This is the official code repository for the ECCV 2020 paper: "GroSS: Group-Size Series Decomposition for Grouped Architecture Search"
Currently, the released code supports decomposition, training and evaluation of the configurations included in the results section of the paper. We aim to release scripts to perform search soon.
- Create and activate a new conda environment:
conda create -n gross python=3.7 & conda activate gross - Install pytorch with the correct version of CUDA for your machine (tested on PyTorch v1.3.1 and CUDA 10.1), eg:
conda install pytorch=1.3.1 torchvision=0.4.2 cudatoolkit=10.1 -c pytorch - Install the requirements:
pip install -r requirements.txt
Various configurations use either CIFAR-10 or ImageNet.
For greatest convenience, download these datasets in a directory called data, such that the project structure is:
GroSS
├- configs
├- data
├- cifar
└- imagenet
├- dataset
⋮
Alternatively, you can modify the DATASET.ROOT_DIR in the configuration files.
To decompose and/or train a configuration run the command: python decompose_and_ft.py -c PATH/TO/CHOSEN/CONFIGURATION/FILE
For evaluation of a configuration, run the command: python test_config.py -c PATH/TO/CHOSEN/CONFIGURATION/FILE
@misc{howardjenkins2019gross,
title={GroSS: Group-Size Series Decomposition for Grouped Architecture Search},
author={Henry Howard-Jenkins and Yiwen Li and Victor A. Prisacariu},
year={2019},
eprint={1912.00673},
archivePrefix={arXiv},
primaryClass={cs.LG}
}