Anytime Classification

Code for paper Towards Anytime Classification in Early-Exit Architectures by Enforcing Conditional Monotonicity.

Code for pretrained models is taken from the following repos:

• MSDNet: https://github.com/kalviny/MSDNet-PyTorch
• IMTA: https://github.com/kalviny/IMTA

Main Dependencies

• Python = 3.6
• Pytorch = 1.7

Setup

1. Clone or download this repo. cd yourself to it's root directory.
2. Create and activate python conda enviromnent: conda create --name anytime-class python=3.8
3. Activate conda environment: conda activate anytime-class
4. Install dependencies, using pip install -r requirements.txt

Code

• Main (anytime) functions are implemented in anytime_predictors.py
• Figures from the main paper are reproduced in paper_plots.ipynb. For code to run, the following steps need to be performed:
  • Download ImageNet validation dataset and store it in data\ImageNet\val
  • Download pretrained models (e.g., MSDNet) from here and store them in pretrained_models directory. Alternatively, use repos of pretrained models to train those from scratch

Acknowledgements

The Robert Bosch GmbH is acknowledged for financial support.