Welcome fastai to the Hugging Face Hub

Few have done as much because the fast.ai ecosystem to make Deep Learning accessible. Our mission at Hugging Face is to democratize good Machine Learning. Let’s make exclusivity in access to Machine Learning, including pre-trained models, a thing of the past and let’s push this amazing field even further.

fastai is an open-source Deep Learning library that leverages PyTorch and Python to offer high-level components to coach fast and accurate neural networks with state-of-the-art outputs on text, vision, and tabular data. Nevertheless, fast.ai, the corporate, is greater than only a library; it has grown right into a thriving ecosystem of open source contributors and folks learning about neural networks. As some examples, try their book and courses. Join the fast.ai Discord and forums. It’s a guarantee that you’re going to learn by being a part of their community!

Due to all this, and more (the author of this post began his journey because of the fast.ai course), we’re proud to announce that fastai practitioners can now share and upload models to Hugging Face Hub with a single line of Python.

👉 On this post, we are going to introduce the combination between fastai and the Hub. Moreover, you possibly can open this tutorial as a Colab notebook.

We would like to thank the fast.ai community, notably Jeremy Howard, Wayde Gilliam, and Zach Mueller for his or her feedback 🤗. This blog is heavily inspired by the Hugging Face Hub section within the fastai docs.

Why share to the Hub?

The Hub is a central platform where anyone can share and explore models, datasets, and ML demos. It has essentially the most extensive collection of Open Source models, datasets, and demos.

Sharing on the Hub amplifies the impact of your fastai models by making them available for others to download and explore. You can too use transfer learning with fastai models; load another person’s model as the premise in your task.

Anyone can access all of the fastai models within the Hub by filtering the hf.co/models webpage by the fastai library, as within the image below.

Along with free model hosting and exposure to the broader community, the Hub has built-in version control based on git (git-lfs, for giant files) and model cards for discoverability and reproducibility. For more information on navigating the Hub, see this introduction.

Joining Hugging Face and installation

To share models within the Hub, you will have to have a user. Create it on the Hugging Face website.

The huggingface_hub library is a light-weight Python client with utility functions to interact with the Hugging Face Hub. To push fastai models to the hub, it’s worthwhile to have some libraries pre-installed (fastai>=2.4, fastcore>=1.3.27 and toml). You possibly can install them robotically by specifying [“fastai”] when installing huggingface_hub, and your environment is sweet to go:

pip install huggingface_hub["fastai"]

Making a fastai Learner

Here we train the first model within the fastbook to discover cats 🐱. We fully really useful reading all the fastbook.

from fastai.vision.all import *
path = untar_data(URLs.PETS)/'images'

def is_cat(x): return x[0].isupper()
dls = ImageDataLoaders.from_name_func(
    path, get_image_files(path), valid_pct=0.2, seed=42,
    label_func=is_cat, item_tfms=Resize(224))

learn = vision_learner(dls, resnet34, metrics=error_rate)
learn.fine_tune(1)

Sharing a Learner to the Hub

A Learner is a fastai object that bundles a model, data loaders, and a loss function. We’ll use the words Learner and Model interchangeably throughout this post.

First, log in to the Hugging Face Hub. You will want to create a write token in your Account Settings. Then there are three options to log in:

1. Type huggingface-cli login in your terminal and enter your token.

2. If in a python notebook, you need to use notebook_login.

from huggingface_hub import notebook_login

notebook_login()

3. Use the token argument of the push_to_hub_fastai function.

You possibly can input push_to_hub_fastai with the Learner you ought to upload and the repository id for the Hub within the format of “namespace/repo_name”. The namespace could be a person account or a corporation you’ve got write access to (for instance, ‘fastai/stanza-de’). For more details, check with the Hub Client documentation.

from huggingface_hub import push_to_hub_fastai


repo_id = "espejelomar/identify-my-cat"

push_to_hub_fastai(learner=learn, repo_id=repo_id)

The Learner is now within the Hub within the repo named espejelomar/identify-my-cat. An automatic model card is created with some links and next steps. When uploading a fastai Learner (or every other model) to the Hub, it is useful to edit its model card (image below) in order that others higher understand your work (check with the Hugging Face documentation).

if you ought to learn more about push_to_hub_fastai go to the Hub Client Documentation. There are some cool arguments you is likely to be fascinated by 👀. Remember, your model is a Git repository with all of the benefits that this entails: version control, commits, branches…

Loading a Learner from the Hugging Face Hub

Loading a model from the Hub is even simpler. We’ll load our Learner, “espejelomar/identify-my-cat”, and test it with a cat image (🦮?). This code is customized from
the first chapter of the fastbook.

First, upload a picture of a cat (or possibly a dog?). The Colab notebook with this tutorial uses ipywidgets to interactively upload a cat image (or not?). Here we are going to use this cute cat 🐅:

Now let’s load the Learner we just shared within the Hub and test it.

from huggingface_hub import from_pretrained_fastai


repo_id = "espejelomar/identify-my-cat"

learner = from_pretrained_fastai(repo_id)

It really works 👇!

_,_,probs = learner.predict(img)
print(f"Probability it is a cat: {100*probs[1].item():.2f}%")

Probability it's a cat: 100.00%

The Hub Client documentation includes addtional details on from_pretrained_fastai.

Blurr to combine fastai and Hugging Face Transformers (and share them)!

[Blurr is] a library designed for fastai developers who wish to train and deploy Hugging Face transformers – Blurr Docs.

We’ll:

1. Train a blurr Learner with the high-level Blurr API. It would load the distilbert-base-uncased model from the Hugging Face Hub and prepare a sequence classification model.
2. Share it to the Hub with the namespace fastai/blurr_IMDB_distilbert_classification using push_to_hub_fastai.
3. Load it with from_pretrained_fastai and take a look at it with learner_blurr.predict().

Collaboration and open-source are improbable!

First, install blurr and train the Learner.

git clone https://github.com/ohmeow/blurr.git
cd blurr
pip install -e ".[dev]"

import torch
import transformers
from fastai.text.all import *

from blurr.text.data.all import *
from blurr.text.modeling.all import *

path = untar_data(URLs.IMDB_SAMPLE)
model_path = Path("models")
imdb_df = pd.read_csv(path / "texts.csv")

learn_blurr = BlearnerForSequenceClassification.from_data(imdb_df, "distilbert-base-uncased", dl_kwargs={"bs": 4})
learn_blurr.fit_one_cycle(1, lr_max=1e-3)

Use push_to_hub_fastai to share with the Hub.

from huggingface_hub import push_to_hub_fastai


repo_id = "fastai/blurr_IMDB_distilbert_classification"

push_to_hub_fastai(learn_blurr, repo_id)

Use from_pretrained_fastai to load a blurr model from the Hub.

from huggingface_hub import from_pretrained_fastai


repo_id = "fastai/blurr_IMDB_distilbert_classification"

learner_blurr = from_pretrained_fastai(repo_id)

Try it with a pair sentences and review their sentiment (negative or positive) with learner_blurr.predict().

sentences = ["This integration is amazing!",
             "I hate this was not available before."]

probs = learner_blurr.predict(sentences)

print(f"Probability that sentence '{sentences[0]}' is negative is: {100*probs[0]['probs'][0]:.2f}%")
print(f"Probability that sentence '{sentences[1]}' is negative is: {100*probs[1]['probs'][0]:.2f}%")

Again, it really works!

Probability that sentence 'This integration is amazing!' is negative is: 29.46%
Probability that sentence 'I hate this was not available before.' is negative is: 70.04%

What’s next?

Take the fast.ai course (a new edition is coming soon), follow Jeremy Howard and fast.ai on Twitter for updates, and begin sharing your fastai models on the Hub 🤗. Or load considered one of the models which might be already within the Hub.

📧 Be at liberty to contact us via the Hugging Face Discord and share if you’ve got an idea for a project. We might love to listen to your feedback 💖.

Would you wish to integrate your library to the Hub?

This integration is made possible by the huggingface_hub library. If you ought to add your library to the Hub, we now have a guide for you! Or just tag someone from the Hugging Face team.

A shout out to the Hugging Face team for all of the work on this integration, specifically @osanseviero 🦙.

Thanks fastlearners and hugging learners 🤗.