2023-AICOSS

2023 AICOSS hackathon competition with multi-GPU training

For single-GPU training and more information about the competition, you can use this repository.

Training

Virtual Environment Settings

You should install Anaconda for the following codes.

git clone https://github.com/junpark-ai/AICOSS
cd AICOSS/
conda env create -f ddp.yaml

conda activate ddp

You can change the name of your environment by using the -n option such as conda env create -f ddp.yaml -n 'your_env_name'.

Training example

python main.py \
	--epochs=10 \
	--lr=1e-4 \
	--min_lr=1e-7 \
	--model_name=cvt384_q2l \
	--loss_name=PartialSelectiveLoss \
	--img_size=384 \
	--batch_size=8 \
	--grad_accumulation=8 \
	--path=/data

You should change --path to your data path.

If you have less/more GPUs than 4, you can change --gpu=0,1,2,3 to whatever you want.

And batch_size means for each GPU device, so if you set --batch_size=8 --gpu=0,1 your total batch size will be 16.

We used 4 x Nvidia RTX 3090 GPUs for training.

Reference

• ML_Decoder
• TResNet
• timm
• query2label
• Pytorch Lightning

Papers

• RandAugment: Practical automated data augmentation with a reduced search space [NeurIPS 2020]
• TResNet: High Performance GPU-Dedicated Architecture [WACV 2021]
• Multi-label Classification with Partial Annotations using Class-aware Selective Loss [CVPR 2022]
• ML-Decoder: Scalable and Versatile Classification Head [WACV 2023]