PythonLayer is not implemented in Multi-GPU training

[RSly]

Hi
I am using the latest nv-caffe and get the error
PythonLayer is not implemented in Multi-GPU training.

It seems that this issue has been fixed in bvlc caffe in
BVLC#4563
and
BVLC#2936

is there a special reason this change has not been merged in nv-caffe?

I think this is a great improvement to caffe and can save a huge amount of training time with python loss and layers.

Ref: NVIDIA/DIGITS#1570

[drnikolaev]

@RSly it's here: https://github.com/NVIDIA/caffe/tree/caffe-0.16

[RSly]

@drnikolaev thanks a lot! I will test it soon

Does it work out of the box or will I need a caffe.py as in
https://github.com/BVLC/caffe/blob/master/python/train.py
to train on multi-gpus with python data layer?

[RSly]

Hi @drnikolaev
I tested this and it all works like a charm! bravo!
Also no need for caffe.py

[drnikolaev]

Cool, thanks for confirming!

[LinHungShi]

@RSly @drnikolaev How do you enable training with multiple gpu in Python? I can't find any tutorial for multi-gpu training in Python, and the only thing that I found relevent in _caffe.cpp is set_devices which seems not useful for mutl-gpu training. I know that caffe.cpp has support for python layers, but I'd like to see if there any way I can do it in Python rather than C++. Thanks