can the decode part be implemented inside the tensorrt model?

[Libaishun]

it would be nice if the decode part is treated as part of the tensorrt model rather than postprocess outside, i tried to embed it, for that i use torch.nonzero() which unfortunately is a unsupported op by tensorrt

[CaoWGG]

The post-processing part can be develop as a custom op of onnx like dcnv2.