What is the main modifications of py-R-FCN-multiGPU compared to its precedent? about py-r-fcn-multigpu

[bharatsingh430]

You can look into the train_multi_gpu.py file in the fast_rcnn folder for the multi gpu part. I use the latest caffe branch which includes NCCL and then replace the batch norm layer in it with the one in py-r-fcn. ResNet would otherwise take 70-80% more memory if you use batch norm in the main repo. Other missing layers are also added, like ohem etc. So that each network on a different gpu operates on different set of images, random seed is set based on the GPU_ID in roi_data_layer/layer.py.

[crazylyf]

@bharatsingh430 Very nice of you, thank you!

[zimenglan-sysu-512]

@bharatsingh430 how do you modify the python code/layers to support the multi-gpu training and testing?