Exploiting Epistemic Uncertainty at Inference Time

This repository contains the code used for the Paper titled Exploiting epistemic uncertainty at inference time for early-exit power saving.

It also contains some of the code used in the paper Adapting Branched Networks to Enable Progressive Intelligence, namely in the folder main/models.

The repository is organised as such:

• main
  • sub directories:
    • models/:
      • code for producing ResNet models of any depth/width, and attaching any number of branches to them. There is also code for producing 'slimmable' models, with 3 operating widths at inference time.
    • notebooks/:
      • Analysis notebooks
    • utils/:
      • Utility files for calling models, dataset, etc.
    • ptflops/:
      • Adapted code from this repository used for calculating model parameter counts and MAC usage.
  • scripts:
    • train.py
      • Code for training the branched models
    • slimmable_train.py
      • Code for training 'slimmable' branched models
    • get_embeddings.py
      • Code for retrieving the embeddings on the train/validation set
    • knn_OOD.py
      • This performs the knn measurements using the OOD test sets
    • get_predictions.py
      • This performs the inference on the test set, to obtain the accuracy on each ID set, it also saves the raw outputs
• bash-scripts
  • This contains bash scripts for running experiments using SLURM workload manager
  • These can be called to run the experiments of the paper
• trained-models
  • Trained models are stored here, along with any of their evaluation results