search

rbgirshick / py-faster-rcnn

Faster R-CNN (Python implementation) -- see https://github.com/ShaoqingRen/faster_rcnn for the official MATLAB version
Other
8.1k stars 4.11k forks source link

Training time for ZF #192

Open lynetcha opened 8 years ago

lynetcha commented 8 years ago

How long can I expect faster_rcnn to take to train ZF on VOC2007? I am getting about 4.5hours on a TitanX, which seems pretty long. (The training rate is 0.242s / iter)

happyharrycn commented 8 years ago

If you are using end2end training, have a good SSD and compiled Caffe with latest cuDNN, training should be much faster than 200 ms per iteration.

Austriker commented 8 years ago

@happyharrycn Which version of caffe do you use ? I tried to build the caffe fork of this repo with cuda 7.5 and cuDNN v5 and it didn't work.

happyharrycn commented 8 years ago

I am using the caffe main branch (cuDNN v5 is supported on May 17th). You will need to merge the RoI pooling and the smoothed l1 loss layers from Ross's caffe version.

Austriker commented 8 years ago

@happyharrycn I have done a PR 4163 that merge the both layers to caffe and I am working on porting the code to python 3.4 py-faster-rcnn fork. I still have an issue with the test in my PR. I can't manage to solve it yet !

Can I have the link to Ross's caffe version please ?

happyharrycn commented 8 years ago

@Austriker I have commented on your PR 4163. It should compile and work now.

Austriker commented 8 years ago

@happyharrycn Great thank you ! I will test the python 3 port ! It will avoid a lot of maintenance problem having only one caffe !