Awesome

MAPLE

Code for "Gaussian Latent Representations for Uncertainty Estimation Using Mahalanobis Distance in Deep Classifiers", ICCV Workshop 2023.

Dataset

The dataset is arranged such that each class has a directory with the corresponding images placed in them. An example directory structure is shown below.

├── dataset
│   ├── train_data
│   │   ├── class1
│   │   ├── class2
...
│   │   ├── classN
│   ├── test_data
│   │   ├── class1
│   │   ├── class2
...
│   │   ├── classN

Each dataset is followed by a csv file containing the class name and the corresponding classification label. An example for CIFAR10 is given in data/cifar10.csv.

Before launching the training, please make sure that the dataset paths and the id paths (csv files) are included in the config.py.

Training

The necessary python libraries and their versions to run the code are specified in requirements.txt. The hyperparameters and arguments needed for training the network are available in config.py. Depending on the dataset used, please make sure to change the respective hyperparameters. To launch the training, run

python3 train.py

The code automatically splits the dataset into train and validation.

Inference

To launch the inference, run

python3 mahalanobis_calculation.py

This calculates the Mahalanobis distance and the prediction probability for both the in distribution and out-of-distribution dataset, and computes the in distribution and out-of-distribution metrics.

If you use this code, please cite the following paper:

Aishwarya Venkataramanan, Assia Benbihi, Martin Laviale, Cedric Pradalier. Gaussian Latent Representations for Uncertainty Estimation using Mahalanobis Distance in Deep Classifiers. In Proceedings of the IEEE/CVF International Conference on Computer Vision. 2023. p. 4488-4497.

@inproceedings{venkataramanan2023gaussian,
  title={Gaussian Latent Representations for Uncertainty Estimation using Mahalanobis Distance in Deep Classifiers},
  author={Venkataramanan, Aishwarya and Benbihi, Assia and Laviale, Martin and Pradalier, C{\'e}dric},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={4488--4497},
  year={2023}
}