Serverless functions are a great choice to deploy inference models! In this example, we deploy a small neural network to recognize hand-written digits.
In addition, this example shows how to load objects from Scaleway's S3 in rust.
You can test the sample application at: https://rust-example-www.s3-website.fr-par.scw.cloud/
- Install and configure Cargo
- Create an S3 bucket
- Install and configure Scaleway's Serverless Framework provider
- Setup your local environment:
| Variable | Description |
|---|---|
SCW_ACCESS_KEY |
Access key to use for S3 operations. |
SCW_SECRET_KEY |
Secret key to use for S3 operations. |
S3_BUCKET |
Name of the bucket to store the model into. |
SCW_DEFAULT_REGION |
Region of your bucket and function. Default: fr-par |
The model is taken from dfdx examples.
If you want to train the model locally you'll need to download the MNIST dataset. The deziped files can then be placed in ./datasets/mnist.
To train the model and upload it to your bucket, you can run:
cd training && cargo run -rOnce finished, the model will be serialized with numpy and uploaded to your S3 bucket.
A test server is available to make sure your function is working:
cargo run --bin test-server
$ curl localhost:3000 -d '{"data": [0, 1, ...]}'
>> {"output": [0.1233, 0.122, 0.12, ...]}You can deploy the function directly with Serverless Framework.
serverless deployIf you want to play around with the model, you can try the interface!
Replace <my-function-url> with your serverless function endpoint or localhost:3000 when using the local test server.
export VITE_SLS_FUNCTION_URL=<my-function-url>
pnpm run dev