FPS in 1080ti just 30-50fps for single_class detection

[xxy90]

[RangiLyu]

Make sure you turn on cudnn acceleration.
Pull the latest version of nanodet or change

    torch.backends.cudnn.deterministic = True
    torch.backends.cudnn.benchmark = False

in demo.py to

    torch.backends.cudnn.enabled = True
    torch.backends.cudnn.benchmark = True

By the way, pytorch is a training framework, not designed for inference. If you want to get better inference speed on GPU, try to use C++ demo. C++ demo using ncnn vulkan as backend.