How to see the speed up on GPU? about caffe HOT 5 OPEN

Random sparse neural networks with crs sparse computation are slow. Please use structured sparsity and conv_mode: LOWERED_CCNMM in our deploy protobuf. More details in tutorials.

from caffe.

Thanks for you reply. The model in this case is the "caffenet_SSL_0.4469.caffemodel", which should be trained with structured sparsity.

I can see cpu speed up (by setting caffe.set_mode_cpu()) using "conv_mode: LOWERED_CCNMM". But I couldn't see speed up on GPU (by setting caffe.set_mode_gpu()) when applying "conv_mode: LOWERED_CSRMM".

Could you help me see the speed up on GPU? Thanks in advance.

from caffe.

@wenwei202 @liaocs2008
I would like to know whether the command ‘’./build/tools/caffe time -model=xx/xx.protoxtx -weights=xx/xx.caffemodel‘’ can get test time and compute speeded-up.

from caffe.

@jiaqun123 Using "caffe time" command measures the end to end time of a layer. But I think in this implementation, @wenwei202 measures the time only for matrix multiplication. You can check the python script "examples/caffenet_classifier.py" to measure speedup.

from caffe.

@liaocs2008
Thanks for your reply。I tried to use examples/cifar10_classifier.py
python examples/cifar10_classifier.py models/cifar10/cifar10_full.prototxt cifar10_full_iter_200000.caffemodel. But I get the the following result and didn't get the time for each conv layer.

from caffe.