Computer Vision Demo - DevOps for ML

This repository contains everything needed to automate training, testing and deployment of an example ML model with DevOps tools and CI/CD.

The original code can be found in Kaggle.com. Feel free to copy the notebook and run it on the cloud to get familiar with the demo:

https://www.kaggle.com/code/tomasfern/cats-or-dogs-classifier

The Model

The base model is resnet34, a 34-layer Convolutional Neural Network (CNN) for Deep Residual Learning for Image Recognition. This repository fine-tunes the base model to classify images as dogs or cats.

The Dataset

For fine-tuning, we'll use a subset of the Oxfort IIIT Pets dataset. The subset uses about 1800 labeled images of dogs and cats. The labels are taken from the name of the file. For example:

yorkshire_terrier_85.jpg - The filename begins with lowercase, indicating it's a dog. - The words indicate the breed of the dog. - The number indicates the sample item.

Russian_Blue_111.jpg - The filename begins with Uppercase, indicating it's a cat. - The words indicate the breed of the cat. - The number indicates the sample item.

The Application

We use streamlit to run a web application on top of the model.

Branches

main: the final state of the demo with CI/CD, DVC and ML pipelines.
noremote: same as main but without the DVC remote (no AWS S3 bucket required).
initial: the bare minimum to get started. No pipelines, no dvc installed.

Prerequisites

Before starting, you'll need the folliwing tools:

DVC
Python 3 and pip
Docker Desktop or Docker Engine
Git and Git-LFS
AWS CLI

It is also recommended to sign up for free accounts on the following websites:

Setup

Fork and clone this repository (branch=initial): git clone -b initial https://github.com/<USERNAME>/semaphore-demo-mlops.git
Create a virtualenv: python -m venv .venv
Activate it: source .venv/bin/activate
Install dependencies: pip install -r requirements.txt
Initialize the DVC repository: dvc init
Download the sample dataset: wget https://huggingface.co/datasets/tomfern/oxford-pets-subset/resolve/main/images.tar.gz -O data/images.tar.gz (alternative link: https://www.kaggle.com/datasets/tomasfern/oxford-iiit-pets-subset)
Ensure the downloaded tarball is located in data/images.tar.gz

Manual finetuning and deployment

To train the model manually:

Run data prepare script (unpack tarball): python src/prepare.py
Run training script: python src/train.py
Run test script: python src/test.py

You should now have the models files in the models/ directory.

You can now run the application with:

$ streamlit run src/app.py

Streamlit should open a browser window, you can upload pictures and get the model to classify them.

To run the application in a container:

$ docker build -t cats-and-dogs .
$ docker run -d -it -p 8501:8501 cats-and-dogs

Open your browser to https://localhost:8501 to use the application.

To deploy the application to HugginFace Spaces:

Create a HuggingFace account.
Create a SSH keypair and upload the public key to HugginFace.
Create a Streamlit Space on HuggingFace

Run the deployment script:

# eg ./deploy.sh https://huggingface.co/spaces/tomfern/cats-and-dogs /home/semaphore/.ssh/id_ed25519 
./deploy <huggingface_https_git_repo> <path_to_priv_key>

After a few minutes the application should be running in your Space.

DVC Workflow

We'll use DVC to track datasets and automate the whole process.

First, download

To setup a DVC ML Pipeline, use dvc stage add like this:

# prepare stage
$ dvc stage add -n prepare \
    -d src/prepare.py \
    -o data/images \
    python src/prepare.py

# train stsage
$ dvc stage add -n train \
    -d src/train.py -d data/images \
    -o models/model.pkl -o models/model.pth \
    -m metrics/classification.md \
    --plots metrics/confusion_matrix.png \
    --plots metrics/top_losses.png \
    --plots metrics/finetune_results.png \
    python src/train.py

# test stage
$ dvc stage add -n test \
    -d src/test.py -d models/model.pkl -d models/model.pth \
    python src/test.py

This will create dvc.yaml. You can see the dependecy graph with:

$ dvc dag

+---------+
| prepare |
+---------+
      *
      *
      *
 +-------+
 | train |
 +-------+
      *
      *
      *
  +------+
  | test |
  +------+

To run the pipeline:

$ dvc repro

This will execute the required steps (ala Makefile) only. After each execution you should commit dvc.yaml and dvc.lock to preserve the state of the training in Git.

Using Remote Storage

As of now, only the hashes of the trained model is stored in the GitHub. The actual model files are automatically added to .gitignore when the DVC stages are created, so they only exist in your machine. This makes sense because both the data and the model might turn up being too large and unwieldy for a GitHub repository.

To push the to the cloud you need to add remote storage. DVC supports multiple storage providers including Google Cloud and AWS S3 Buckets.

In the following example, we use an S3 Bucket for remote storage:

Go to your AWS Console and create an S3 Bucket (eg. mybucket-dvc)
Check you have access to the buckets from your machine: aws s3 ls
Add the AWS remote to DVC: dvc remote add mybucket-dvc s3://mybucket/main
Set the bucket as default: dvc remote default mybucket-dvc
Push the files to the bucket: dvc push

Now each time you make change you should both commit the changes to GitHub and then do a dvc push. You can get the latest files from the remote storage with dvc pull

Use the ML pipeline in another machine

You can reproduce the training in another machine by leveraging Git and the remote storage. Once you configure the AWS CLI to access your S3 bucket you can:

Pull from dvc remote repo: dvc pull
Run workflow: dvc repro
Push data and artifacts to remote repo: dvc push
Commit changes: git add -A && git commit

CI/CD

To setup a CI/CD pipeline you'll need a few things:

A token with access to the remote storage. For example, an IAM account with restricted access to the S3 bucket (read/write).
For containers, a token to access the Docker registry. For example, the user/password for a hub.docker.com account.
For running the app in HuggingFace Spaces, you'll need to upload your SSH pubkey and install GIT LFS.

Example configuration with Semaphore CI/CD:

Sign up with GitHub for a 15-day trial StartUp Semaphore account (the free account won't be enough)
Create secrets for:
- aws-s3-mlops: variables AWS_ACCESS_KEY_ID and AWS_SECRET_ACCESS_KEY with S3 bucket access (it's a good idea to create a programmatic IAM user)
- dockerhub: variables DOCKER_USERNAME and DOCKER_PASSWORD
- huggingface: upload private SSH key to folder /home/semaphore/.ssh/ (e.g id_ed25519)
- github: variable GITHUB_ACCESS_TOKEN with write permission to public repos.
Add your project to Semaphore
In the test block of the CI pipeline enable the github secret.
Ensure the secrets names are correct
Update the environment varibles in the deploy pipeline. They must point to your priv SSH key and HuggingFace Git repository.
Push changes and see your pipeline flow.

The main branch includes an example pipeline to train, test, containerize and deploy your application.

License

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the “Software”), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED “AS IS”, WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.

taltaf913 / semaphore-demo-mlops