Understanding Hyperparamters for finetuning

[akirs82]

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