This release contains pretrained models used by the various experiments to be used with the training scripts. The code in the linked tag matches the code used in the TMLR experiments.
There are four types of models uploaded:
- Mean: cached average value in the dataset for mean imputation.
- Baseline: classifier mapping from clean 3 channel images with no missing values to probabilities.
- Robust: classifier mapping from mutated 4 channel images with missing values to probabilities.
- DMV: classifier mapping from mutated 4 channel images with missing values to Dirichlet alpha parameters.
Arguments used to create these models are included in the repository at the relative path in the log folder or in the attached logs zip file.
CelebA
Models for usage on the CelebA HD faces dataset. See scripts/setup/unpack_celeba.sh.
Baseline
celeba-mean.pklz: Originally located atmodels/mean/celeba.pklz.celeba-baseline-Blond_Hair-20250508-133714.pklz: Originally located atmodels/celeba/Blond_Hair/celeba-20250508-133714.pklz.celeba-baseline-Eyeglasses-20250508-134104.pklz: Originally located atmodels/celeba/Eyeglasses/celeba-20250508-134104.pklz.celeba-baseline-Smiling-20250508-133956.pklz: Originally located atmodels/celeba/Smiling/celeba-20250508-133956.pklz
DMV
celeba-dmv-Blond_Hair-20250514-005354.pklz: Originally located atmodels/dirichlet-celeba/Blond_Hair/celeba-20250514-005354.pklz.celeba-dmv-Eyeglasses-20250514-004153.pklz: Originally located atmodels/dirichlet-celeba/Eyeglasses/celeba-20250514-004153.pklz.celeba-dmv-Smiling-20250513-220427.pklz: Originally located atmodels/dirichlet-celeba/Smiling/celeba-20250513-220427.pklz.
MNIST
Models for use on MNIST digits. Included via the Python package.
mnist-mean.pklz: Originally located atmodels/mean/mnist.pklz.mnist-baseline-20250919-194130.pklz: Originally located atmodels/mnist/mnist-20250919-194130.pklz.mnist-robust-20251003-153019.pklz: Originally located atmodels/robust/mnist/mnist-20251003-153019.pklz.mnist-dmv-20250919-194130.pklz: Originally located atmodels/dirichlet-mnist/mnist-20250919-235704.pklz.
CIFAR10
Models for use on CIFAR10 multiclass. Included via the Python package.
cifar10-mean.pklz: Originally located atmodels/mean/cifar10.pklz.cifar10-baseline-20250515-150106.pklz: Originally located atmodels/cifar10/cifar10-20250515-150106.pklz.cifar10-robust-20251003-153001.pklz: Originally located atmodels/robust/cifar10/cifar10-20251003-153001.pklz.cifar10-dmv-20250611-062342.pklz: Originally located atmodels/dirichlet-cifar10/cifar10-20250611-062342.pklz.
StarCraftCIFAR10
Models for use on StarCraftCIFAR10 multiclass. Included via the Python package. Robust and DMV include 3 variants:
- MCAR: trained on MCAR data.
- MNAR high: trained on MNAR data using a sharpness of
-1.5, causing it to be more likely to drop when units are present. - MNAR low: trained on MNAR data using a sharpness of
1.5, causing it to be more likely to drop when units are absent.
Baseline
starcraft-cifar10-mean.pklz: Originally located atmodels/mean/starcraft.pklz.starcraft-cifar10-baseline-20250515-150053.pklz: Originally located atmodels/starcraft-cifar10/starcraft-20250515-150053.pklz.
Robust
starcraft-cifar10-robust-mcar-20251008-021131.pklz: Originally located atmodels/robust/starcraft-cifar10/starcraft-20251008-021131.pklz.starcraft-cifar10-robust-mnar-high-20260430-194422.pklz: Originally located atmodels/robust/starcraft-cifar10/starcraft-20260430-194422.pklz.starcraft-cifar10-robust-mnar-low-20260430-194429.pklz: Originally located atmodels/robust/starcraft-cifar10/starcraft-20260430-194429.pklz.
DMV
starcraft-cifar10-dmv-mcar-20250613-033643.pklz: Originally located atmodels/dirichlet-starcraft-cifar10/starcraft-20250613-033643.pklz.starcraft-cifar10-dmv-mnar-high-20260430-194354.pklz: Originally located atmodels/dirichlet-starcraft-cifar10/starcraft-20260430-194354.pklz.starcraft-cifar10-dmv-mnar-low-20260430-194415.pklz: Originally located atmodels/dirichlet-starcraft-cifar10/starcraft-20260430-194415.pklz.
Other files
training-logs.zip: Arguments and training logs for training all above models.results.zip: Experiment logs and result CSV files.