What is the most effective way to migrate an existing model to a newer version of Pytorch?

I have some models currently in production, using torch’s C++ API and traced torchscripts, ala these instructions: https://pytorch.org/docs/master/generated/torch.jit.trace.html
I also have the model’s state dicts saved away.
Overall this solution works great.

The models and torchscripts were generated using pytorch 1.3
I want to update the next version of the system to pytorch 1.5

Retraining is possible, but not preferred, and certainly not a long term solution.
Is there a procedure for updating my existing models, from either the saved torchscripts or the state dicts?


The easiest way is to instantiate the model in Python, load the state dict and trace the model again.
This is something to know, even though you can run the traced model without the source, you’ll regret not keeping it around at some point.

Best regards