-
Notifications
You must be signed in to change notification settings - Fork 508
How to evaluate TAR FAR on your dataset
(1)先清洗数据How-to-clean-your-dataset
(2)创建一个批处理文件比如evaluate-tar-far.bat 用记事本/写字板/Notepad++打开,填入以下命令
set proto_file=model\mobilefacenet-res4-8-16-4-dim128.zqparams
set model_file=model\mobilefacenet-res4-8-16-4-dim128-emore.nchwbin
set blob_name=fc5
set src_root=webface1000X50
set thread_num=10
SampleFaceDatabaseZQCNN.exe make_compact %src_root% tar-far-database.feats tar-far-database.names %proto_file% %model_file% %blob_name% 2 %thread_num%
SampleFaceDatabaseZQCNN.exe compute_similarity_compact tar-far-score.raw tar-far-flag.raw tar-far-info.txt tar-far-database.feats tar-far-database.names %thread_num%
SampleFaceDatabaseZQCNN.exe evaluate_tar_far tar-far-score.raw tar-far-flag.raw tar-far-info.txt
最终tar-far数据会在控制台显示。你也可以在cmd里运行
call evaluate-tar-far.bat>log.txt
这样运行过程及最终结果都会写在log.txt里
注意1:
如果你想测caffe格式的模型,把SampleFaceDatabaseZQCNN.exe
换成SampleFaceDatabaseOpenCV.exe
, 参数make_compact
换成make112X112_compact
或者make112X96_compact
注意2: 如果你要测的库有5万张图,需要50000 x 49999 / 2 x (4+1) x 3 Bytes = 17.46 GB的硬盘空间(还得考虑硬盘空间是不是连续); 如果你要测的库有10万张图,则需要约70 GB的硬盘空间(还得考虑硬盘空间是不是连续);如果你要测的库有20万张图,则需要约280 GB的硬盘空间(还得考虑硬盘空间是不是连续)。