Custom Estimators
The tfestimators framework makes it easy to construct and build 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.
But what if none of the 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. The figure on the right illustrates the basic components of an estimator. Users can implement custom behaviors and or architecture inside the model_fn
of the estimator.
This tutorial covers how to create your own Estimator
using the building blocks provided in tfestimators
package, 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
The complete code for this tutorial can be found here.
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) 
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.
Downloading and Loading Abalone CSV Data
We first write a function that downloads the training, testing, and evaluation data from TensorFlow website if we haven’t downloaded them before.
library(tfestimators)
maybe_download_abalone < function(train_data_path, test_data_path, predict_data_path, column_names_to_assign) {
if (!file.exists(train_data_path)  !file.exists(test_data_path)  !file.exists(predict_data_path)) {
cat("Downloading abalone data ...")
train_data < read.csv("http://download.tensorflow.org/data/abalone_train.csv", header = FALSE)
test_data < read.csv("http://download.tensorflow.org/data/abalone_test.csv", header = FALSE)
predict_data < read.csv("http://download.tensorflow.org/data/abalone_predict.csv", header = FALSE)
colnames(train_data) < column_names_to_assign
colnames(test_data) < column_names_to_assign
colnames(predict_data) < column_names_to_assign
write.csv(train_data, train_data_path, row.names = FALSE)
write.csv(test_data, test_data_path, row.names = FALSE)
write.csv(predict_data, predict_data_path, row.names = FALSE)
} else {
train_data < read.csv(train_data_path, header = TRUE)
test_data < read.csv(test_data_path, header = TRUE)
predict_data < read.csv(predict_data_path, header = TRUE)
}
return(list(train_data = train_data, test_data = test_data, predict_data = predict_data))
}
COLNAMES < c("length", "diameter", "height", "whole_weight", "shucked_weight", "viscera_weight", "shell_weight", "num_rings")
downloaded_data < maybe_download_abalone(
file.path(getwd(), "train_abalone.csv"),
file.path(getwd(), "test_abalone.csv"),
file.path(getwd(), "predict_abalone.csv"),
COLNAMES
)
train_data < downloaded_data$train_data
test_data < downloaded_data$test_data
predict_data < downloaded_data$predict_data
Next, we construct the input function as follows:
constructed_input_fn < function(dataset) {
input_fn(dataset, features = num_rings, response = num_rings, num_epochs = NULL)
}
train_input_fn < constructed_input_fn(train_data)
test_input_fn < constructed_input_fn(test_data)
predict_input_fn < constructed_input_fn(predict_data)
Instantiating an Estimator
When defining a model using one of tf.estimator’s provided classes, such as linear_dnn_combined_classifier
, you supply all the configuration parameters right in the constructor, e.g.:
diameter < column_numeric("diameter")
height < column_numeric("height")
model < dnn_linear_combined_classifier(
linear_feature_columns = feature_columns(diameter),
dnn_feature_columns = feature_columns(height),
dnn_hidden_units = c(100L, 50L)
)
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 linear_dnn_combined_classifier
.
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
:
model < estimator(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
.
Note: Just like tfestimators
’ predefined regressors and classifiers, the estimator
initializer also accepts the general configuration arguments model_dir
and config
.
For the abalone age predictor, the model will accept one hyperparameter: learning rate. Here, learning_rate
is set to 0.001
, but you can tune this value as needed to achieve the best results during model training.
The following code creates the list model_params
containing the learning rate and instantiates the Estimator
:
# Set model params
model_params < list(learning_rate = 0.001)
# Instantiate Estimator
model < estimator(model_fn, params = model_params)
Constructing the model_fn
The basic skeleton for an Estimator
API model function looks like this:
model_fn < function(features, labels, mode, params, config) {
# 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 estimator_spec object
}
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 followingmode_keys()
string values indicating the context in which the model_fn was invoked:
"train"
Themodel_fn
was invoked in training mode, namely via atrain()
call. 
"eval"
. Themodel_fn
was invoked in evaluation mode, namely via anevaluate()
call. 
"infer"
. 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) and a config
that represents the configurations used in a model, including GPU percentage, cluster information, etc.
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 an estimator_spec
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 ininfer
mode). A dict that maps key names of your choice toTensor
s containing the predictions from the model, e.g.:
predictions < list(results = tensor_of_predictions)
In infer
mode, the dict that you return in estimator_spec
will then be returned by predict()
, so you can construct it in the format in which you’d like to consume it.
loss
(required ineval
andtrain
modes). 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$accuracy
:
eval_metric_ops < list(accuracy = tf$metrics$accuracy(labels, predictions))
If you do not specify eval_metric_ops
, only loss
will be calculated during evaluation.
Configuring a neural network with feature_column and 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 input_layer
function:
input_layer < input_layer(
features = features, feature_columns = c(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:
hidden_layer < tf$layers$dense(inputs = input_layer, units = 10L, 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:
second_hidden_layer < tf$layers$dense(
inputs = hidden_layer, units = 20L, activation = tf$nn$relu)

NULL
. The following code creates a layer ofunits
nodes fully connected to the previous layersecond_hidden_layer
with no activation function, just a linear transformation:
output_layer < tf$layers$dense(inputs = second_hidden_layer,
units = 3L, activation = NULL)
Other activation functions are possible, e.g.:
output_layer < tf$layers$dense(inputs = second_hidden_layer,
units = 10L, 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
.
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$reshape
to a onedimensional tensor to capture the model’s predictions, which are stored in predictions_dict
.
Defining loss for the model
The estimator_spec
returned by the model_fn
must contain loss
: a Tensor
representing the loss value, which quantifies how well the model’s predictions reflect the label values during training and evaluation runs. The tf$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()
:
loss < tf$losses$mean_squared_error(labels, predictions)
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 < list(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())
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 estimator_spec
object containing mode
, predictions_dict
, loss
, and train_op
:
model_fn < function(features, labels, mode, params, config) {
# Connect the first hidden layer to input layer
first_hidden_layer < tf$layers$dense(features, 10L, activation = tf$nn$relu)
# Connect the second hidden layer to first hidden layer with relu
second_hidden_layer < tf$layers$dense(first_hidden_layer, 10L, activation = tf$nn$relu)
# Connect the output layer to second hidden layer (no activation fn)
output_layer < tf$layers$dense(second_hidden_layer, 1L)
# Reshape output layer to 1dim Tensor to return predictions
predictions < tf$reshape(output_layer, list(1L))
predictions_list < list(ages = predictions)
# Calculate loss using mean squared error
loss < tf$losses$mean_squared_error(labels, predictions)
eval_metric_ops < list(rmse = tf$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())
return(estimator_spec(
mode = mode,
predictions = predictions_list,
loss = loss,
train_op = train_op,
eval_metric_ops = eval_metric_ops
))
}
model_params < list(learning_rate = 0.001)
model < estimator(model_fn, params = model_params)
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.
The following code fits the neural network to the training data and evaluates the model performance based on the eval_metric_ops
that we have defined: