Introduce Tag Listing

[muellerzr]

Still a draft

This PR introduces the ability to fetch all available tags for models or datasets, and returns them as a nested namespace object.

Rather than taking the approach of #263, where we introduce a few Enum classes, I'm opting to copy fastcore's AttrDict class, which allows us to write nested objects from dictionaries quickly, allowing a user to perform both x.y.z and x[y][z]. The benefit here is it also comes with tab completion.

The repr for this class will show available attributes they can search by.

For a quick example, here's what this could look like:
tags = api.get_model_tags()

If we just look at the repr of tags it looks like so:

Available Attributes:
 * benchmark
 * language_creators
 * languages
 * licenses
 * multilinguality
 * size_categories
 * task_categories
 * task_ids

We can then look in tags.benchmark to see:

Available Attributes:
 * raft
 * superb
 * test

Before finally we look in tags.benchmark.raft to see:

benchmark:raft

Allowing a friendly user interface, without them having to remember benchmark:raft

Testing still needs to be done, but want your opinion on this introduction @julien-c @nateraw @LysandreJik

[muellerzr]

We could then use this AttrDict anywhere we want from the api calls as well later on, so it wouldn't necessarily be just for this thing.

(We could also call this AttributeDictionary so that it's not shorthanded)

[muellerzr]

Decided to go ahead with AttributeDictionary, and added some tests in.

One small tidbit of feedback I'd like is since testing both the API function and the ModelTags itself will look exactly the same, should we duplicate them? Or just include them in one vs the other.

[julien-c]

I wouldn't duplicate tests given that the test suite is already super slow. (we should aim to make it faster)

[muellerzr]

Makes sense. (Refactoring/making them faster is also on my radar). Will look at why the tests don't seem to like being run through ghactions, they pass locally for me which is problematic

[LysandreJik]

Great, I like the idea and implementation! As you know, I'm particularly fond of the autocompletion it offers which makes it easy to never leave the IDE/console/notebook and still know the exact wording of the different tags.

There's a small issue with it right now, however:

>>> tags.license
Available Attributes:
 * ???
 * academicfreelicensev3.0
 * agpl_3.0
 * apache2.0
 * apache_2
 * apache_2.0
 * bsd_3
 * bsd_3_clause
 * cc
 * cc0.0
 * cc0_1.0
 * cc_by_4.0
 * cc_by_nc_4.0
 * cc_by_nc_nd_4.0
 * cc_by_nc_sa
 * cc_by_nc_sa_3.0
 * cc_by_nc_sa_4.0
 * cc_by_sa_4.0
 * gpl
 * gpl_2.0
 * gpl_3.0
 * lgpl_lr
 * mit
 * mitlicense
 * pddl

>>> tags.license.cc_by_4.0
  File "<input>", line 1
    tags.license.cc_by_4.0
                         ^
SyntaxError: invalid syntax

Overall super excited about this!

[muellerzr]

@osanseviero would like your review on the endpoints.md + README please :)

[osanseviero]

Great work! I'm also very excited about this and I can already see cool things we could have apart from the search api. 🔥

Leaving some notes while playing around (not action items here, just thoughts as they come)

• I'm not a fan of multilinguality tag since it's misleading (internal issue here). On the other hand this is consistent with the UI, so it seems fine as is!

• I've always wondered which values exist for language_creators, so it's great to have this!

• These APIs will really be useful for explorations as well (cc @severo)

• Is tags.size_categories working as intended? The _repr returns this

Available Attributes:
 * n<1K
 * n>1M
 * unknown

but doing tags.size_categories.keys() gives me many other options, so something is off. My guess is because if not key[0].isdigit() here

• Should we sort alphabetically the output? For example, when doing tags.library having this sorted might be more readable for our users.

Available Attributes:
 * AdapterTransformers
 * Asteroid
 * ESPnet
 * Flair
 * JAX
 * Joblib
 * Keras
 * ONNX
 * PyTorch
 * Pyannote
 * Rust
 * Scikit_learn
 * SentenceTransformers
 * Stanza
 * TFLite
 * TensorBoard
 * TensorFlow
 * TensorFlowTTS
 * Timm
 * Transformers
 * allennlp
 * fastText
 * fastai
 * spaCy
 * speechbrain

[muellerzr]

Today I need to make two more changes and then it will be good to go, these are things I found through further toying last night:

• If there are any special characters at all we don't show it as an attribute, not just if it starts with a number. Since we can't do a.2<b, as python would read it as a.2 < b
• Figure out what tests are failing

[muellerzr]

@osanseviero once tests pass, we should be good to go I believe :) (in CI/CD)

[osanseviero]

This looks great! 🚀 🚀

[LysandreJik]

This looks good to me! How long does the test suite take to run?

[muellerzr]

Perfect! It only takes 1.7 seconds to run it