Add Huggingface Integration #916
Conversation
|
MLCommons CLA bot All contributors have signed the MLCommons CLA ✍️ ✅ |
There was a problem hiding this comment.
Before I start reviewing this in earnest, I would need at least the following 2 pieces of information to be added to the PR:
- Documentation: (it is absolutely fine to have a bullet point list of items that link to the main HF docs)
- Tests
I believe both of these were present in the previous PR.
|
Hi @Wauplin and @NielsRogge - this PR looks good from my end. Do you have any feedback? |
There was a problem hiding this comment.
Hey there 👋 Thanks for the ping!
I left a few comments from an outsider point of view. I do think that the CLI should be more opinionated (understand "have less options and decide things for the user") otherwise we pretty much end up with a CLI close to what huggingface-cli upload and huggingface-cli download do.
| from pathlib import Path | ||
| from GANDLF.utils import get_git_hash | ||
|
|
||
| readme_template = """ |
There was a problem hiding this comment.
Is this a simple copy of the model card template found here? If yes, I can suggest to either:
- directly reuse the template from
huggingface_hub(i.e.ModelCard.from_template(card_data)without thetemplate_str). - or define your own template but in this case you should only put the relevant fields and descriptions for your library (instead of having all fields as empty)
There was a problem hiding this comment.
Thanks @Wauplin for making me aware of this ,I will definitely go through it and make required changes as you mentioned
There was a problem hiding this comment.
Hi @Wauplin,
We had an internal discussion on what would be the best way for us to showcase potential model uploaders with a specific set of required options for the model card. Thus far, we have landed on using a custom model card. The reason to have all the fields present is provide the ability for a user to put in more information than what we require.
Here, we have put the string "REQUIRED_FOR_GANDLF" for the fields that are explicitly needed for the user to populate, and the rest have been left as present in the template.
In the code, we plan to add 2 checks:
- If
"REQUIRED_FOR_GANDLF"is found, we present an error to the user saying that this field needs to be populated with appropriate information. - The
Repositorykey should always behttps://github.com/mlcommons/GaNDLF.
Thoughts?
There was a problem hiding this comment.
Brilliant, thanks for the confirmation! We'll get on it right away. 😄
There was a problem hiding this comment.
@sarthakpati @Wauplin so how can we test this file if we propose the upload functionality as we only have entry points tests, do we have to mention a specific directory there
There was a problem hiding this comment.
Perhaps you can leverage one of the existing training tests to test the upload. I would recommend this one, since this would only upload a single model.
Ensure you put an appropriate description for it (such as Unit testing model or something) to make it clear for anyone viewing it. Is there a way to update an existing model, @Wauplin?
| tags += [git_hash] | ||
|
|
||
| card_data = ModelCardData(library_name="GaNDLF", tags=tags) | ||
| card = ModelCard.from_template(card_data, template_str=readme_template) |
| def download_from_hub( | ||
| repo_id: str, | ||
| revision: Union[str, None] = None, | ||
| cache_dir: Union[str, None] = None, | ||
| local_dir: Union[str, None] = None, | ||
| force_download: bool = False, | ||
| token: Union[str, None] = None, | ||
| ): | ||
| snapshot_download( | ||
| repo_id=repo_id, | ||
| revision=revision, | ||
| cache_dir=cache_dir, | ||
| local_dir=local_dir, | ||
| force_download=force_download, | ||
| token=token, | ||
| ) |
There was a problem hiding this comment.
I am not sure this alias is really needed. I would simply call snapshot_download in other places in the code.
There was a problem hiding this comment.
I still think the alias is not needed and that snapshot_download could be used by default
There was a problem hiding this comment.
@Wauplin actually this alias is use for the alignment of hugging face downloading feature with Gandlf command line design pattern ,change it to by default may abruptly conflict the command line argument
Co-authored-by: Lucain <[email protected]>
|
@pranayasinghcsmpl some lint fixes (unused variables and whatnot) will be needed for this PR. Thanks for taking care of it! Thanks for your comments and suggestions, @Wauplin! |
Codecov ReportAttention: Patch coverage is
Additional details and impacted files@@ Coverage Diff @@
## master #916 +/- ##
==========================================
+ Coverage 94.58% 94.61% +0.03%
==========================================
Files 161 164 +3
Lines 9567 9701 +134
==========================================
+ Hits 9049 9179 +130
- Misses 518 522 +4
Flags with carried forward coverage won't be shown. Click here to find out more. ☔ View full report in Codecov by Sentry. |
|
Support ticket generated with Codacy to explore the coverage issue. |
|
Codacy folks suggested not to use coverage reporter for anything coming in from other forks 🙄 Anyway, we should be good to go from my end. @Wauplin is this PR good to merge for you? |
| def download_from_hub( | ||
| repo_id: str, | ||
| revision: Union[str, None] = None, | ||
| cache_dir: Union[str, None] = None, | ||
| local_dir: Union[str, None] = None, | ||
| force_download: bool = False, | ||
| token: Union[str, None] = None, | ||
| ): | ||
| snapshot_download( | ||
| repo_id=repo_id, | ||
| revision=revision, | ||
| cache_dir=cache_dir, | ||
| local_dir=local_dir, | ||
| force_download=force_download, | ||
| token=token, | ||
| ) |
There was a problem hiding this comment.
I still think the alias is not needed and that snapshot_download could be used by default
| @click.option( | ||
| "--hf-template", | ||
| "-hft", | ||
| help="Adding the template path for the model card it is Required during Uploaing a model", | ||
| type=click.Path(exists=True, file_okay=True, dir_okay=False), | ||
| ) |
There was a problem hiding this comment.
Maybe default it to hugging_face.md to reduce friction? Users are free to provide another template if they want but having one by default should reduce friction and help grow usage.
There was a problem hiding this comment.
So we have to provide a default path for the hugging face template so that when ever the user want to upload a ,model by default that template will be uploaded ?
There was a problem hiding this comment.
That's what I suggest to do yes. There is already a hugging_face.md template in this PR so I was suggesting to reuse it.
There was a problem hiding this comment.
Ok thanks for you suggestion @Wauplin we will do the same
Co-authored-by: Lucain <[email protected]>
Co-authored-by: Lucain <[email protected]>
Co-authored-by: Lucain <[email protected]>
Co-authored-by: Lucain <[email protected]>
Fixes #727
Proposed Changes
Checklist
CONTRIBUTINGguide has been followed.typingis used to provide type hints, including and not limited to usingOptionalif a variable has a pre-defined value).pip installstep is needed for PR to be functional), please ensure it is reflected in all the files that control the CI, namely: python-test.yml, and all docker files [1,2,3].