Is Torch faster than Caffe?

[nitish11]

Hi,

I am using the gender detection model in Torch and in Caffe for detection from live camera.

Running the code on CPU and keeping the same models file, I am getting different prediction times. For Caffe, it is ~1.30 seconds per frame. For Torch, it is ~0.45 seconds per frame.

What could be the possible reason for the time difference? Is Torch faster than Caffe?

[admercs]

LuaJIT is faster than Python and Torch is generally faster than Caffe from what I've seen in benchmarks, which are a bit outdated: https://github.com/soumith/convnet-benchmarks. I hear Nvidia gives the best support to Torch, which they use for much of their own work (e.g., autonomous car demonstrations), as others like Google and Nervana/Intel compete on hardware. cuDNN 5 speeds up Torch quite a bit: https://devblogs.nvidia.com/parallelforall/optimizing-recurrent-neural-networks-cudnn-5/.

[nitish11]

@adam-erickson : :+1: thanks.

But, the huge difference in only one frame is the concern. The benchmark has not compared LuaJIT with Caffe.

[GilLevi]

Hi @nitish11 ,

Thanks for your interest in our work and thanks for this cool repository: https://github.com/nitish11/GenderRecognition

It's embarrassing to admit it, but I never worked with Torch and I really can't say which one is faster (or in this case - why you get faster run-times in Torch).

Best, Gil.

[nitish11]

I checked torch and Caffe computation engine called BLAS.

In Ubuntu 14.04,

ldd /home/nitish/caffe/build/lib/libcaffe.so
ldd /home/nitish/torch/install/lib/libTH.so

From the output, I observed that torch is linked against openblas, and caffe is linked against libcblas, which might be the reason for slower Caffe.

Solution : Build Caffe with OpenBlas

[admercs]

Are you sure it's not simply the difference in looping speed between LuaJIT and Python? It can be quite large. Similar to Julia, LuaJIT is closer to C.

[nitish11]

I am not sure about looping speed between LuaJIT and Python.

[MohanaRC]

There is a better way to call torch from python code using wrapper. Please check