search

chinakook / hr101_mxnet

A MXNet tiny face detector
MIT License
95 stars 26 forks source link

Nice work! #1

Closed ujjaldas1997 closed 7 years ago

ujjaldas1997 commented 7 years ago

Can you give the details of gpu, and running time? I had run both mxnet and matconvnet version and both were taking huge amount of time(about 250 sec for selfie.jpg). Also will you release the training code of the same?

chinakook commented 7 years ago

@ujjaldas1997 Both implementations are less than 3 sec on my GTX 1070(with CUDNN 5.1). No training code in mxnet until now.

ujjaldas1997 commented 7 years ago

Ahh, in my case it is -- 01:00.0 3D controller: NVIDIA Corporation GM108M [GeForce 940M] (rev a2) I think performance issue is because of this. Are you working on training mxnet for hr_res_101? Thank you.

chinakook commented 7 years ago

@ujjaldas1997 I think that training is trivial because the tiny face is virtually a RPN of Faster RCNN(just with a FCN skip connection added). I have no plan to write training scripts in mxnet recently, but I think you can refer to a RPN of FRCNN training.

ujjaldas1997 commented 7 years ago

Yes and thanks. I have almost finished training code(using keras library the traditional way), and thinking of switching to mxnet as it is faster. And thanks again for the direction.