wiseodd / compound-density-networks

Implementation of: Kristiadi, Agustinus, and Asja Fischer. "Predictive Uncertainty Quantification with Compound Density Networks." (2019).

Compound Density Networks

General information

1. The codes are meant to be run on a GPU.
2. The default arguments for each codes are already set so that running the codes without argument will replicate the results shown in the paper.

Instruction

1. Install the dependencies contained in requirements.txt. Remember to install pytorch with GPU support, manually if necessary.
2. Create new folder called data and run extract_features_cifar10.py.
3. Run the code on a GPU, e.g.: CUDA_VISIBLE_DEVICES=0 python ml_cdn_mnist.py.
4. Trained models will be saved in models/{dataset} directory.
5. Experiment results will be saved in results/{dataset} directory in Numpy format, i.e. use np.load to load the results.