Closed dirkgr closed 2 years ago
TransformerQA and Naqanet don't converge when training with the instructions in the demo :-(
@osanseviero, why did you write the push-to-hf
command so that it renames the weights file to model.th
? We expect best.th
or better weights.th
. Like this we don't load properly when I try to use hf://
URL.
As discussed offline indeed this is an issue in the push-to-hf command. Fixing this is a minor change in the pushing script.
@osanseviero, I just uploaded a second version, and it did this: https://huggingface.co/allenai/bidaf/blob/main/README.md
That's probably not intended?
Hi @dirkgr. The implementation of the push command respects the existing model card to avoid deleting user-written model card data or metadata, so it just adds the template before. We're working in some wrapper for the metadata that will include validation and allow merging metadata, but this is still not done. If you prefer, we can modify the command such that it does not add anything to the model card when there is already a model card, but moving forward we'll probably want to add things such as metrics to the metadata.
@dirkgr this is just a friendly ping to make sure you haven't forgotten about this issue 😜
@osanseviero, the tensor board stuff is there, but it doesn't seem to render?
Hi @dirkgr, I actually see them in the training metrics tab
Ah, works in Chrome, not in Safari.
@severo from the team is looking into it! :hugs:
It seems to be an error with the last version of tensorboard. I reported it here: https://github.com/tensorflow/tensorboard/issues/5489.
Fixed!
Thanks for reporting
Thanks for fixing!