Code for the visualization experiment of "Dimensionality Reduction for Representing the Knowledge of Probabilistic Models" (ICLR 2019)

The code was tested on pytorch 0.3.0

Run the following code to train low-dimensionality representations for the test set of CIFAR-100:

 python DRPR_train.py

The dataset name can be changed with the variable called dataset which can take the following values: "cifar10", "cifar100", "mnist" or "stl10".

It saves the low-dimensional representations (2d by default) in the created folder dataset_learned_representations where dataset is the name of the dataset.

Dataset folders

Each dataset folder "cifar10", "cifar100", "mnist" and "stl10" contains the following files:

• original_data.txt contains the output representations of a CNN (before the softmax activation).
• proba_data_tau.npy where tau is the value of the temperature. More details can be found in the appendix of the paper (Section C.1.2). It corresponds to applying the script called "proba_extraction.m" in the scripts folder.
• labels.txt contains the labels of the examples.

Plotting the learned low-dimensional representations

We provide plot functions in matlab in the scripts folder (their name prefix is plot).