/BIMP

Code to reproduce the experiments of ICLR2023-paper: How I Learned to Stop Worrying and Love Retraining

Primary LanguagePython

[ICLR2023] How I Learned to Stop Worrying and Love Retraining

Authors: Max Zimmer, Christoph Spiegel, Sebastian Pokutta

This repository contains the code to reproduce the experiments from the ICLR2023 paper "How I Learned to Stop Worrying and Love Retraining". The code is based on PyTorch 1.9 and the experiment-tracking platform Weights & Biases. The code to reproduce semantic segmentation as well as NLP experiments will be added soon.

Structure and Usage

Experiments are started from the following file:

  • main.py: Starts experiments using the dictionary format of Weights & Biases.

The rest of the project is structured as follows:

  • strategies: Contains all used sparsification methods.
  • runners: Contains classes to control the training and collection of metrics.
  • metrics: Contains all metrics as well as FLOP computation methods.
  • models: Contains all model architectures used.
  • utilities: Contains useful auxiliary functions and classes.

Citation

In case you find the paper or the implementation useful for your own research, please consider citing:

@inproceedings{zimmer2023how,
title={How I Learned to Stop Worrying and Love Retraining},
author={Max Zimmer and Christoph Spiegel and Sebastian Pokutta},
booktitle={The Eleventh International Conference on Learning Representations },
year={2023},
url={https://openreview.net/forum?id=_nF5imFKQI}
}