search

TreB1eN / InsightFace_Pytorch

Pytorch0.4.1 codes for InsightFace
MIT License
1.72k stars 418 forks source link

About the accuracy tested in LFW AgeDB-30 Cfp-fp #142

Open haolinlyou opened 4 years ago

haolinlyou commented 4 years ago

I have trained the repo with ms1m. After 30 epochs, the accuracy is 0.993、0.933、0.934，lower the pretrain model. My hyper-parameters is: LR 0.001 batch_size 384 (96*4) IR-se50 Can you tell me what can i do for imporving accuracy

ruiming46zrm commented 4 years ago

set lr = 0.1 at the begining.

haolinlyou commented 4 years ago

set lr = 0.1 at the begining.

OK Thank you for your reply. By the way, i foud the loss is nearly 0.004 when training in epoch 30.

1icas commented 4 years ago

I have trained the repo with ms1m. After 30 epochs, the accuracy is 0.993、0.933、0.934，lower the pretrain model. My hyper-parameters is: LR 0.001 batch_size 384 (96*4) IR-se50 Can you tell me what can i do for imporving accuracy

你好，请问可以看一下您的loss下降曲线吗

xszg007 commented 3 years ago

@lk123400 May I ask how to verify the acc on those three datasets? I trained my model but there was no acc output during training. I also could't find any usable verify program in this prj.

Fanzhenxiong commented 10 months ago

我已经用 ms1m 训练了回购。经过30个周期后，精度为0.993、0.933、0.934，降低预训练模型。我的超参数是： LR 0.001 batch_size 384 （96*4） IR-se50 你能告诉我我能做些什么来提高准确性吗 请问怎么测试lfw数据集和cpf数据集啊？运行哪个文件啊？