AutoMix tutorial that is independent of MMClassification

[lofrienger]

Dear Authors,

Inspiring work!
Could you please provide a simplified demo/tutorial of how to use AutoMix in general PyTorch training? I have no experience of using OpenMMLab-style framework.
A small demo like google colab notebook would be of great help!

Thanks!

[Lupin1998]

Hi,
Thanks for your advice @lofrienger! We are working on a simplified version of AutoMix for more practical usage, but it might require quite a long time. Therefore, we plan to provide a demo of AutoMix based on the Jupyter notebook in OpenMixup. Currently, you can install OpenMixup with quick installation steps, download a small-scale dataset (e.g., CIFAR-100), and start training AutoMix.

Taking AutoMix on CIFAR-100 as an example, we provide installation and training scripts as follows:

# install openmixup
conda create -n openmixup python=3.8 pytorch=1.10 cudatoolkit=11.3 torchvision -c pytorch -y
conda activate openmixup
pip3 install openmim
mim install mmcv-full
git clone https://github.com/Westlake-AI/openmixup.git
cd openmixup
python setup.py develop
# download CIFAR-100 dataset and put it under './data/cifar100'.
mkdir data
python -c "import torchvision; torchvision.datasets.CIFAR100(download=True, root='data/cifar100');"
# start training
PORT=29001 bash tools/dist_train.sh configs/classification/cifar100/automix/basic/r18_l2_a2_near_L1_01_mlr5e_2.py 1

During training, some mixed samples will be saved in the experiment directory. We hope this would help you!

Siyuan Li