline.py --dataset "mnist" -use_det "true" -use_lat "false" --max_context_points 50 --model_type "CNP"

Additionally, various hyperparameters and settings can be specified using the utilFiles/get_args.py file

Reference codes used for the repository

• ANP anc CNP Official Codes - https://github.com/deepmind/neural-processes