BoB-Classification

This repository is the official implementation of Image Classification task in the Battle of the Backbones: A Large-Scale Comparison of Pretrained Models across Computer Vision Tasks.

Dependencies

Version Control of Python libraries in environment.yml file. To create a virtual environment:

conda env create -f environment.yml

Running instructions

Following instructions introduce how to evaluate pretrained backbones with different classficaition tasks. We provide various datasets and pre-trained models.

Finetuning (convnext_xl pretrained with vicregl)

python3 -m torch.distributed.launch --nproc_per_node=8 train.py /path/to/ImageNet/ILSVRC2012 --dataset ImageNet --config ./model_configs/convnext_xl_vicreg_ft.yaml --lr 1e-3 --experiment convnext_xl_vicreg_ft_lr1e-3

1% and 10% of ImageNet training (stable diffusion pretrained model)

python3 -m torch.distributed.launch --nproc_per_node=8 train_semi.py /path/to/ImageNet/ILSVRC2012 --dataset semiimagenet --n_shot 10 --config ./model_configs/vit_small_patch16_224_dino_semi.yaml -b 16 --experiment semi_sd_adamw_lr2.5e-4_wd5e-2_ld65_shot10_da --weight-decay 0.05 --lr 2.5e-4 --drop-path 0.1 --epochs 60 --model stable_diffusion_v1 --layer-decay 0.65 --aa rand-m9-mstd0.5-inc1 --accum_iter 4

Linear Probing (CLIP trained resnet50)

python3 -m torch.distributed.launch --nproc_per_node=4 linear_probe.py --data_path /fs/cml-datasets/ImageNet/ILSVRC2012/ --batch_size 512 --epochs 90 --blr 0.1 --weight_decay 0.0 --dist_eval --model resnet50_clip

This code is based on the implementations of MAE, TIMM