3.2

Package a Model Image from a Reusable Base Image

Enterprise Applicable to Enterprise tier only

This doc shows how to build a base image and use it. Using the base image has some benefits:

  1. If the way you load and use models is the same, these models can share the same base image
  2. If you just want to update the model file, a base image can speed up the building process

The idea is that you write a general model serving code and assume the model file is placed under a certain path. And build it as the base image by s2i.

When a model file is ready, use docker to copy the model file into the correct path and build the model deployment image.

In the following section, this doc uses Tensorflow 2 as a simple showcase. The code is under Github.

Prerequisites

Check everything is ready to go by running:

s2i usage seldonio/seldon-core-s2i-python3:0.18

Build the Base Image (Python)

  • Please use Python 3.6 (Recommended)

  • Write a general model serving code Model.py

    import tensorflow as tf

class Model:
    def __init__(self):
        self._model = tf.keras.models.load_model('model')

    def predict(self, X, feature_names=None, meta=None):
        output = self._model.predict(X)
        return output

  • Create a requirements.txt file and write down all required packages

    tensorflow==2.1.0

  • Create a .s2i folder and create a .s2i/environment file with the following content:

    MODEL_NAME=Model
API_TYPE=REST
SERVICE_TYPE=MODEL
PERSISTENCE=0

Build the base image by:

 s2i build . seldonio/seldon-core-s2i-python3:0.18 tensorflow2-prepackage

(Using seldonio/seldon-core-s2i-python3 instead if using Python 3 rather than Python 3.6)

Use the Base Image to Build the Model Deployment Image

Based on our previous base image, whenever you have a model outputted by:

model.save(export_path)

You can use this base image to build your model deployment image.

First, create a Dockerfile:

FROM tensorflow2-prepackage
COPY export_path model

(Please replace the export_path to your path)

This means you copy your model files into the path that you pre-defined in the base image code.

Then, you can build the model deployment image by:

docker build -t tensorflow2-prepackage-model .

Verify Your Model Deployment Image

In order to verify the image, you can run it:

docker run -p 5000:5000 --rm tensorflow2-prepackage-model

And send a post request by the following format:

curl -X POST localhost:5000/api/v1.0/predictions \
    -H 'Content-Type: application/json' \
    -d '{ "data": {"ndarray": [${INPUT_DATA}] } }'

The ${INPUT_DATA} is the data that you can feed into deployed model for prediction.

The dimension of input data must be same as model's input shape.

For example, if we create our model with a specified input_shape=(4,) by the following definition:

model = tf.keras.models.Sequential([
    keras.layers.Dense(64, activation='relu', input_shape=(4,)),
    ...
    ...
])

Then, we can send a post request that ${INPUT_DATA} with shape 4.

curl -X POST localhost:5000/api/v1.0/predictions \
    -H 'Content-Type: application/json' \
    -d '{ "data": {"ndarray": [[5.1, 3.3, 1.7, 0.5]] } }'

Or if we create our model with a specified input_shape=(2,2) by the following definition:

model = tf.keras.models.Sequential([
    keras.layers.Dense(64, activation='relu', input_shape=(2,2)),
    ...
    ...
])

Then, we can also send a post request that ${INPUT_DATA} with shape (2,2).

curl -X POST localhost:5000/api/v1.0/predictions \
    -H 'Content-Type: application/json' \
    -d '{ "data": {"ndarray": [[[5.1, 3.3], [1.7, 0.5]]] } }'

After sending the post request, we can obtain the response output in the following format:

{"data":{"names":[],"ndarray":[${PREDICTION_RESULT}]},"meta":{}}

The ${PREDICTION_RESULT} is a list to represent the prediction value.

For example, the following output shows three prediction values in each class.

{"data":{"names":[],"ndarray":[[3.093,-0.519,-8.918]]},"meta":{}}

After verifying your model deployment image, now you can use this image in the PrimeHub model deployment feature.

Share Your Base Image

Share your base image by pushing it to a docker registry.

Therefore, others can re-use the model serving code again. They can share the same base image and build a model deployment image by docker.

← Package a Model Image for RPersistence →