Multi-model endpoints

It is possible to serve multiple models in the same Cortex API using any type of Cortex Predictor. In this guide we'll show the general outline of a multi-model deployment. The section for each predictor type is based on a corresponding example that can be found in the examples directory of the Cortex project.
Python Predictor
Specifying models in API config
The following template is based on the live-reloading/python/mpg-estimator example.
cortex.yaml
Even though it looks as if there's only a single model served, there are actually 4 different versions saved in s3://cortex-examples/sklearn/mpg-estimator/linreg/.
1
- name: mpg-estimator
2
  kind: RealtimeAPI
3
  predictor:
4
    type: python
5
    path: predictor.py
6
    model_path: s3://cortex-examples/sklearn/mpg-estimator/linreg/
Copied!
predictor.py
1
import mlflow.sklearn
2
import numpy as np
3
​
4
​
5
class PythonPredictor:
6
    def __init__(self, config, python_client):
7
        self.client = python_client
8
​
9
    def load_model(self, model_path):
10
        return mlflow.sklearn.load_model(model_path)
11
​
12
    def predict(self, payload, query_params):
13
        model_version = query_params.get("version")
14
​
15
        # process the input
16
        # ...
17
​
18
        model = self.client.get_model(model_version=model_version)
19
        result = model.predict(model_input)
20
​
21
        return {"prediction": result, "model": {"version": model_version}}
Copied!
Making predictions
For convenience, we'll export our API's endpoint (yours will be different from mine):
1
$ api_endpoint=http://a36473270de8b46e79a769850dd3372d-c67035afa37ef878.elb.us-west-2.amazonaws.com/mpg-estimator
Copied!
Next, we'll make a prediction using the sentiment analyzer model by specifying the model version as a query parameter:
1
$ curl "${api_endpoint}?version=1" -X POST -H "Content-Type: application/json" -d @sample.json
2
​
3
{"prediction": 26.929889872154185, "model": {"version": "1"}}
Copied!
Then we'll make a prediction using the 2nd version of the model (since they are just duplicate models, it will only return the same result):
1
$ curl "${api_endpoint}?version=2" -X POST -H "Content-Type: application/json" -d @sample.json
2
​
3
{"prediction": 26.929889872154185, "model": {"version": "2"}}
Copied!
Without specifying models in API config
For the Python Predictor, the API configuration for a multi-model API is similar to single-model APIs. The Predictor's config field can be used to customize the behavior of the predictor.py implementation.
The following template is based on the pytorch/multi-model-text-analyzer example.
cortex.yaml
1
- name: multi-model-text-analyzer
2
  kind: RealtimeAPI
3
  predictor:
4
    type: python
5
    path: predictor.py
6
    config: {...}
7
    ...
Copied!
predictor.py
Models should be loaded within the predictor's constructor. Query parameters are encouraged to be used when selecting the model for inference.
1
# import modules here
2
​
3
class PythonPredictor:
4
    def __init__(self, config):
5
        # prepare the environment, download/load models/labels, etc
6
        # ...
7
​
8
        # load models
9
        self.analyzer = initialize_model("sentiment-analysis")
10
        self.summarizer = initialize_model("summarization")
11
​
12
    def predict(self, query_params, payload):
13
        # preprocessing
14
        model_name = query_params.get("model")
15
        model_input = payload["text"]
16
        # ...
17
​
18
        # make prediction
19
        if model_name == "sentiment":
20
            results = self.analyzer(model_input)
21
            predicted_label = postprocess(results)
22
            return {"label": predicted_label}
23
        elif model_name == "summarizer":
24
            results = self.summarizer(model_input)
25
            predicted_label = postprocess(results)
26
            return {"label": predicted_label}
27
        else:
28
            return JSONResponse({"error": f"unknown model: {model_name}"}, status_code=400)
Copied!
Making predictions
For convenience, we'll export our API's endpoint (yours will be different from mine):
1
$ api_endpoint=http://a36473270de8b46e79a769850dd3372d-c67035afa37ef878.elb.us-west-2.amazonaws.com/multi-model-text-analyzer
Copied!
Next, we'll make a prediction using the sentiment analyzer model by specifying the model name as a query parameter:
1
$ curl "${api_endpoint}?model=sentiment" -X POST -H "Content-Type: application/json" -d @sample-sentiment.json
2
​
3
{"label": "POSITIVE", "score": 0.9998506903648376}
Copied!
Then we'll make a prediction using the text summarizer model:
1
$ curl "${api_endpoint}?model=summarizer" -X POST -H "Content-Type: application/json" -d @sample-summarizer.json
2
​
3
Machine learning is the study of algorithms and statistical models that computer systems use to perform a specific task. It is seen as a subset of artificial intelligence. Machine learning algorithms are used in a wide variety of applications, such as email filtering and computer vision. In its application across business problems, machine learning is also referred to as predictive analytics.
Copied!
TensorFlow Predictor
For the TensorFlow Predictor, a multi-model API is configured by placing the list of models in the Predictor's models field (each model will specify its own unique name). The predict() method of the tensorflow_client object expects a second argument that represents the name of the model that will be used for inference.
The following template is based on the tensorflow/multi-model-classifier example.
cortex.yaml
1
- name: multi-model-classifier
2
  kind: RealtimeAPI
3
  predictor:
4
    type: tensorflow
5
    path: predictor.py
6
    models:
7
      paths:
8
        - name: inception
9
          model_path: s3://cortex-examples/tensorflow/image-classifier/inception/
10
        - name: iris
11
          model_path: s3://cortex-examples/tensorflow/iris-classifier/nn/
12
        - name: resnet50
13
          model_path: s3://cortex-examples/tensorflow/resnet50/
14
      ...
Copied!
predictor.py
1
# import modules here
2
​
3
class TensorFlowPredictor:
4
    def __init__(self, tensorflow_client, config):
5
        # prepare the environment, download/load labels, etc
6
        # ...
7
​
8
        self.client = tensorflow_client
9
​
10
    def predict(self, payload, query_params):
11
        # preprocessing
12
        model_name = query_params["model"]
13
        model_input = preprocess(payload["url"])
14
​
15
        # make prediction
16
        results = self.client.predict(model_input, model_name)
17
​
18
        # postprocess
19
        predicted_label = postprocess(results)
20
​
21
        return {"label": predicted_label}
Copied!
Making predictions
For convenience, we'll export our API's endpoint (yours will be different from mine):
1
$ api_endpoint=http://a36473270de8b46e79a769850dd3372d-c67035afa37ef878.elb.us-west-2.amazonaws.com/multi-model-classifier
Copied!
Next, we'll make a prediction using the iris classifier model by specifying the model name as a query parameter:
1
$ curl "${ENDPOINT}?model=iris" -X POST -H "Content-Type: application/json" -d @sample-iris.json
2
​
3
{"label": "setosa"}
Copied!
Then we'll make a prediction using the resnet50 model:
1
$ curl "${ENDPOINT}?model=resnet50" -X POST -H "Content-Type: application/json" -d @sample-image.json
2
​
3
{"label": "sports_car"}
Copied!
Finally we'll make a prediction using the inception model:
1
$ curl "${ENDPOINT}?model=inception" -X POST -H "Content-Type: application/json" -d @sample-image.json
2
​
3
{"label": "sports_car"}
Copied!
ONNX Predictor
For the ONNX Predictor, a multi-model API is configured by placing the list of models in the Predictor's models field (each model will specify its own unique name). The predict() method of the onnx_client object expects a second argument that represents the name of the model that will be used for inference.
The following template is based on the onnx/multi-model-classifier example.
cortex.yaml
1
- name: multi-model-classifier
2
  kind: RealtimeAPI
3
  predictor:
4
    type: onnx
5
    path: predictor.py
6
    models:
7
      paths:
8
        - name: resnet50
9
          model_path: s3://cortex-examples/onnx/resnet50/
10
        - name: mobilenet
11
          model_path: s3://cortex-examples/onnx/mobilenet/
12
        - name: shufflenet
13
          model_path: s3://cortex-examples/onnx/shufflenet/
14
      ...
Copied!
predictor.py
1
# import modules here
2
​
3
class ONNXPredictor:
4
    def __init__(self, onnx_client, config):
5
        # prepare the environment, download/load labels, etc
6
        # ...
7
​
8
        self.client = onnx_client
9
​
10
    def predict(self, payload, query_params):
11
        # process the input
12
        model_name = query_params["model"]
13
        model_input = preprocess(payload["url"])
14
​
15
        # make prediction
16
        results = self.client.predict(model_input, model_name)
17
​
18
        # postprocess
19
        predicted_label = postprocess(results)
20
​
21
        return {"label": predicted_label}
Copied!
Making predictions
For convenience, we'll export our API's endpoint (yours will be different from mine):
1
$ api_endpoint=http://a36473270de8b46e79a769850dd3372d-c67035afa37ef878.elb.us-west-2.amazonaws.com/multi-model-classifier
Copied!
Next, we'll make a prediction using the resnet50 model by specifying the model name as a query parameter:
1
$ curl "${ENDPOINT}?model=resnet50" -X POST -H "Content-Type: application/json" -d @sample.json
2
​
3
{"label": "tabby"}
Copied!
Then we'll make a prediction using the mobilenet model:
1
$ curl "${ENDPOINT}?model=mobilenet" -X POST -H "Content-Type: application/json" -d @sample.json
2
​
3
{"label": "tabby"}
Copied!
Finally we'll make a prediction using the shufflenet model:
1
$ curl "${ENDPOINT}?model=shufflenet" -X POST -H "Content-Type: application/json" -d @sample.json
2
​
3
{"label": "Egyptian_cat"}
Copied!