caffe prediction speed in GTX980

[Zahoor Mohamed J]

Hi

I am trying to bulk predict images (a million ) using bvlc_alexnet model using a single GTX 980 card. 

The python interface seem to be pretty slow even when i enable gpu_mode (1 image per sec).

So i switched to c++ version using MemoryData layer with a batch size of 256... 

I am able to classify about 5 images per sec with this ...

My Question is:

1) What classification speed can i expect for 256x256 images in GTX 980 with a 256 batch size? 

     Caffe paper claims ~2.5ms per image using k40... can i expect similar speed ?

2) Can i expect the similar speed in python interface also with Oversample = False? 

./Zahoor

[Andriy Lysak]

The speed for your python tests seems very slow!

I am running a slightly modified bvlc_alexnet model (nothing that should speed up or slow down the classifications drastically) on a GTX980 on an Ubuntu machine and i am getting ~5ms per image.

If your speeds are that much slower than something is wrong.

if you post a snippet of your python code maybe someone can help.

Best Regards,
Andriy

[Zahoor Mohamed J]

Thanks for the reply Andriy.. i went back and looked in to the python code again..

i had switched ON the Oversample which was fiddling with the speed..i made it false now...and now i am getting somewhere 40 images per sec...

how do i increase the batch size to speed it up more...? 

./zahoor