The OCR approach is rephrased as Segmentation Transformer: https://arxiv.org/abs/1909.11065. This is an official implementation of semantic segmentation for HRNet. https://arxiv.org/abs/1908.07919
the original conf_mat calculation method costs too much when transfer data back to cpu.
And this makes v100 falls into some kind of stall, when GPU Mem is used more than 20GB.
I added the the conf_mat calculation method by torch api using GPU. This imporves the efficiency in each testing or validating.
Hope you can merge this into your official repo.
the original conf_mat calculation method costs too much when transfer data back to cpu. And this makes v100 falls into some kind of stall, when GPU Mem is used more than 20GB. I added the the conf_mat calculation method by torch api using GPU. This imporves the efficiency in each testing or validating. Hope you can merge this into your official repo.