Training

(1) Stage 1: Supervised Contrastive Training

Please refer this repository for PyTorch implementation of Supervised Contrastive Learning paper.

(2) Stage 2: Virtual Outlier Synthesis

Following the paper and the repository, we generate virtual outliers. To synthesize a new set of virtual outliers, use --generate_ood while running the train script.

(3) Stage 3: OOD Score Optimization

python train_virtual.py --epochs 20
  --learning_rate 0.1
  --batch_size 128
  --momentum 0.9
  --save 0.9 Folder_to_save_checkpoints
  --generate_ood # to generate a new set of virtual outliers
  --generate_centroids  # find ID class centroids

(4) Stage 4: ID Classification Head Training

python stage4_ID_classifier.py --train

Testing

(1) OOD Classification

To run the results for FPR95, AUROC and AUPR please run:

        python test.py --cifar_root <root path of the cifar dataset> (required)
                       --weights_path <path to weight of the model> (required)
                       --centroids_path <path to the centroids file> (required)
                       --batch_size <batch size of test batches>
                       --lsun_root <path to root directory of lsun dataset>
                       --places365_root <path to root directory of places dataset>
                       --isun_root <path to root directory of isun dataset>
                       --dtd_root <path to root directory of dtd dataset>
                       --svhn_root <path to root directory of svhn dataset>

For faster results it is better to run the script once per OOD dataset i.e with only one of the OOD datset path arguments at a time.

(2) ID Classification To use the model as a classifier for ID data:

python stage4_ID_classifier.py --test

Visualization

python visualize_embeddings.py

Ensure we have the datasets for Textures (DTD), SVHN, iSUN, LSUN and Places 365 for testing and visualization.

The pretrained models and PyTorch objects are present in drive.