Copyright (c) 2020-2022 the authors : Neural Processes with Stochastic Attention

Description of experiments : Neural Processes with Stochastic Attention

There are 4 experiments in this code.

• gp / gp_test : 1D regression task.
• recommend_sys : movieLenz-10k.
• lv_model : predator-prey model.
• celebA : image completion task.

The structure of directory

• recommend_sys / lv_model / celebA include train, test codes
  "./save_models/" : location of model checkpoints in training. "./test_save_models/" : model path in testing.

• In gp / gp_test, we separate train, test codes. gp : train code.
  "./save_models/" : location of model checkpoints in training.
  gp_test : test code.
  "./save_models/" : model path in testing.

Dataset

• gp : it is synthetic dataset, so it dose not explicitly requires the data file.
• recommend_sys : "./data/movielens".
• lv_model : only test dataset (hudson's hare lynx) you can assess to "./data/dataset/LynxHare.txt"
  For the training dataset, you should refer "https://github.com/juho-lee/bnp". . you have to run "/regression/data/lotka_volterra.py" and then obtain "train.tar" / "val.tar"
  . you must place "train.tar" and "val.tar" on "./lv_model/data/dataset/".
• celebA : you can download at "https://www.kaggle.com/jessicali9530/celeba-dataset".
  . you should write the absolute root path of celebA dataset on the line 50 and 123 in "./celebA/main.py".

Run

- At each directory, there is "readme.txt", which describe how to run train and test codes.

Pre-trained checkpoints

- gp_test / lv_model / recommend_sys : there exist pre-trained checkpoints. 
- celebA : There are no pre-trained checkpoint due to space constraints.