I’m curious why we move stuffs out of CUDA here.
I’m trying to speed up object detection’s mAP calculation using on_epoch_end
callback. This requires calling non-maximum suppression which has CUDA-optimized implementations.
I’m curious why we move stuffs out of CUDA here.
I’m trying to speed up object detection’s mAP calculation using on_epoch_end
callback. This requires calling non-maximum suppression which has CUDA-optimized implementations.
grab_idx
is used by show_results
which doesn’t care about the GPU. You shouldn’t be using it inside a metric that you want computed on the GPU.