ydwen / opensphere

A hyperspherical face recognition library based on PyTorch
https://opensphere.world/
MIT License
274 stars 29 forks source link

Understanding Hyperparamters for finetuning #20

Open akirs82 opened 1 year ago

akirs82 commented 1 year ago

Hello ydwen,

Can you brief us once on the training hyperparameters like s,m,lw. For my custom dataset that I have collected in my university, ACC is around 59 to 60 and the validation accuracies are reducing in every iterations. The Mag_std,Mag_mean values are so high. Sample logs as follows for different learning rates,

train.log train.log train.log train.log train.log train.log

My custom dataset contains surveillance images with avergae widthxheight approx as 30x30. I am using spherefacer as below,

  head:
    net:
      type: SphereFaceR_H
      s: 60
      m: 1.4
      magn_type: v2
      lw: 50.

It seems little difficult to train these - any help will be greatly helpful and appreciated.

regards akirs