Created by: ghost

Hi. I managed to run mask_rcnn_resnet50_1 with both fp32 and fp16. I run it on Nvidia V100 card, 16GB version. I'm running on a single card.

I get around 4 images per second using fp32, and when I switch to fp16 (by adding --fp16) the performance does not increase, it's still around 4 images per second.

Is this expected?

I do everything following your readmes, so using coco dataset you linked, your pretrained rn50 weights and so on...

For the sake of completeness, here is the command I used for fp32:

python3 train.py -c mask_rcnn_r50_1x.yml -o max_iters=200 use_gpu=True

and here is the version for fp16 (I changed affinity to bn in the config)

python3 train.py -c mask_rcnn_r50_1x.yml -o max_iters=200 use_gpu=True --fp16