Skip to content

Deployments

Deploying a trained model is straightforward. From the model page in the user interface a model can be deployed as an API. A name can be given together with the initial amount of replicas.

After a model is deployed, spinning up the actual service with the model can take anywhere from a few seconds to a few minutes. The deployment will show the number of replicas that are online together with the total desired number of replicas.

Next to the unique endpoint assigned to the deployment, you can create and swap static endpoints from the deployment page.

Endpoints

Routes

The available routes depend on the project. The /predict route is always available.

If a project contains a transformer with a transform_input method, this /predict route sends the input through the transform_input method before passing it to the model. In this case, there is also a /predict_raw route where raw features can be sent.

In case there is no transformer, /predict will pass the features to the trained model and send back the response.

Test deployment

The deployment page contains an option to interact with the API. On the left is a form input in case that is applicable for the input shape of your features. This is linked to the JSON input to make it easier to understand what the request looks like to call the model. On the right side, there is a button to make the actual request where the response is shown.

On top of the test deployment section, you can select whether to use the transformer input or a raw feature input if there is a transformer present.

Test deployment

On the right, there is a code generator that can generate code in multiple languages and frameworks to make connecting downstream tasks even faster.

Code generation