search

bertinetto / cfnet

[CVPR'17] Training a Correlation Filter end-to-end allows lightweight networks of 2 layers (600 kB) to high performance at fast speed..
MIT License
523 stars 155 forks source link

evaluation is slow #10

Closed gongbudaizhe closed 7 years ago

gongbudaizhe commented 7 years ago

Hi,

good work as always : )

The speed of CFNet-conv2 reported in the paper is 75 fps. But when I evaluate this model in OTB-2013, it ran less than 3 fps. I am using NVIDIA GTX 1080, cuda 8.0, cudnn 5.1. Matconvnet is compiled with GPU.

Is there anything I might be missing?

bertinetto commented 7 years ago

I think you might have the visualization on? That speed has been reported (clearly) without visualization, on a Titan-X and 4GHz machine.

Otherwise maybe there is a problem with your matconvnet.

gongbudaizhe commented 7 years ago

The visualization flag is off. Maybe this is not the problem.

Matconvnet is installed following the instructions in the official documentation. No errors during installation.

I evaluated the tracker using the python version of OTB evaluation toolkit though. It uses Matlab engine to invoke code written in Matlab. Maybe this causes the slow down?

bertinetto commented 7 years ago

Yep that might be the case.

gongbudaizhe commented 7 years ago

It turns out that Matlab 2016a doesn't support GTX 1080 and recompiles GPU libraries every time a new session is started. More discussions can be found in this thread

foolwood commented 6 years ago

@gongbudaizhe Solution is found here. https://github.com/apache/incubator-mxnet/issues/3239

export CUDA_CACHE_MAXSIZE=2147483647 export CUDA_CACHE_DISABLE=0

gongbudaizhe commented 6 years ago

@foolwood thanks