Hosting Machine Learning Model Demos with Hugging Face Spaces and Streamlit
This article will describe the steps required to set up a demo for a machine / deep learning model on Hugging Face Spaces using Streamlit. The article is divided into the following parts:
- Introduction to the components
- How to access the camera on the user's device from the Hugging Face Space environment.
- How to set up the demo on HF Spaces
- Issues with the STUN server in the Streamlit app and possible solutions.
- How to manage Spaces using GitHub workflows and related issues.
Introduction to the components
What is Hugging Face Spaces?
Hugging Face provides an amazing tool to create and host your machine learning demos. The feature that provides this is Hugging Face Spaces. One can go to https://huggingface.co/spaces and find several hosted ML demos built by people from all over the world.
To create a new Space, click on the “Create new Space” button on the top right corner of the page. Space is similar to a GitHub repository, and the interface is also very similar to GitHub.
Steps:
- Enter a name for your repository and select the license you want your demo to have.
- Then select the library you want to use for the Space SDK. The input, output, and interface of the demo will be done using the chosen library. For this article, I will be using Streamlit.
- The next step is to choose the hardware you want your demo to run on. This can be changed later in the settings of your Space. The default is
“CPU basic · 2 vCPU · 16 GB · FREE”. You can switch to different hardware at any time in your Space settings. - Finally, choose the privacy setting for your space.
What is Streamlit?
Streamlit provides a faster way to build and share data apps. Streamlit turns data scripts into shareable web apps in minutes. You can find out more about it here: https://streamlit.io/
On your local machine, you can install Streamlit using the following command:
pip install streamlit
In the Hugging Face Spaces environment, Streamlit comes pre-installed.
How to access the camera on the user’s device from the Hugging Face Spaces environment
How do I access the camera?
If you want to show a demo on real-time data using the webcam on your laptop or even the camera on your mobile, you need to install the streamlit-webrtc library.
The streamlit-webrtc library is built for handling real-time video and audio over the network using Streamlit. It can be installed using the following command:
pip install streamlit-webrtc
A few examples on using Streamlit and streamlit-webrtc for different applications are also available in this GitHub repository
How to set up the demo on HF Spaces
What to do next to setup the demo?
Let's take a look at the steps to set up the demo, listed below
Set metadata
First of all, we need to set the metadata in the README.md file.
In my demo, I used the following lines
---
title: Object Detection MobileNetSSD Demo
emoji: 📷
colorFrom: red
colorTo: purple
sdk: streamlit
sdk_version: 1.26.0
app_file: app.py
pinned: false
license: lgpl-3.0
---
# MobileNet-SSD Object Detection Demo with Web App using Hugging Face Spaces
On my profile, it shows like this
Create a requirements.txt file.
In the requirements.txt file, add the libraries that must be installed in the environment. For example, in my demo, I had the following in my requirements.txt file:
numpy
opencv-python-headless==4.8.0.76
streamlit_webrtc==0.47.0
pydub==0.25.1
twilio~=8.5.0
Create a packages.txt file
In the packages.txt file, add the following line:
ffmpeg
Create a runtime.txt file
You can also create a runtime.txt file containing the Python version on which you want to run your code. However, these depend on the compatibility with other libraries being used. The runtime.txt file should contain the following line:
python-3.8.7
Create an app.py file
The app.py file will contain all the working code, which will control everything from taking input to showing the outputs using Streamlit. Of course, you can have other supporting files for your model. However, when the app is launched, you see the output of the following command:
streamlit run app.py
For running your app on your local machine, you can use the same command.
- Try copying one of the example demos from
to your app.py and commit.
As soon as you commit, it will start building the container. You can see the logs by clicking the “Logs” button in the top panel. You can see the libraries being installed and also track if any errors occur during the build or the app's execution.
Once the status changes from “Building” to “Running”, click on the “App” tab on the top panel. Here, you can see the interface of your app as per your instructions or code.
For hosting you own demo
To host your own demo, commit all the required files (supporting files containing the model definition and helper functions) to the repository and make the necessary changes to the app.py file. Before committing, you also need to change the requirements.txt, packages.txt, and runtime.txt files.
- For example, if you need a specific version of Torch or Cuda runtime, add the following lines:
torch==1.13.1+cu117
torchvision==0.14.1+cu117
torchaudio==0.13.1
--extra-index-url https://download.pytorch.org/whl/cu117
- If you are only using the free CPU tier, you won’t need Cuda. In that case, add the following lines (as per https://pytorch.org/get-started/previous-versions/)
torch==1.13.1+cpu
torchvision==0.14.1+cpu
torchaudio==0.13.1
--extra-index-url https://download.pytorch.org/whl/cpu
- If you need to commit large model files (> 10 MB), be sure to use git lfs.
Issues with the STUN server in the Streamlit app and possible solutions
With every application comes an Issue
The streamlit-webrtc uses Twilio STUN/TURN servers. If the credentials are not set, it falls to the free STUN servers provided by Google. However, the free STUN servers provided by Google are not always stable or working.
To use a Twilio STUN/TURN server, create an account on Twilio. Twilio provides free STUN servers and paid TURN servers. Twilio also provides an initial credit of about $15.50.
- Go to Settings of your HF Space
- Scroll down to “Variables and Secrets”
- Click on “New secret”
- Type in “TWILIO_ACCOUNT_SID” in the Name textbox. In the VALUE textbox, paste the Account SID from your Twilio > Accounts > API Keys & Tokens tab.
- Click on “New secret” again
- Type in “TWILIO_AUTH_TOKEN” in the Name textbox. In the VALUE textbox, paste the Auth Token from your Twilio > Accounts > API Keys & Tokens tab.
The container will start rebuilding. You can see your usage summary in Twilio > Billing > Usage Summary. To continue using the TURN server after the given credit runs out, you need to buy TURN server access from Twilio.
How to manage Spaces using GitHub workflows and related issues
Managing HF Spaces using GitHub workflows
You can also manage your Spaces from GitHub using GitHub Actions. Good documentation is provided here
Steps
- Create a repository on GitHub and name it anything you want. Subsequently, create an empty Space in Hugging Face Spaces.
- Add your Spaces app as an additional remote to your existing Git repository
git clone https://github.com/GIT_USERNAME/REPO_NAME
cd REPO_NAME
git remote add space https://huggingface.co/spaces/HF_USERNAME/SPACE_NAME
This is done to set up your GitHub repository and Spaces app together.
3. Then force push to sync everything for the first time:
git push --force space main
4. Create a GitHub Secret ( REPO > Settings > Secrets and variables > Actions ) with your HF_TOKEN
. You can find your Hugging Face API token under Access Tokens on your Hugging Face profile (If you don’t already have one, create a new token).
5. Configure a GitHub workflow and add the contents (also given below) in
to the workflow file. Replace HF_USERNAME
with your username and SPACE_NAME
with your Space name.
name: Sync to Hugging Face hub
on:
push:
branches: [main]
# to run this workflow manually from the Actions tab
workflow_dispatch:
jobs:
sync-to-hub:
runs-on: ubuntu-latest
steps:
- uses: actions/checkout@v3
with:
fetch-depth: 0
lfs: true
- name: Push to hub
env:
HF_TOKEN: ${{ secrets.HF_TOKEN }}
run: git push https://HF_USERNAME:$HF_TOKEN@huggingface.co/spaces/HF_USERNAME/SPACE_NAME main
You can add the second workflow given in the above link to check for large files.
Steps after editing the files
Note that, at this stage the Spaces repository is empty. After adding the required files run the following commands:
git add .
git commit -m “commit message”
git push origin main
You can see the changes made to the files in the GitHub repository being updated in the Hugging Face Spaces repository too.
Problems with large files
When trying to upload large model files (> 10MB), you need to use Git LFS. However, there is a limit to the bandwidth and storage in a free tier.
For uploading large files, before adding and committing, you need to run the following commands in order:
git lfs install
git lfs track *.[ext]
git add *.[ext]
git commit -m "Added large model file"
git push origin main
Exceeding the limits can cause syncing errors between Space and your GitHub repo.
A sample HF Spaces demo of a MobileNet SSD Object Detection is hosted here
The corresponding GitHub repository is
To embed your space on any website click the button indicated by the red rectangle in the picture below
Then click “Embed this space” from the drop-down menu. You can use the Iframe script or the Direct URL in your website.
The above HF Spaces demo is embedded below (click to open in a new window)
Clap and share if you like this article. Follow for more.