search

Tianxiaomo / pytorch-YOLOv4

PyTorch ,ONNX and TensorRT implementation of YOLOv4
Apache License 2.0
4.47k stars 1.49k forks source link

tensorrt conversion problem #117

Open ou525 opened 4 years ago

ou525 commented 4 years ago

@ersheng-ai Thank you very much for your work，does tensorrt already support the mish function, do you need to write plug-in support? Is it convenient to provide a demo of c++ tensorrt?

ersheng-ai commented 4 years ago

ONNX standard is now being main focus of TensorRT parsers. You can use *`x Tanh( Softplus( x ) )** instead of **Mish( x )`**.

If you want to convert a PyTorch model containing mish function into ONNX, torch.onnx.export() will transform Mish( x ) into *`x Tanh( Softplus( x ) )`** automatically.

C++ TensorRT for YoloV4 such as deepstream pipeline is still under development but the demo may be available a few weeks later.