Project Repo: https://github.com/athletedecoded/rusty-pipe

Rusty Pipe [WiP]

E2E ML Pipeline -- from data to distroless deploy

Train

Download example dataset

$ cd train
$ make dataset

Train model

$ cargo run hymenoptera_data

Convert model for Deploy

$ python3 -m venv ~/.venv
$ source ~/.venv/bin/activate
$ pip install -r requirements.txt
$ make models

Test Locally

$ cd ../deploy
$ make run

Deploy to Azure Distroless Container

Provision Container App rusty-pipe in Azure. Update Container App > Settings > Ingress > Port = 8080.
From Azure CLI, generate Service Principle Credentials. Capture output and add to Github Repo as AZURE_CREDENTIALS Actions Repository Secret.

az ad sp create-for-rbac --name "RustyPipeAccess" --role contributor --scopes /subscriptions/$AZURE_SUBSCRIPTION_ID --sdk-auth

Generate GitHub PAT with write:packages access. Capture output and add to Github Repo as PAT Actions Repository Secret.
Push then manually trigger from GitHub Actions > Trigger Auto Deploy from branch: deploy-distro

git checkout -b deploy-distro
git push origin deploy-distro

Gotchas:

Rusty-Pipe Repo > Settings > Actions > General > Actions Permissions > Allow all actions
Rusty-Pipe Repo > Settings > Actions > General > Workflow Permissions > Read & Write

Local base URL https://localhost:8080

Example deployed URL https://rusty-pipe.livelyisland-7ce892fe.eastus.azurecontainerapps.io/

GET / -- Homepage

POST /predict -- Predict Image

curl -X POST -H "Content-Type: multipart/form-data" -F "image=@assets/ant.jpg" <base_url>/predict

Train

Dataset: create tch dataloader that takes train_val split with class subdirectories
Models: improve CNN, fix VGG, pass model as CLI param
Dynamic class generation --> txt file --> deploy

Deploy