Your copula is a classifier in disguise!

This repository contains code from my paper 'Your copula is a classifier in disguise: classification-based copula density estimation' accepted at AISTATS 25.

Paper: https://arxiv.org/abs/2411.03014

Instruction to reproduce experiments from paper

• Figure 1: Our model estimates copula densities through classification. The code for this is in the Figure 1/figure1.ipynb notebook, and can be run in less than a minute on a CPU. It can also serve as a good starting point to using Ratio copulas on your own!

• Figure 2: 2D copula models trained on monochromatic images. Use the Figure 2/2d_single_image.ipynb notebook. It will load the relevant samples and densities and plot the figure. The notebook also includes the code for the IGC and TLL/vine copulas. For the Ratio Copula, training is done with Figure 2/2d_image_cop_simpleNNET.py and Figure 2/2d_image_einstein_simpleNNET.py.py for each of the pictures.

• Figure 3: Box plots showing the average LL across 25 fits on samples from different parametric copulas. The files for this figure are contained in 2D copulas. The 2d_experiments_plot.ipynb notebook can be run to produce the figure; it will load all experiment data and reproduce the exact same figure as in the paper. Python scripts (2dexperiments_L_ratio.py and 2dexperiments_NNet_ratio.py) can be run for the ratio copula models while the parametric copula benchmarks are run from the 2d_experiments_plot.ipynb notebook directly (you will need to uncomment those parts).

• Figure 4: Example equivalent classifiers for six parametric copulas. There is a notebook in Figure 4/Figure 4.ipynb with all the code to produce this figure.

• Figure 5: Using other losses for better tail modelling. There is a single notebook Figure 5/Rebuttal_exp copy.ipynb with all the code needed to reproduce the figure.

• Figure 6: Digits samples from copula models. The notebook Figure 6/digits_exp.ipynb has the code to load samples and produce the plot. The samples can also be found in the folder.

• Figure 7: MNIST samples from copula models. The notebook Figure 7/Figure 7.ipynb has the code to load samples and produce the plot. The samples can also be found in the folder.

• Table 1: Log-likelihoods and Wasserstein-2 on high-dimensional datasets. There are two folders inside: Table 1/MNIST_exp and Table 1/Digits_exp. Each contains python files for training and sampling from the copula models.