Feature request: easy to use pre-trained models

[oliverdain]

In order to use these models you have to clone a few git repos, manually install a bunch of stuff, use a specific, old version of Python, etc. The only instructions for use are calling various command line scripts. It'd be really nice if there was a smaller package that let you (1) instantiate a model instance and (2) call load_state_dict. From there anyone familiar with PyTorch knows how to use it in their code. Ideally this library would have minimal extra dependencies, especially those that can't be simply pip-install'd.

A torch.jit export of a complete model including architecture and weights might be really nice too.

Currently the easiest way to get a model that could be used with scripts, mmcv, etc. is to set everthing up, then export to onnx or torchseve and go from there. But the ONNX export I tried failed due to https://github.com/ViTAE-Transformer/ViTPose/issues/73 and the torchserve route fails with errors as well.