[RFC] Support exporting to torchscript

[zhiqwang]

Describe the feature

Add support for exporting the models released on mmlab to torchscript.

Motivation

1. The training framework currently used by mmlab is mainly PyTorch, so it would be relatively easy to export the model to torchscript.

2. If we can export the models to torchscript, it will be easy to deploy it with libtorch, we can also use more tools released by the PyTorch teams, such as torchserve.

3. Seems that torchscript is attracting more and more attention in industrial production environment. For example, we can find that some edge inference libraries such as coreML, ncnn, TVM, TensorRT and MNN have added or are adding support for exporting torchscript to their own IR models.

Related resources

• PNNX is an open standard for PyTorch model interoperability Tencent/ncnn#3262
• https://github.com/alibaba/MNN/releases/tag/1.2.0
• http://github.com/NVIDIA/Torch-TensorRT
• https://github.com/apache/tvm/releases/tag/v0.8.0
• https://coremltools.readme.io/docs/pytorch-conversion
• https://github.com/pytorch/serve

[grimoire]

Thanks for the advice!
We are doing some technical verification about torchscript. And I believe it will be supported in future once we find a good way to integrate it in OpenMMLab.

[zhiqwang]

Close due to open-mmlab/mmdeploy#159 , Thanks for this great enhancement!