org.tribuo.interop.tensorflow

Class TensorflowTrainer<T extends Output<T>>

java.lang.Object

org.tribuo.interop.tensorflow.TensorflowTrainer<T>

All Implemented Interfaces:

com.oracle.labs.mlrg.olcut.config.Configurable, com.oracle.labs.mlrg.olcut.provenance.Provenancable<TrainerProvenance>, Trainer<T>

public final class TensorflowTrainer<T extends Output<T>>
extends Object
implements Trainer<T>

Trainer for Tensorflow. Expects the underlying Tensorflow graph to have specific placeholders and targets listed below.

• TensorflowModel.INPUT_NAME - the input minibatch.
• TensorflowModel.OUTPUT_NAME - the predicted output.
• TARGET - the output to predict.
• TRAIN - the train function to run (usually a single step of SGD).
• TRAINING_LOSS - the loss tensor to extract for logging.
• EPOCH - the current epoch number, used for gradient scaling.
• IS_TRAINING - a boolean placeholder to turn on dropout or other training specific functionality.
• INIT - the function to initialise the graph.

This trainer only works with graphs setup for minibatches. To recover single example training just use a batch size of 1.

This trainer uses the serialisation functionality in TensorflowUtil, as opposed to a SavedModel or a checkpoint.

N.B. Tensorflow support is experimental and may change without a major version bump.

Nested Class Summary

Nested Classes

Modifier and Type	Class and Description
static class	TensorflowTrainer.TensorflowTrainerProvenance

Field Summary

Fields

Modifier and Type	Field and Description
static String	EPOCH
static String	INIT
static String	IS_TRAINING
static String	TARGET
static String	TRAIN
static String	TRAINING_LOSS

Fields inherited from interface org.tribuo.Trainer

DEFAULT_SEED

Constructor Summary

Constructors

Constructor and Description
TensorflowTrainer(byte[] graphDef, ExampleTransformer<T> exampleTransformer, OutputTransformer<T> outputTransformer, int minibatchSize, int epochs, int testBatchSize) Constructs a Trainer for a tensorflow graph.
TensorflowTrainer(Path graphPath, ExampleTransformer<T> exampleTransformer, OutputTransformer<T> outputTransformer, int minibatchSize, int epochs, int testBatchSize) Constructs a Trainer for a tensorflow graph.

Method Summary

Modifier and Type	Method and Description
int	getInvocationCount() The number of times this trainer instance has had it's train method invoked.
TrainerProvenance	getProvenance()
void	postConfig() Used by the OLCUT configuration system, and should not be called by external code.
String	toString()
Model<T>	train(Dataset<T> examples, Map<String,com.oracle.labs.mlrg.olcut.provenance.Provenance> runProvenance) Trains a predictive model using the examples in the given data set.

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, wait, wait, wait

Methods inherited from interface org.tribuo.Trainer

train

Field Detail

TARGET

public static final String TARGET

See Also:

Constant Field Values

TRAIN

public static final String TRAIN

See Also:

Constant Field Values

TRAINING_LOSS

public static final String TRAINING_LOSS

See Also:

Constant Field Values

EPOCH

public static final String EPOCH

See Also:

Constant Field Values

IS_TRAINING

public static final String IS_TRAINING

See Also:

Constant Field Values

INIT

public static final String INIT

See Also:

Constant Field Values

Constructor Detail

TensorflowTrainer

public TensorflowTrainer(Path graphPath,
                         ExampleTransformer<T> exampleTransformer,
                         OutputTransformer<T> outputTransformer,
                         int minibatchSize,
                         int epochs,
                         int testBatchSize)
                  throws IOException

Constructs a Trainer for a tensorflow graph.

Parameters:

graphPath - The path to the graph protobuf. Must have the targets and placeholders specified above.

exampleTransformer - The example transformer to convert a Tribuo Example into a Tensor.

outputTransformer - The output transformer to convert a Tribuo Output into a Tensor and back. This encodes the output type.

minibatchSize - The minibatch size to use in training.

epochs - The number of SGD epochs to run.

testBatchSize - The minibatch size to use at test time.

Throws:

IOException - If the graphPath is invalid or failed to load.

TensorflowTrainer

public TensorflowTrainer(byte[] graphDef,
                         ExampleTransformer<T> exampleTransformer,
                         OutputTransformer<T> outputTransformer,
                         int minibatchSize,
                         int epochs,
                         int testBatchSize)

Constructs a Trainer for a tensorflow graph.

Parameters:

graphDef - The graph definition as a byte array. Must have the targets and placeholders specified above.

exampleTransformer - The example transformer to convert a Tribuo Example into a Tensor.

outputTransformer - The output transformer to convert a Tribuo Output into a Tensor and back. This encodes the output type.

minibatchSize - The minibatch size to use in training.

epochs - The number of SGD epochs to run.

testBatchSize - The minibatch size to use at test time.

Method Detail

postConfig

public void postConfig()
                throws IOException

Used by the OLCUT configuration system, and should not be called by external code.

Specified by:

postConfig in interface com.oracle.labs.mlrg.olcut.config.Configurable

Throws:

IOException

train

public Model<T> train(Dataset<T> examples,
                      Map<String,com.oracle.labs.mlrg.olcut.provenance.Provenance> runProvenance)

Description copied from interface: Trainer

Trains a predictive model using the examples in the given data set.

Specified by:

train in interface Trainer<T extends Output<T>>

Parameters:

examples - the data set containing the examples.

runProvenance - Training run specific provenance (e.g., fold number).

Returns:

a predictive model that can be used to generate predictions for new examples.

toString

public String toString()

Overrides:

toString in class Object

getInvocationCount

public int getInvocationCount()

Description copied from interface: Trainer

The number of times this trainer instance has had it's train method invoked.

This is used to determine how many times the trainer's RNG has been accessed to ensure replicability in the random number stream.

Specified by:

getInvocationCount in interface Trainer<T extends Output<T>>

Returns:

The number of train invocations.

getProvenance

public TrainerProvenance getProvenance()

Specified by:

getProvenance in interface com.oracle.labs.mlrg.olcut.provenance.Provenancable<TrainerProvenance>