Equiangular-Basis-Vectors

Paper is available at: http://arxiv.org/abs/2303.11637

EBVs pretrain on ImageNet-1K can be download at: https://github.com/aassxun/Equiangular-Basis-Vectors/releases/download/untagged-72ef569681e789dd1bb8/EBV_ResNet_dim1000_SGD_epoch605.pth

Data for the 10w classes classification task can be download from: https://github.com/aassxun/Equiangular-Basis-Vectors/releases/download/10w_classes/10w_classes.zip

It contains 120w images for training and 60w images for validation. (Resize to 224*224)

Environment

Python 3.8.11
Pytorch 1.11.0
torchvision 0.12.0
numpy 1.22.4
timm 0.6.11

Generate EBVs

Please refer to /Generate_EBV/Generate_EBV.py

Experiments on ImageNet

Training logs and code can be found at: /ImageNet_Validation_Experiment/ and /ImageNet_Ablation_Study/Log/

The random seed for all the experiments is 42.

Experiments on COCO

All the code and training logs can be found at: /coco_detection/work_dirs/ebv_det/

The random seeds for each experiment are exactly the same as the corresponding experimental settings in MMdetection.

Experiments on ADE20K

All the code and training logs can be found at: /ADE20k_segmentation/work_dirs/ebv_seg/

The random seeds for each experiment are exactly the same as the corresponding experimental settings in MMSegmentation.