Official implementation of MINDE: Mutual Information Neural Diffusion Estimation

List of packages (see also requirements.txt)

bmi==1.0.0
jax==0.4.25
numpy==1.26.4
pytorch_lightning==2.2.1
scikit_learn==1.4.1.post1
torch==2.2.1
torchvision==0.17.1

MI estimation is conducted in src\minde\minde_cond.py and src\minde\minde_j.py with the functions mi_compute_sigma() and mi_compute() for both the variants (See Algorithm 2 and Algorithm 4 ).

demo.ipynb is a quick jupyter notebook to launch MINDE.

The training is conducted in src\libs\train_step_cond.py and src\libs\SDE.train_step.py( See Algorithm 1 and 3)

Make sure the requirements are satisfied in your environment. To run minde_c and minde_j and baselines, you can specify the number of tasks by choosing task_0 and task_n which corresponds to the number of the task in in paper.

python -m scripts.eval_minde_c --task_0 0 --task_n 2
python -m scripts.eval_minde_j --task_0 0 --task_n 2
python -m scripts.eval_baseline --task_0 0 --task_n 2

Please refer to src/consistency_tests/ to run the consistency tests. Please note that it's first required to train the autoencoder by running python -m src.consistency_tests.Autoencoder --rows 10.