About a Arcface training strategy

Question

I'm training Arcface with CASIA-Webface dataset with refering This github page(https://github.com/ZhaoJ9014/face.evoLVe.PyTorch).

I've aligned the CASIA-Webface dataset using the face alignment codes from that github page and fixed them to 128x128(So I changed some layers in ResNet50). And next I began the training Arcface with 256 batch size during few days. Now, the current training has exceeded more than 160 epochs. However, the loss doesn't seem to be lower than 1.0 and the accuracy in the LFW datset is still under 99%. The example code from that github page is considered to use MS-Celeb-1M for training.

• (1) Has anybody tried training ArcFace with CASIA-Webface and evaluate them on LFW Dataset? In this case, how much accuracy does it record?

• (2) I'm using 256 batch size and current epoch is 168. Do we need more training? If so, how much epochs and batch size are required for Arcface to perform sufficient accuracy?

• (3) I modified the image size to 128x128 and ResNet50's fully connected layer. Does it affect to Arcface accuracy? (The github page uses 112x112)

My status of training Arcface on CASIA (Validation is performed on LFW only now)