Vastlab / SSFiOWL

Self-Supervised Features Improve Open-World Learning

This repo reproduces results from the paper "Self-Supervised Features Improve Open-World Learning"

Please use the following BibTex to cite this work.

article{dhamija2021self,
  title={Self-Supervised Features Improve Open-World Learning},
  author={Dhamija, Akshay Raj and Ahmad, Touqeer and Schwan, Jonathan and Jafarzadeh, Mohsen and Li, Chunchun and Boult, Terrance E},
  journal={arXiv preprint arXiv:2102.07848},
  year={2021}
}

Dependencies

This repo is dependent on the repo https://github.com/Vastlab/vast that contains some useful functionality for various projects at VastLab. Please install it using pip install git+https://github.com/Vastlab/vast.git.

Feature Extraction

We pre-extract features from the self supervised networks and use them for all our experiments. For extracting the features please use either FromCSV.py or FromDirectoryStructures.py based on how your data is structured. The scripts are present at https://github.com/Vastlab/vast/tree/main/vast/scripts/FeatureExtractors.

Non-Backpropagating Incremental Learning (NIL)

Sample command used to run incremental learning experiments using the NIL approach

time python NIL.py --training_feature_files {Feature_Path}/resnet50/imagenet_1000_train.hdf5 \
 --validation_feature_files {Feature_Path}/resnet50/imagenet_1000_val.hdf5 \
 --layer_names avgpool --OOD_Algo EVM --tailsize 1. --distance_metric euclidean \
 --initialization_classes 50 --total_no_of_classes 100 --new_classes_per_batch 10 \ 
 --output_dir /tmp/ --distance_multiplier 0.7 --no_of_exemplars 20

Non-backpropagting Open World Learning (NOWL)

Sample command used to run open world learning experiments using the NOWL approach

time python NOWL.py --training_feature_files {Feature_Path}/resnet50/imagenet_1000_train.hdf5 \
 --validation_feature_files {Feature_Path}/resnet50/imagenet_1000_val.hdf5 \
 --layer_names avgpool --OOD_Algo EVM --tailsize 1. --distance_metric euclidean \
 --initialization_classes 50 --total_no_of_classes 100 --new_classes_per_batch 10 \
 --output_dir /tmp/ --distance_multiplier 0.7 --no_of_exemplars 20 --cover_threshold 0.7 \
 --known_sample_per_batch 2500 --unknown_sample_per_batch 2500 --initial_no_of_samples 15000