Deploy machine learning models in production

Cortex is an open source platform that takes machine learning models—trained with nearly any framework—and turns them into production web APIs in one command.



Below, we'll walk through how to use Cortex to deploy OpenAI's GPT-2 model as a service on AWS. You'll need to install Cortex on your AWS account before getting started.

Step 1: Configure your deployment

Define a deployment and an api resource. A deployment specifies a set of APIs that are deployed together. An api makes a model available as a web service that can serve real-time predictions. The configuration below will download the model from the cortex-examples S3 bucket. You can run the code that generated the model here.

# cortex.yaml
- kind: deployment
name: text
- kind: api
name: generator
model: s3://cortex-examples/text-generator/gpt-2/124M

Step 2: Add request handling

The model requires encoded data for inference, but the API should accept strings of natural language as input. It should also decode the inference output. This can be implemented in a request handler file using the pre_inference and post_inference functions:

from encoder import get_encoder
encoder = get_encoder()
def pre_inference(sample, metadata):
context = encoder.encode(sample["text"])
return {"context": [context]}
def post_inference(prediction, metadata):
response = prediction["sample"]
return encoder.decode(response)

Step 3: Deploy to AWS

Deploying to AWS is as simple as running cortex deploy from your CLI. cortex deploy takes the declarative configuration from cortex.yaml and creates it on the cluster. Behind the scenes, Cortex containerizes the model, makes it servable using TensorFlow Serving, exposes the endpoint with a load balancer, and orchestrates the workload on Kubernetes.

$ cortex deploy
deployment started

You can track the status of a deployment using cortex get. The output below indicates that one replica of the API was requested and one replica is available to serve predictions. Cortex will automatically launch more replicas if the load increases and spin down replicas if there is unused capacity.

$ cortex get generator --watch
status up-to-date available requested last update avg latency
live 1 1 1 8s 123ms
url: http://***

Step 4: Serve real-time predictions

Once you have your endpoint, you can make requests:

$ curl http://*** \
-X POST -H "Content-Type: application/json" \
-d '{"text": "machine learning"}'
Machine learning, with more than one thousand researchers around the world today, are looking to create computer-driven machine learning algorithms that can also be applied to human and social problems, such as education, health care, employment, medicine, politics, or the environment...

Any questions? chat with us.

More examples

Key features

  • Autoscaling: Cortex automatically scales APIs to handle production workloads.

  • Multi framework: Cortex supports TensorFlow, Keras, PyTorch, Scikit-learn, XGBoost, and more.

  • CPU / GPU support: Cortex can run inference on CPU or GPU infrastructure.

  • Rolling updates: Cortex updates deployed APIs without any downtime.

  • Log streaming: Cortex streams logs from deployed models to your CLI.

  • Prediction monitoring: Cortex monitors network metrics and tracks predictions.

  • Minimal declarative configuration: Deployments are defined in a single cortex.yaml file.