Mixup-Data-Dependency

Code associated with the paper "Towards Understanding the Data Dependency of Mixup-style Training".

Running Alternating Line Experiments

In order to generate the plots found in Section 2.3 ("A Mixup Failure Case"), one can run the following command for different values of alpha.

python3 tasks/train_models.py --task-name NCAL --alpha 128 --num-runs 10

If running using slurm, it is also possible to just run:

./tasks/run_task_with_erm.sh NCAL 128 10 0

The generated output files can be found under runs/ and plots/ with file names based on the provided parameters.

Running Image Classification Experiments

In order to generate the plots found in Section 2.4 ("Sufficient Conditions for Minimizing the Original Risk"), one can run the following commands for different values of alpha.

python3 tasks/train_models.py --task-name MNIST --alpha 1024 --num-runs 5
python3 tasks/train_models.py --task-name CIFAR10 --alpha 1024 --num-runs 5
python3 tasks/train_models.py --task-name CIFAR100 --alpha 1024 --num-runs 5

Once again, if running using slurm it is possible to instead run ./tasks/run_task_with_erm.sh with the same arguments as above and an additional fourth argument set to 0. As before, output files can be found in runs/ and plots/.

Running Angular Distance Analysis

To recreate the approximate epsilon computation found in Section 2.4 (in the discussion of application of sufficient conditions), one can run the following command after manually setting subset_prop and alpha in analysis/mixup_point_analysis.py.

python3 analysis/mixup_point_analysis.py

Running Two Moons Experiments

To recreate the two moons experiments found in Section 3.1 ("The Margin of Mixup Classifiers"), set alpha_1 and alpha_2 in tasks/two_moons/py to the mixing parameters to be compared and then run the following command.

python3 tasks/two_moons.py