huggingface_hub
huggingface_hub copied to clipboard
Guide HF transformer users to use the corresponding hub functions
Is your feature request related to a problem? Please describe.
We have push_to_hub_keras
to push Keras models. However, HF transformer architectures have a more comprehensive model hub push in save_pretrained
. It would be nice to nudge users there if they try to use push_to_hub_keras
with an HF transformer.
Describe the solution you'd like A loud warning nudging users to the right place :D
Describe alternatives you've considered An exception, but might be too disruptive.
Additional context As discussed here. I don't know if there is a similar problem for PyTorch -- if there is, this issue would also be applicable there.
cc @merveenoyan
I think most transformers TF users will just be using existing tools in transformers, I don't know if anyone has tried push_to_hub_keras
with :hugs: transformers
The transformer models all have the push_to_hub
method which pushes everything required to the hub; it's shared by models, tokenizers, configurations, feature extractors and processors. If I understand correctly, you'd like to push for users to use push_to_hub_keras
instead of that, but I think it would break the API shared across all frameworks in transformers
.
@LysandreJik no, the other way around, my concern is that users might attempt use push_to_hub_keras
with transformer models, when there are better methods (as you mentioned).
@osanseviero I haven't seen related issues, so maybe I'm simply overthinking :D
This conversation started because push_to_hub_keras
has a config
argument, but it is not an HF model config -- it is a Keras model config. The mixed terminology may cause confusion for users:
-
push_to_hub_keras
pushes to hub Keras models, but it is not meant to be used with Keras HF transformers models; -
push_to_hub_keras
has aconfig
argument, but it has nothing to do with a standard config object in HF transformers models.
I think this might be a bit of an overthinking; push_to_hub_keras
is not as widely know as transformers
push_to_hub
, and everywhere in transformers
docs we show push_to_hub
, so I don't think people would try using push_to_hub_keras
for Transformers models.
EDIT: I do see how the two different use of config can cause confusions, although we also do that with the PT mixin I think.
@osanseviero I agree. To my observation, people do model.save()
and then when they try to load they are asked for config or tokenizer and fail in loading, no one uses push_to_hub_keras()
.
I'd say it's also related to https://github.com/huggingface/huggingface_hub/issues/824
(closing this issue as the API evolved over time and push_to_hub_keras
still has very limited usage. Anyone landing on this issue, feel free to comment and we can reopen :))