This is a PyTorch implementation of the paper titled SelectiveNet: A Deep Neural Network with an Integrated Reject Option [Geifman and El-Yaniv, ICML2019]
A deep neural network architecture with an integrated reject option that can be trained end-to-end for classification and regression tasks.
Install requirements using pip install -r requirements.txt
I run the code with Pytorch 1.10.0, CUDA 10.2
Note: In the default version, you need Weights and Biases for logging the metrics and saving checkpoints when running train.py. In addition, the default path to load checkpoints from is Weights and Biases log path. You can disable Weights and Biases in training by using --unobserve as an input argument to train.py and changing log_path to a desired local directory for metric logging and checkpoint saving. Following this, you can disable Weights and Biases in test time by using --unobserve as an input argument. If checkpoints are saved locally, set input argument --checkpoint to the local directory and set --weight to the name of the checkpoint in test.py.
Use scripts/train.py to train the model. Example usage:
# Example usage
cd scripts
python train.py --dataset cifar10 --coverage 0.7 Use scripts/test.py to test the network. Example usage:
# Example usage (test single weight)
cd scripts
python test.py --dataset cifar10 --exp_id ${id_of_training_experminet} --weight ${name_of_saved_model}--coverage 0.7
# Example usage (test multiple weights)
cd scripts
python test.py --dataset cifar10 --exp_id 2fkl0ib7 --coverage 0.7- Implementation borrows from https://github.com/gatheluck/pytorch-SelectiveNet.