how to do conv-forward one CNN network on multiple GPU (matcaffe - caffe master)

[himora...@gmail.com]

Dear all,

I'm using matcaffe (caffe master) to extract feature of one CNN network

Its able to train CNN on multiple GPU using ./builld/tools/caffe train -gpu 0,1 ....

However, it seem that the matcaffe cannot do conv-forward a CNN on multiple GPU

I have use the matcaffe command: 

  caffe.set_mode_gpu()

  caffe.set_device(0)

  caffe.set_device(1)

  net = caffe.Net(model,weights, 'test')

  net.forward(input_data)

However, the system seem to run the network testing on 1 GPU. 

Hope to get your support on this problem.

Best regards 

[Alfredo Nava]

I have this issue as well and have not gotten any help on getting the net.forward command to process all data with multiple GPUs.

Seems as if the closest way to do this is to run your script in multiprocess and set each process to each GPU, and that will process your data in parallel if you do it in python. Not sure if you can do this in mattcaffe.

[Vu Hong Thuan]

As I know that caffe now supports multi-gpu only for training

https://github.com/BVLC/caffe/blob/master/docs/multigpu.md