Add TF weights
Model converted by the transformers' pt_to_tf CLI -- all converted model outputs and hidden layers were validated against its Pytorch counterpart. Maximum crossload output difference=3.433e-05; Maximum converted output difference=3.433e-05.
Hi Joao, thanks for adding the tf model!
That was a fast reply 😂
For context: I'm a TF maintainer at Hugging Face, and this is your most downloaded model whose weights can be automatically converted into TensorFlow, using our tools. We believe that having TF weights would be of interest to the community, and will further boost the visibility of the model.
I also don't want to be a source of spam! Let me know if you would like me to open PRs with TF weights for other models that you own. Alternatively, if you'd like to have the TF weights but no hub notifications, I can also push the weights using admin privileges 🤗
I am hanging out all day on hugggingface.co 😂 I don't mind if you push the TF models directly, since I would merge them anyway 🤗.
Can you tell me, how the process works if I need to update the PyTorch model? Do I need to delete the "old" TF model in this case?
Can you tell me, how the process works if I need to update the PyTorch model? Do I need to delete the "old" TF model in this case?
That's a great question. I actually have three answers for it -- yes, the old weights must be updated as well:
- Our long-term goal is to automate everything -- each user/org would have some flag in their options to allow a bot on our end to push updated weights (or open a PR) whenever the original weights are updated. Then everyone would have the right weights for all frameworks all the time, or would get warnings if there are conversion issues 🤗
- Our short-term goal is to enable users/orgs to be self-sufficient in these matters, i.e. release the
pt-to-tfCLI to manually update weights and open PRs. However, Hub PRs are quite new, and the functionality to programmatically open PRs is still on a branch, so the CLI is cumbersome to use unless you checkout to a special branch. - As of now, you can ping me whenever you'd like a conversion. I'm going through Hub models by reverse number of downloads, in the hopes to help as many TF users as possible!
Thanks for the feedback. Honestly, I never had TF on my radar since I am working with PT and I don't know how many people are actively using TF.
So the " just works out of the box" option 1 sounds great :) Thanks for your work on this 🤗