onnx is nice in principle, but pretty limited. core ops like non-max-suppression can't get properly converted. also model deployment is not great, memory consum

onnx is nice in principle, but pretty limited. core ops like non-max-suppression can't get properly converted. also model deployment is not great, memory consumption and control thereof worse than with tensorflow.