The tf.estimator framework makes it easy to construct and train machine
learning models via its highlevel Estimator API. Estimator
offers classes you can instantiate to quickly configure common model types such
as regressors and classifiers:
tf.estimator.LinearClassifier
: Constructs a linear classification model.tf.estimator.LinearRegressor
: Constructs a linear regression model.tf.estimator.DNNClassifier
: Construct a neural network classification model.tf.estimator.DNNRegressor
: Construct a neural network regression model.tf.estimator.DNNLinearCombinedClassifier
: Construct a neural network and linear combined classification model.tf.estimator.DNNLinearCombinedRegressor
: Construct a neural network and linear combined regression model.
But what if none of tf.estimator
's predefined model types meets your needs?
Perhaps you need more granular control over model configuration, such as
the ability to customize the loss function used for optimization, or specify
different activation functions for each neural network layer. Or maybe you're
implementing a ranking or recommendation system, and neither a classifier nor a
regressor is appropriate for generating predictions.
This tutorial covers how to create your own Estimator
using the building
blocks provided in tf.estimator
, which will predict the ages of
abalones based on their physical
measurements. You'll learn how to do the following:
 Instantiate an
Estimator
 Construct a custom model function
 Configure a neural network using
tf.feature_column
andtf.layers
 Choose an appropriate loss function from
tf.losses
 Define a training op for your model
 Generate and return predictions
Prerequisites
This tutorial assumes you already know tf.estimator API basics, such as
feature columns, input functions, and train()
/evaluate()
/predict()
operations. If you've never used tf.estimator before, or need a refresher,
you should first review the following tutorials:
 tf.estimator Quickstart: Quick introduction to training a neural network using tf.estimator.
 TensorFlow Linear Model Tutorial: Introduction to feature columns, and an overview on building a linear classifier in tf.estimator.
 Building Input Functions with tf.estimator: Overview of how to construct an input_fn to preprocess and feed data into your models.
An Abalone Age Predictor
It's possible to estimate the age of an abalone (sea snail) by the number of rings on its shell. However, because this task requires cutting, staining, and viewing the shell under a microscope, it's desirable to find other measurements that can predict age.
The Abalone Data Set contains the following feature data for abalone:
Feature  Description 

Length  Length of abalone (in longest direction; in mm) 
Diameter  Diameter of abalone (measurement perpendicular to length; in mm) 
Height  Height of abalone (with its meat inside shell; in mm) 
Whole Weight  Weight of entire abalone (in grams) 
Shucked Weight  Weight of abalone meat only (in grams) 
Viscera Weight  Gut weight of abalone (in grams), after bleeding 
Shell Weight  Weight of dried abalone shell (in grams) 
The label to predict is number of rings, as a proxy for abalone age.
“Abalone shell” (by Nicki Dugan Pogue, CC BYSA 2.0)
Setup
This tutorial uses three data sets.
abalone_train.csv
contains labeled training data comprising 3,320 examples.
abalone_test.csv
contains labeled test data for 850 examples.
abalone_predict
contains 7 examples on which to make predictions.
The following sections walk through writing the Estimator
code step by step;
the full, final code is available
here.
Loading Abalone CSV Data into TensorFlow Datasets
To feed the abalone dataset into the model, you'll need to download and load the
CSVs into TensorFlow Dataset
s. First, add some standard Python and TensorFlow
imports, and set up FLAGS:
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import argparse
import sys
import tempfile
# Import urllib
from six.moves import urllib
import numpy as np
import tensorflow as tf
FLAGS = None
Enable logging:
tf.logging.set_verbosity(tf.logging.INFO)
Then define a function to load the CSVs (either from files specified in commandline options, or downloaded from tensorflow.org):
def maybe_download(train_data, test_data, predict_data):
"""Maybe downloads training data and returns train and test file names."""
if train_data:
train_file_name = train_data
else:
train_file = tempfile.NamedTemporaryFile(delete=False)
urllib.request.urlretrieve(
"http://download.tensorflow.org/data/abalone_train.csv",
train_file.name)
train_file_name = train_file.name
train_file.close()
print("Training data is downloaded to %s" % train_file_name)
if test_data:
test_file_name = test_data
else:
test_file = tempfile.NamedTemporaryFile(delete=False)
urllib.request.urlretrieve(
"http://download.tensorflow.org/data/abalone_test.csv", test_file.name)
test_file_name = test_file.name
test_file.close()
print("Test data is downloaded to %s" % test_file_name)
if predict_data:
predict_file_name = predict_data
else:
predict_file = tempfile.NamedTemporaryFile(delete=False)
urllib.request.urlretrieve(
"http://download.tensorflow.org/data/abalone_predict.csv",
predict_file.name)
predict_file_name = predict_file.name
predict_file.close()
print("Prediction data is downloaded to %s" % predict_file_name)
return train_file_name, test_file_name, predict_file_name
Finally, create main()
and load the abalone CSVs into Datasets
, defining
flags to allow users to optionally specify CSV files for training, test, and
prediction datasets via the command line (by default, files will be downloaded
from tensorflow.org):
def main(unused_argv):
# Load datasets
abalone_train, abalone_test, abalone_predict = maybe_download(
FLAGS.train_data, FLAGS.test_data, FLAGS.predict_data)
# Training examples
training_set = tf.contrib.learn.datasets.base.load_csv_without_header(
filename=abalone_train, target_dtype=np.int, features_dtype=np.float64)
# Test examples
test_set = tf.contrib.learn.datasets.base.load_csv_without_header(
filename=abalone_test, target_dtype=np.int, features_dtype=np.float64)
# Set of 7 examples for which to predict abalone ages
prediction_set = tf.contrib.learn.datasets.base.load_csv_without_header(
filename=abalone_predict, target_dtype=np.int, features_dtype=np.float64)
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.register("type", "bool", lambda v: v.lower() == "true")
parser.add_argument(
"train_data", type=str, default="", help="Path to the training data.")
parser.add_argument(
"test_data", type=str, default="", help="Path to the test data.")
parser.add_argument(
"predict_data",
type=str,
default="",
help="Path to the prediction data.")
FLAGS, unparsed = parser.parse_known_args()
tf.app.run(main=main, argv=[sys.argv[0]] + unparsed)
Instantiating an Estimator
When defining a model using one of tf.estimator's provided classes, such as
DNNClassifier
, you supply all the configuration parameters right in the
constructor, e.g.:
my_nn = tf.estimator.DNNClassifier(feature_columns=[age, height, weight],
hidden_units=[10, 10, 10],
activation_fn=tf.nn.relu,
dropout=0.2,
n_classes=3,
optimizer="Adam")
You don't need to write any further code to instruct TensorFlow how to train the
model, calculate loss, or return predictions; that logic is already baked into
the DNNClassifier
.
By contrast, when you're creating your own estimator from scratch, the
constructor accepts just two highlevel parameters for model configuration,
model_fn
and params
:
nn = tf.estimator.Estimator(model_fn=model_fn, params=model_params)

model_fn
: A function object that contains all the aforementioned logic to support training, evaluation, and prediction. You are responsible for implementing that functionality. The next section, Constructing themodel_fn
covers creating a model function in detail. 
params
: An optional dict of hyperparameters (e.g., learning rate, dropout) that will be passed into themodel_fn
.
For the abalone age predictor, the model will accept one hyperparameter:
learning rate. Define LEARNING_RATE
as a constant at the beginning of your
code (highlighted in bold below), right after the logging configuration:
tf.logging.set_verbosity(tf.logging.INFO)
# Learning rate for the model
LEARNING_RATE = 0.001
Then, add the following code to main()
, which creates the dict model_params
containing the learning rate and instantiates the Estimator
:
# Set model params
model_params = {"learning_rate": LEARNING_RATE}
# Instantiate Estimator
nn = tf.estimator.Estimator(model_fn=model_fn, params=model_params)
Constructing the model_fn
The basic skeleton for an Estimator
API model function looks like this:
def model_fn(features, labels, mode, params):
# Logic to do the following:
# 1. Configure the model via TensorFlow operations
# 2. Define the loss function for training/evaluation
# 3. Define the training operation/optimizer
# 4. Generate predictions
# 5. Return predictions/loss/train_op/eval_metric_ops in EstimatorSpec object
return EstimatorSpec(mode, predictions, loss, train_op, eval_metric_ops)
The model_fn
must accept three arguments:
features
: A dict containing the features passed to the model viainput_fn
.labels
: ATensor
containing the labels passed to the model viainput_fn
. Will be empty forpredict()
calls, as these are the values the model will infer.mode
: One of the followingtf.estimator.ModeKeys
string values indicating the context in which the model_fn was invoked:tf.estimator.ModeKeys.TRAIN
Themodel_fn
was invoked in training mode, namely via atrain()
call.tf.estimator.ModeKeys.EVAL
. Themodel_fn
was invoked in evaluation mode, namely via anevaluate()
call.tf.estimator.ModeKeys.PREDICT
. Themodel_fn
was invoked in predict mode, namely via apredict()
call.
model_fn
may also accept a params
argument containing a dict of
hyperparameters used for training (as shown in the skeleton above).
The body of the function performs the following tasks (described in detail in the sections that follow):
 Configuring the model—here, for the abalone predictor, this will be a neural network.
 Defining the loss function used to calculate how closely the model's predictions match the target values.
 Defining the training operation that specifies the
optimizer
algorithm to minimize the loss values calculated by the loss function.
The model_fn
must return a tf.estimator.EstimatorSpec
object, which contains the following values:

mode
(required). The mode in which the model was run. Typically, you will return themode
argument of themodel_fn
here. 
predictions
(required inPREDICT
mode). A dict that maps key names of your choice toTensor
s containing the predictions from the model, e.g.:python predictions = {"results": tensor_of_predictions}
In
PREDICT
mode, the dict that you return inEstimatorSpec
will then be returned bypredict()
, so you can construct it in the format in which you'd like to consume it. 
loss
(required inEVAL
andTRAIN
mode). ATensor
containing a scalar loss value: the output of the model's loss function (discussed in more depth later in Defining loss for the model) calculated over all the input examples. This is used inTRAIN
mode for error handling and logging, and is automatically included as a metric inEVAL
mode. 
train_op
(required only inTRAIN
mode). An Op that runs one step of training. 
eval_metric_ops
(optional). A dict of name/value pairs specifying the metrics that will be calculated when the model runs inEVAL
mode. The name is a label of your choice for the metric, and the value is the result of your metric calculation. Thetf.metrics
module provides predefined functions for a variety of common metrics. The followingeval_metric_ops
contains an"accuracy"
metric calculated usingtf.metrics.accuracy
:python eval_metric_ops = { "accuracy": tf.metrics.accuracy(labels, predictions) }
If you do not specify
eval_metric_ops
, onlyloss
will be calculated during evaluation.
Configuring a neural network with tf.feature_column
and tf.layers
Constructing a neural network entails creating and connecting the input layer, the hidden layers, and the output layer.
The input layer is a series of nodes (one for each feature in the model) that
will accept the feature data that is passed to the model_fn
in the features
argument. If features
contains an ndimensional Tensor
with all your feature
data, then it can serve as the input layer.
If features
contains a dict of feature columns passed to
the model via an input function, you can convert it to an inputlayer Tensor
with the tf.feature_column.input_layer
function.
input_layer = tf.feature_column.input_layer(
features=features, feature_columns=[age, height, weight])
As shown above, input_layer()
takes two required arguments:
features
. A mapping from string keys to theTensors
containing the corresponding feature data. This is exactly what is passed to themodel_fn
in thefeatures
argument.feature_columns
. A list of all theFeatureColumns
in the model—age
,height
, andweight
in the above example.
The input layer of the neural network then must be connected to one or more
hidden layers via an activation
function that performs a
nonlinear transformation on the data from the previous layer. The last hidden
layer is then connected to the output layer, the final layer in the model.
tf.layers
provides the tf.layers.dense
function for constructing fully
connected layers. The activation is controlled by the activation
argument.
Some options to pass to the activation
argument are:

tf.nn.relu
. The following code creates a layer ofunits
nodes fully connected to the previous layerinput_layer
with a ReLU activation function (tf.nn.relu
):python hidden_layer = tf.layers.dense( inputs=input_layer, units=10, activation=tf.nn.relu)

tf.nn.relu6
. The following code creates a layer ofunits
nodes fully connected to the previous layerhidden_layer
with a ReLU 6 activation function (tf.nn.relu6
):python second_hidden_layer = tf.layers.dense( inputs=hidden_layer, units=20, activation=tf.nn.relu)

None
. The following code creates a layer ofunits
nodes fully connected to the previous layersecond_hidden_layer
with no activation function, just a linear transformation:python output_layer = tf.layers.dense( inputs=second_hidden_layer, units=3, activation=None)
Other activation functions are possible, e.g.:
output_layer = tf.layers.dense(inputs=second_hidden_layer,
units=10,
activation_fn=tf.sigmoid)
The above code creates the neural network layer output_layer
, which is fully
connected to second_hidden_layer
with a sigmoid activation function
(tf.sigmoid
). For a list of predefined
activation functions available in TensorFlow, see the API docs.
Putting it all together, the following code constructs a full neural network for the abalone predictor, and captures its predictions:
def model_fn(features, labels, mode, params):
"""Model function for Estimator."""
# Connect the first hidden layer to input layer
# (features["x"]) with relu activation
first_hidden_layer = tf.layers.dense(features["x"], 10, activation=tf.nn.relu)
# Connect the second hidden layer to first hidden layer with relu
second_hidden_layer = tf.layers.dense(
first_hidden_layer, 10, activation=tf.nn.relu)
# Connect the output layer to second hidden layer (no activation fn)
output_layer = tf.layers.dense(second_hidden_layer, 1)
# Reshape output layer to 1dim Tensor to return predictions
predictions = tf.reshape(output_layer, [1])
predictions_dict = {"ages": predictions}
...
Here, because you'll be passing the abalone Datasets
using numpy_input_fn
as shown below, features
is a dict {"x": data_tensor}
, so
features["x"]
is the input layer. The network contains two hidden
layers, each with 10 nodes and a ReLU activation function. The output layer
contains no activation function, and is
tf.reshape
to a onedimensional
tensor to capture the model's predictions, which are stored in
predictions_dict
.
Defining loss for the model
The EstimatorSpec
returned by the model_fn
must contain loss
: a Tensor
representing the loss value, which quantifies how well the model's predictions
reflect the label values during training and evaluation runs. The tf.losses
module provides convenience functions for calculating loss using a variety of
metrics, including:

absolute_difference(labels, predictions)
. Calculates loss using the absolutedifference formula (also known as L_{1} loss). 
log_loss(labels, predictions)
. Calculates loss using the logistic loss forumula (typically used in logistic regression). 
mean_squared_error(labels, predictions)
. Calculates loss using the mean squared error (MSE; also known as L_{2} loss).
The following example adds a definition for loss
to the abalone model_fn
using mean_squared_error()
(in bold):
def model_fn(features, labels, mode, params):
"""Model function for Estimator."""
# Connect the first hidden layer to input layer
# (features["x"]) with relu activation
first_hidden_layer = tf.layers.dense(features["x"], 10, activation=tf.nn.relu)
# Connect the second hidden layer to first hidden layer with relu
second_hidden_layer = tf.layers.dense(
first_hidden_layer, 10, activation=tf.nn.relu)
# Connect the output layer to second hidden layer (no activation fn)
output_layer = tf.layers.dense(second_hidden_layer, 1)
# Reshape output layer to 1dim Tensor to return predictions
predictions = tf.reshape(output_layer, [1])
predictions_dict = {"ages": predictions}
# Calculate loss using mean squared error
loss = tf.losses.mean_squared_error(labels, predictions)
...
See the API guide for a full list of loss functions and more details on supported arguments and usage.
Supplementary metrics for evaluation can be added to an eval_metric_ops
dict.
The following code defines an rmse
metric, which calculates the root mean
squared error for the model predictions. Note that the labels
tensor is cast
to a float64
type to match the data type of the predictions
tensor, which
will contain real values:
eval_metric_ops = {
"rmse": tf.metrics.root_mean_squared_error(
tf.cast(labels, tf.float64), predictions)
}
Defining the training op for the model
The training op defines the optimization algorithm TensorFlow will use when
fitting the model to the training data. Typically when training, the goal is to
minimize loss. A simple way to create the training op is to instantiate a
tf.train.Optimizer
subclass and call the minimize
method.
The following code defines a training op for the abalone model_fn
using the
loss value calculated in Defining Loss for the Model, the
learning rate passed to the function in params
, and the gradient descent
optimizer. For global_step
, the convenience function
tf.train.get_global_step
takes care of generating an integer variable:
optimizer = tf.train.GradientDescentOptimizer(
learning_rate=params["learning_rate"])
train_op = optimizer.minimize(
loss=loss, global_step=tf.train.get_global_step())
For a full list of optimizers, and other details, see the API guide.
The complete abalone model_fn
Here's the final, complete model_fn
for the abalone age predictor. The
following code configures the neural network; defines loss and the training op;
and returns a EstimatorSpec
object containing mode
, predictions_dict
, loss
,
and train_op
:
def model_fn(features, labels, mode, params):
"""Model function for Estimator."""
# Connect the first hidden layer to input layer
# (features["x"]) with relu activation
first_hidden_layer = tf.layers.dense(features["x"], 10, activation=tf.nn.relu)
# Connect the second hidden layer to first hidden layer with relu
second_hidden_layer = tf.layers.dense(
first_hidden_layer, 10, activation=tf.nn.relu)
# Connect the output layer to second hidden layer (no activation fn)
output_layer = tf.layers.dense(second_hidden_layer, 1)
# Reshape output layer to 1dim Tensor to return predictions
predictions = tf.reshape(output_layer, [1])
# Provide an estimator spec for `ModeKeys.PREDICT`.
if mode == tf.estimator.ModeKeys.PREDICT:
return tf.estimator.EstimatorSpec(
mode=mode,
predictions={"ages": predictions})
# Calculate loss using mean squared error
loss = tf.losses.mean_squared_error(labels, predictions)
# Calculate root mean squared error as additional eval metric
eval_metric_ops = {
"rmse": tf.metrics.root_mean_squared_error(
tf.cast(labels, tf.float64), predictions)
}
optimizer = tf.train.GradientDescentOptimizer(
learning_rate=params["learning_rate"])
train_op = optimizer.minimize(
loss=loss, global_step=tf.train.get_global_step())
# Provide an estimator spec for `ModeKeys.EVAL` and `ModeKeys.TRAIN` modes.
return tf.estimator.EstimatorSpec(
mode=mode,
loss=loss,
train_op=train_op,
eval_metric_ops=eval_metric_ops)
Running the Abalone Model
You've instantiated an Estimator
for the abalone predictor and defined its
behavior in model_fn
; all that's left to do is train, evaluate, and make
predictions.
Add the following code to the end of main()
to fit the neural network to the
training data and evaluate accuracy:
train_input_fn = tf.estimator.inputs.numpy_input_fn(
x={"x": np.array(training_set.data)},
y=np.array(training_set.target),
num_epochs=None,
shuffle=True)
# Train
nn.train(input_fn=train_input_fn, steps=5000)
# Score accuracy
test_input_fn = tf.estimator.inputs.numpy_input_fn(
x={"x": np.array(test_set.data)},
y=np.array(test_set.target),
num_epochs=1,
shuffle=False)
ev = nn.evaluate(input_fn=test_input_fn)
print("Loss: %s" % ev["loss"])
print("Root Mean Squared Error: %s" % ev["rmse"])
Then run the code. You should see output like the following:
...
INFO:tensorflow:loss = 4.86658, step = 4701
INFO:tensorflow:loss = 4.86191, step = 4801
INFO:tensorflow:loss = 4.85788, step = 4901
...
INFO:tensorflow:Saving evaluation summary for 5000 step: loss = 5.581
Loss: 5.581
The loss score reported is the mean squared error returned from the model_fn
when run on the ABALONE_TEST
data set.
To predict ages for the ABALONE_PREDICT
data set, add the following to
main()
:
# Print out predictions
predict_input_fn = tf.estimator.inputs.numpy_input_fn(
x={"x": prediction_set.data},
num_epochs=1,
shuffle=False)
predictions = nn.predict(input_fn=predict_input_fn)
for i, p in enumerate(predictions):
print("Prediction %s: %s" % (i + 1, p["ages"]))
Here, the predict()
function returns results in predictions
as an iterable.
The for
loop enumerates and prints out the results. Rerun the code, and you
should see output similar to the following:
...
Prediction 1: 4.92229
Prediction 2: 10.3225
Prediction 3: 7.384
Prediction 4: 10.6264
Prediction 5: 11.0862
Prediction 6: 9.39239
Prediction 7: 11.1289
Additional Resources
Congrats! You've successfully built a tf.estimator Estimator
from scratch.
For additional reference materials on building Estimator
s, see the following
sections of the API guides: