Torchscript training from Detectron2

[mdhaisne]

Hello,
I was wondering why it is not possible today to train a model from detectron2 after an export in scripting / tracing mode ?
In this page there is this assert:

assert (
        not model.training
    ), "Currently we only support exporting models in evaluation mode to torchscript"

And on this page about Tracing Adapter, the forward method is using with torch.no_grad(). Maybe it is possible to replace this forward method to enable training ?

Thanks

[zhanghang1989]

torchscript model is used for inference only.