I opened a discussion last year: Packaging pytorch topology first and checkpoints later that didn’t get much traction. It was suggested that I repost it here.
The details are in the previous discussion, so I’ll keep this short. But the gist is I want to:
- Deploy my model as a torch package as normal
- Update the weights in that file (perhaps by unzipping it, adding in a new file with the pickled weight dictionary, and then re-zipping it).
After using torch.package for awhile I’m thinking this isn’t possible given the current design of it, but if that is true I’d like to understand why, and I’d also like to see if this is a feature for torch.package where there might be some interest in developing it.
As mentioned in my previous post, I wrote a torch packager (before torch.package existed) that did something similar to what I’m saying, so I know it’s possible.