quicktune 0.0.4

Quick-Tune: Quickly Learning Which Pretrained Model to Finetune and How

QuickTune (WIP)

Quick-Tune: Quickly Learning Which Pre Trained Model to Fine Tune and How ICLR2024

This repo contains the code for running experiments with QuickTune

Run QuickTune

Prepare Environment

To install QuickTune, you can simply use pip:

pip install quicktune

This project depends on a custom version of timm, which is not available on PyPI. You can install it by running the following command:

pip install git+https://github.com/rapanti/qt_timm

Download the QuickTune Meta-Dataset:

wget https://rewind.tf.uni-freiburg.de/index.php/s/oMxC5sfrkA53ESo/download/qt_metadataset.zip
unzip qt_metadataset.zip

Download the metalearned Optimizer

wget https://rewind.tf.uni-freiburg.de/index.php/s/XBsMjps5n3N9we6

Prepare Custom Dataset

The custom dataset must be in Pytorch's ImageFolder format, e.g. download the Imagenette dataset:

wget https://s3.amazonaws.com/fast-ai-imageclas/imagenette2-320.tgz
tar -xvzf imagenette2-320.tgz

Modify the quicktuning script in the examples folder to your needs.