Type SdcaModel
Namespace tensorflow.contrib.linear_optimizer
Parent PythonObjectContainer
Interfaces ISdcaModel
Stochastic dual coordinate ascent solver for linear models. Loss functions supported: * Binary logistic loss
* Squared loss
* Hinge loss
* Smooth hinge loss
* Poisson log loss This class defines an optimizer API to train a linear model. ### Usage
In the training program you will just have to run the returned Op from
minimize().
Show Example
# Create a solver with the desired parameters.
lr = tf.contrib.linear_optimizer.SdcaModel(examples, variables, options)
min_op = lr.minimize()
opt_op = lr.update_weights(min_op) predictions = lr.predictions(examples)
# Primal loss + L1 loss + L2 loss.
regularized_loss = lr.regularized_loss(examples)
# Primal loss only
unregularized_loss = lr.unregularized_loss(examples) examples: {
sparse_features: list of SparseFeatureColumn.
dense_features: list of dense tensors of type float32.
example_labels: a tensor of type float32 and shape [Num examples]
example_weights: a tensor of type float32 and shape [Num examples]
example_ids: a tensor of type string and shape [Num examples]
}
variables: {
sparse_features_weights: list of tensors of shape [vocab size]
dense_features_weights: list of tensors of shape [dense_feature_dimension]
}
options: {
symmetric_l1_regularization: 0.0
symmetric_l2_regularization: 1.0
loss_type: "logistic_loss"
num_loss_partitions: 1 (Optional, with default value of 1. Number of
partitions of the global loss function, 1 means single machine solver,
and >1 when we have more than one optimizer working concurrently.)
num_table_shards: 1 (Optional, with default value of 1. Number of shards
of the internal state table, typically set to match the number of
parameter servers for large data sets.
}
Methods
Properties
Public instance methods
object predictions(IDictionary<string, object> examples)
Add operations to compute predictions by the model. If logistic_loss is being used, predicted probabilities are returned.
If poisson_loss is being used, predictions are exponentiated.
Otherwise, (raw) linear predictions (w*x) are returned.
Parameters
-
IDictionary<string, object>examples - Examples to compute predictions on.
Returns
-
object - An Operation that computes the predictions for examples.
object regularized_loss(IDictionary<string, object> examples)
Add operations to compute the loss with regularization loss included.
Parameters
-
IDictionary<string, object>examples - Examples to compute loss on.
Returns
-
object - An Operation that computes mean (regularized) loss for given set of examples.
object unregularized_loss(IDictionary<string, object> examples)
Add operations to compute the loss (without the regularization loss).
Parameters
-
IDictionary<string, object>examples - Examples to compute unregularized loss on.
Returns
-
object - An Operation that computes mean (unregularized) loss for given set of examples.