training pvanet in multiple gpu?

[hengck23]

Is it possible to train pvanet in a single machine with multiple gpus?

According to "rbgirshick#107", python layer makes it not possible to use multiple gpu.

However, this paper "Feature Pyramid Networks for Object Detection" (https://arxiv.org/abs/1612.03144) suggests that RPN framework can be trained using synchronized SGD training on 8 GPUs (using caffe2)

I am currently woking on COCO dataset for object detection with pvanet. Any hint to speed training using multiple GPU machine? Or will training on GPU cluster helps?

[songjmcn]

I think you will rewrite python layer using C++ and cuda.The python can not use multiple GPU.

[hengck23]

mxnet faster rcnn (and also mxnet ssd, mxnet yolo) can support multi-gpu in single machine and multiple gpu machines. see: ijkguo/mx-rcnn#42

there is converter to convert mxnet model to caffe model and vice verus.