Estimator

Learn how to use TensorFlow's Estimator API for model development.

We'll cover the following

Chapter Goals:

  • Create an Estimator object for the regression model

A. Creating an Estimator

TensorFlow’s Estimator object provides an organized and simple API for model execution. It handles model training, saving and restoring checkpoints, evaluating a model, and making predictions.

To initialize an Estimator object, we pass in the model function as a required argument. The model function should follow the same template as regressor_fn. The function must return an ExampleSpec object, which specifies the model results for training, evaluation, or prediction.

The two main keyword arguments to know are model_dir and params. The model_dir argument represents the directory we save model checkpoints to. The params argument represents the values we wish to pass into the model function. The argument should be set to a dictionary, which then corresponds to the model function’s params argument.

Press + to interact
import tensorflow as tf
params = {
    'feature_columns': feature_columns,
    'hidden_layers': hidden_layers
}
regressor = tf.estimator.Estimator(
    regressor_fn,
    model_dir=ckpt_dir,
    params=params)

In our example, we initialized the Estimator object with regressor_fn as its model function. We set the checkpoint directory to ckpt_dir, and passed in the feature columns and number of hidden layers through params.

Get hands-on with 1300+ tech skills courses.