tf.estimator.experimental.RNNEstimator

View source on GitHub

An Estimator for TensorFlow RNN models with user-specified head.

Inherits From: Estimator

Example:

token_sequence = sequence_categorical_column_with_hash_bucket(...)
token_emb = embedding_column(categorical_column=token_sequence, ...)

estimator = RNNEstimator(
    head=tf.estimator.RegressionHead(),
    sequence_feature_columns=[token_emb],
    units=[32, 16], cell_type='lstm')

# Or with custom RNN cell:
def rnn_cell_fn(_):
  cells = [ tf.keras.layers.LSTMCell(size) for size in [32, 16] ]
  return tf.keras.layers.StackedRNNCells(cells)

estimator = RNNEstimator(
    head=tf.estimator.RegressionHead(),
    sequence_feature_columns=[token_emb],
    rnn_cell_fn=rnn_cell_fn)

# Input builders
def input_fn_train: # returns x, y
  pass
estimator.train(input_fn=input_fn_train, steps=100)

def input_fn_eval: # returns x, y
  pass
metrics = estimator.evaluate(input_fn=input_fn_eval, steps=10)
def input_fn_predict: # returns x, None
  pass
predictions = estimator.predict(input_fn=input_fn_predict)

Input of train and evaluate should have following features, otherwise there will be a KeyError:

  • if the head's weight_column is not None, a feature with key=weight_column whose value is a Tensor.
  • for each column in sequence_feature_columns:
    • a feature with key=column.name whose value is a SparseTensor.
  • for each column in context_feature_columns:
    • if column is a CategoricalColumn, a feature with key=column.name whose value is a SparseTensor.
    • if column is a WeightedCategoricalColumn, two features: the first with key the id column name, the second with key the weight column name. Both features' value must be a SparseTensor.
    • if column is a DenseColumn, a feature with key=column.name whose value is a Tensor.

Loss and predicted output are determined by the specified head.

head A Head instance. This specifies the model's output and loss function to be optimized.
sequence_feature_columns An iterable containing the FeatureColumns that represent sequential input. All items in the set should either be sequence columns (e.g. sequence_numeric_column) or constructed from one (e.g. embedding_column with sequence_categorical_column_* as input).
context_feature_columns An iterable containing the FeatureColumns for contextual input. The data represented by these columns will be replicated and given to the RNN at each timestep. These columns must be instances of classes derived from DenseColumn such as numeric_column, not the sequential variants.
units Iterable of integer number of hidden units per RNN layer. If set, cell_type must also be specified and rnn_cell_fn must be None.
cell_type A class producing a RNN cell or a string specifying the cell type. Supported strings are: 'simple_rnn', 'lstm', and 'gru'. If set, units must also be specified and rnn_cell_fn must be None.
rnn_cell_fn A function that returns a RNN cell instance that will be used to construct the RNN. If set, units and cell_type cannot be set. This is for advanced users who need additional customization beyond units and cell_type. Note that tf.keras.layers.StackedRNNCells is needed for stacked RNNs.
return_sequences A boolean indicating whether to return the last output in the output sequence, or the full sequence.
model_dir Directory to save model parameters, graph and etc. This can also be used to load checkpoints from the directory into a estimator to continue training a previously saved model.
optimizer An instance of tf.Optimizer or string specifying optimizer type. Defaults to Adagrad optimizer.
config RunConfig object to configure the runtime settings.

ValueError If units, cell_type, and rnn_cell_fn are not compatible.

Eager Compatibility

Estimators are not compatible with eager execution.

config

model_dir

model_fn Returns the model_fn which is bound to self.params.
params

Methods

eval_dir

View source

Shows the directory name where evaluation metrics are dumped.

Args
name Name of the evaluation if user needs to run multiple evaluations on different data sets, such as on training data vs test data. Metrics for different evaluations are saved in separate folders, and appear separately in tensorboard.

Returns
A string which is the path of directory contains evaluation metrics.

evaluate

View source

Evaluates the model given evaluation data input_fn.

For each step, calls input_fn, which returns one batch of data. Evaluates until:

Args
input_fn A function that constructs the input data for evaluation. See Premade Estimators for more information. The function should construct and return one of the following:

  • A tf.data.Dataset object: Outputs of Dataset object must be a tuple (features, labels) with same constraints as below.
  • A tuple (features, labels): Where features is a tf.Tensor or a dictionary of string feature name to Tensor and labels is a Tensor or a dictionary of string label name to Tensor. Both features and labels are consumed by model_fn. They should satisfy the expectation of model_fn from inputs.
steps Number of steps for which to evaluate model. If None, evaluates until input_fn raises an end-of-input exception.
hooks List of tf.train.SessionRunHook subclass instances. Used for callbacks inside the evaluation call.
checkpoint_path Path of a specific checkpoint to evaluate. If None, the latest checkpoint in model_dir is used. If there are no checkpoints in model_dir, evaluation is run with newly initialized Variables instead of ones restored from checkpoint.
name Name of the evaluation if user needs to run multiple evaluations on different data sets, such as on training data vs test data. Metrics for different evaluations are saved in separate folders, and appear separately in tensorboard.

Returns
A dict containing the evaluation metrics specified in model_fn keyed by name, as well as an entry global_step which contains the value of the global step for which this evaluation was performed. For canned estimators, the dict contains the loss (mean loss per mini-batch) and the average_loss (mean loss per sample). Canned classifiers also return the accuracy. Canned regressors also return the label/mean and the prediction/mean.

Raises
ValueError If steps <= 0.

experimental_export_all_saved_models

View source

Exports a SavedModel with tf.MetaGraphDefs for each requested mode.

For each mode passed in via the input_receiver_fn_map, this method builds a new graph by calling the input_receiver_fn to obtain feature and label Tensors. Next, this method calls the Estimator's model_fn in the passed mode to generate the model graph based on those features and labels, and restores the given checkpoint (or, lacking that, the most recent checkpoint) into the graph. Only one of the modes is used for saving variables to the SavedModel (order of preference: tf.estimator.ModeKeys.TRAIN, tf.estimator.ModeKeys.EVAL, then tf.estimator.ModeKeys.PREDICT), such that up to three tf.MetaGraphDefs are saved with a single set of variables in a single SavedModel directory.

For the variables and tf.MetaGraphDefs, a timestamped export directory below export_dir_base, and writes a SavedModel into it containing the tf.MetaGraphDef for the given mode and its associated signatures.

For prediction, the exported MetaGraphDef will provide one SignatureDef for each element of the export_outputs dict returned from the model_fn, named using the same keys. One of these keys is always tf.saved_model.signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY, indicating which signature will be served when a serving request does not specify one. For each signature, the outputs are provided by the corresponding tf.estimator.export.ExportOutputs, and the inputs are always the input receivers provided by the serving_input_receiver_fn.

For training and evaluation, the train_op is stored in an extra collection, and loss, metrics, and predictions are included in a SignatureDef for the mode in question.

Extra assets may be written into the SavedModel via the assets_extra argument. This should be a dict, where each key gives a destination path (including the filename) relative to the assets.extra directory. The corresponding value gives the full path of the source file to be copied. For example, the simple case of copying a single file without renaming it is specified as {'my_asset_file.txt': '/path/to/my_asset_file.txt'}.

Args
export_dir_base A string containing a directory in which to create timestamped subdirectories containing exported SavedModels.
input_receiver_fn_map dict of tf.estimator.ModeKeys to input_receiver_fn mappings, where the input_receiver_fn is a function that takes no arguments and returns the appropriate subclass of InputReceiver.
assets_extra A dict specifying how to populate the assets.extra directory within the exported SavedModel, or None if no extra assets are needed.
as_text whether to write the SavedModel proto in text format.
checkpoint_path The checkpoint path to export. If None (the default), the most recent checkpoint found within the model directory is chosen.

Returns
The path to the exported directory as a bytes object.

Raises
ValueError if any input_receiver_fn is None, no export_outputs are provided, or no checkpoint can be found.

export_saved_model

View source

Exports inference graph as a SavedModel into the given dir.

For a detailed guide, see SavedModel from Estimators.

This method builds a new graph by first calling the serving_input_receiver_fn to obtain feature Tensors, and then calling this Estimator's model_fn to generate the model graph based on those features. It restores the given checkpoint (or, lacking that, the most recent checkpoint) into this graph in a fresh session. Finally it creates a timestamped export directory below the given export_dir_base, and writes a SavedModel into it containing a single tf.MetaGraphDef saved from this session.

The exported MetaGraphDef will provide one SignatureDef for each element of the export_outputs dict returned from the model_fn, named using the same keys. One of these keys is always tf.saved_model.signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY, indicating which signature will be served when a serving request does not specify one. For each signature, the outputs are provided by the corresponding tf.estimator.export.ExportOutputs, and the inputs are always the input receivers provided by the serving_input_receiver_fn.

Extra assets may be written into the SavedModel via the assets_extra argument. This