How to run a pretrained model on CPU-only machine

[palmforest]

Hi,
Does anyone know how to run the pre-trained imagenet model on a machine does not have a cuda-supported GPU? I only need to do the testing phase, eg. to use wrapper,py for prediction.

It seems that in order to compile pycaffe, I need to install Cuda 5.5 or 5.0 to compile the whole caffe package.

Thanks

[tstibor]

Yes, the CUDA package is required. Install CUDA (during installation say no driver and no examples)
and in prototxt file solver_mode: 0. I am running caffe on a latop with ATI card and AMD CPU and it works.

[kloudkl]

#152 and #172 make efforts to resolve #3. You can run in CPU mode as demonstrated here without a GPU. But you still need to install CUDA to build the executable.

[shelhamer]

Right, Caffe works fine without a gpu as long as one still installs the CUDA libraries (without the driver or samples). A CPU-only build without the CUDa dependency is in progress.

[uncommoncode]

CUDA libraries or packages _are not required_ to install at this time with a CPU-only build. In Makefile.config uncomment the line CPU_ONLY := 1. This is documented in http://caffe.berkeleyvision.org/installation.html.

I personally do not have CUDA installed and am successfully able to run in CPU mode and use pycaffe. Note that you will need to change solver_mode: CPU for several examples. Please be aware you will likely be at least ~10x slower than a GPU for training!