Model Composer

Motivation

This use-case prompted the development of model-composer:

• You have two tensorflow models, one trained on weekday data and one trained on weekend data
• You would like to compose a single tensorflow model that can be used to generate predictions for any day of the week.
• You want the composed model to be natively defined in tensorflow - i.e. a single "computational graph" that can be easily loaded and used to make predictions.
• You want a single composed model becasuse:
  • It is easier to maintain than having to implement the logic to compose the models in every service that needs to make predictions.
  • It ensures the performance of the composed model will remain consistent with a native tensorflow model of a similar complexity.
  • It is easier to deploy a single model than multiple models

Documentation

The official documentation is hosted on ReadTheDocs: https://model-composer.readthedocs.io/

Install

Using pip:

pip install model-composer

Extras

Make use of extras to install the model composer implementations that you need:

pip install model-composer[tensorflow]  # compose tensorflow models
pip install model-composer[cloudpathlib]  # load models from cloud storage
pip install model-composer[all]  # all extras

Quick start

Declare your composed model in a yaml file which defines the components and how they should be composed.

name: "ride_share_pricing"
components:
  - name: weekday_model
    path: weekday_model.tf
    type: tensorflow
    where:
      input: is_weekday
      operator: eq
      value: true
  - name: weekend_model
    path: weekend_model.tf
    type: tensorflow
    where:
      input: is_weekday
      operator: eq
      value: false

Each component needs to have the following properties:

• name: The name of the component model
• path: The path to the component model on disk
• type: The type of the component model.
• where: The condition at which the component model should be used.

We build the weekend model and save it to disk.

import tensorflow as tf

# Build the weekend model
weekend_model = tf.keras.Sequential([
    tf.keras.layers.Input(shape=(1,), name="distance"),
    tf.keras.layers.Dense(1, name="price")
])

weekend_model.compile(optimizer="adam", loss="mse")

weekend_model.fit(
    x={"distance": tf.convert_to_tensor([10, 20], dtype=tf.float32)},
    y=tf.convert_to_tensor([10, 20], dtype=tf.float32),
    epochs=10
)
weekend_model.save("weekend_model.tf")

We build the weekday model and save it to disk.

# Build the weekday model
weekday_model = tf.keras.Sequential([
    tf.keras.layers.Input(shape=(1,), name="distance"),
    tf.keras.layers.Dense(1, name="price")
])

weekday_model.compile(optimizer="adam", loss="mse")

weekday_model.fit(
    x={"distance": tf.convert_to_tensor([10, 20], dtype=tf.float32)},
    y=tf.convert_to_tensor([5, 10], dtype=tf.float32),
    epochs=10
)

# Save the models
weekday_model.save("weekday_model.tf")

We can now build our composed model from the example yaml spec.

import tensorflow as tf
from model_composer import TensorflowModelComposer

composed_model = TensorflowModelComposer().from_yaml("example.yaml")

assert isinstance(composed_model, tf.keras.Model)

composed_model.save("composed_model.tf")

loaded_model = tf.keras.models.load_model("composed_model.tf")

composed_model.predict({
  "is_weekday": tf.convert_to_tensor([True, False], dtype=tf.bool),
  "distance": tf.convert_to_tensor([10, 20], dtype=tf.float32)
})

Roadmap

• Support for more ML frameworks:
  • PyTorch
  • Scikit-learn