org.deeplearning4j.nn.transferlearning

Class TransferLearningHelper

java.lang.Object

org.deeplearning4j.nn.transferlearning.TransferLearningHelper

public class TransferLearningHelper
extends java.lang.Object

This class is intended for use with the transfer learning API. Often times transfer learning models have "frozen" layers where parameters are held constant during training For ease of training and quick turn around times, the dataset to be trained on can be featurized and saved to disk. Featurizing in this case refers to conducting a forward pass on the network and saving the activations from the output of the frozen layers. During training the forward pass and the backward pass through the frozen layers can be skipped entirely and the "featurized" dataset can be fit with the smaller unfrozen part of the computation graph which allows for quicker iterations. The class internally traverses the computation graph/MLN and builds an instance of the computation graph/MLN that is equivalent to the unfrozen subset.

Constructor Summary

Constructors

Constructor and Description
TransferLearningHelper(ComputationGraph orig) Expects a computation graph where some vertices are frozen
TransferLearningHelper(ComputationGraph orig, java.lang.String... frozenOutputAt) Will modify the given comp graph (in place!) to freeze vertices from input to the vertex specified.
TransferLearningHelper(MultiLayerNetwork orig) Expects a MLN where some layers are frozen
TransferLearningHelper(MultiLayerNetwork orig, int frozenTill) Will modify the given MLN (in place!) to freeze layers (hold params constant during training) specified and below

Method Summary

Modifier and Type	Method and Description
void	errorIfGraphIfMLN()
org.nd4j.linalg.dataset.DataSet	featurize(org.nd4j.linalg.dataset.DataSet input) During training frozen vertices/layers can be treated as "featurizing" the input The forward pass through these frozen layer/vertices can be done in advance and the dataset saved to disk to iterate quickly on the smaller unfrozen part of the model Currently does not support datasets with feature masks
org.nd4j.linalg.dataset.MultiDataSet	featurize(org.nd4j.linalg.dataset.MultiDataSet input) During training frozen vertices/layers can be treated as "featurizing" the input The forward pass through these frozen layer/vertices can be done in advance and the dataset saved to disk to iterate quickly on the smaller unfrozen part of the model Currently does not support datasets with feature masks
void	fitFeaturized(org.nd4j.linalg.dataset.DataSet input)
void	fitFeaturized(org.nd4j.linalg.dataset.api.iterator.DataSetIterator iter)
void	fitFeaturized(org.nd4j.linalg.dataset.MultiDataSet input)
void	fitFeaturized(org.nd4j.linalg.dataset.api.iterator.MultiDataSetIterator iter) Fit from a featurized dataset.
org.nd4j.linalg.api.ndarray.INDArray	outputFromFeaturized(org.nd4j.linalg.api.ndarray.INDArray input) Use to get the output from a featurized input
org.nd4j.linalg.api.ndarray.INDArray[]	outputFromFeaturized(org.nd4j.linalg.api.ndarray.INDArray[] input) Use to get the output from a featurized input
ComputationGraph	unfrozenGraph() Returns the unfrozen subset of the original computation graph as a computation graph Note that with each call to featurizedFit the parameters to the original computation graph are also updated
MultiLayerNetwork	unfrozenMLN() Returns the unfrozen layers of the MultiLayerNetwork as a multilayernetwork Note that with each call to featurizedFit the parameters to the original MLN are also updated

Methods inherited from class java.lang.Object

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

Constructor Detail

TransferLearningHelper

public TransferLearningHelper(ComputationGraph orig,
                              java.lang.String... frozenOutputAt)

Will modify the given comp graph (in place!) to freeze vertices from input to the vertex specified.

Parameters:

orig - Comp graph

frozenOutputAt - vertex to freeze at (hold params constant during training)

TransferLearningHelper

public TransferLearningHelper(ComputationGraph orig)

Expects a computation graph where some vertices are frozen

Parameters:

orig -

TransferLearningHelper

public TransferLearningHelper(MultiLayerNetwork orig,
                              int frozenTill)

Will modify the given MLN (in place!) to freeze layers (hold params constant during training) specified and below

Parameters:

orig - MLN to freeze

frozenTill - integer indicating the index of the layer and below to freeze

TransferLearningHelper

public TransferLearningHelper(MultiLayerNetwork orig)

Expects a MLN where some layers are frozen

Parameters:

orig -

Method Detail

errorIfGraphIfMLN

public void errorIfGraphIfMLN()

unfrozenGraph

public ComputationGraph unfrozenGraph()

Returns the unfrozen subset of the original computation graph as a computation graph Note that with each call to featurizedFit the parameters to the original computation graph are also updated

unfrozenMLN

public MultiLayerNetwork unfrozenMLN()

Returns the unfrozen layers of the MultiLayerNetwork as a multilayernetwork Note that with each call to featurizedFit the parameters to the original MLN are also updated

outputFromFeaturized

public org.nd4j.linalg.api.ndarray.INDArray[] outputFromFeaturized(org.nd4j.linalg.api.ndarray.INDArray[] input)

Use to get the output from a featurized input

Parameters:

input - featurized data

Returns:

output

outputFromFeaturized

public org.nd4j.linalg.api.ndarray.INDArray outputFromFeaturized(org.nd4j.linalg.api.ndarray.INDArray input)

Use to get the output from a featurized input

Parameters:

input - featurized data

Returns:

output

featurize

public org.nd4j.linalg.dataset.MultiDataSet featurize(org.nd4j.linalg.dataset.MultiDataSet input)

During training frozen vertices/layers can be treated as "featurizing" the input The forward pass through these frozen layer/vertices can be done in advance and the dataset saved to disk to iterate quickly on the smaller unfrozen part of the model Currently does not support datasets with feature masks

Parameters:

input - multidataset to feed into the computation graph with frozen layer vertices

Returns:

a multidataset with input features that are the outputs of the frozen layer vertices and the original labels.

featurize

public org.nd4j.linalg.dataset.DataSet featurize(org.nd4j.linalg.dataset.DataSet input)

During training frozen vertices/layers can be treated as "featurizing" the input The forward pass through these frozen layer/vertices can be done in advance and the dataset saved to disk to iterate quickly on the smaller unfrozen part of the model Currently does not support datasets with feature masks

Parameters:

input - multidataset to feed into the computation graph with frozen layer vertices

Returns:

a multidataset with input features that are the outputs of the frozen layer vertices and the original labels.

fitFeaturized

public void fitFeaturized(org.nd4j.linalg.dataset.api.iterator.MultiDataSetIterator iter)

Fit from a featurized dataset. The fit is conducted on an internally instantiated subset model that is representative of the unfrozen part of the original model. After each call on fit the parameters for the original model are updated

Parameters:

iter -

fitFeaturized

public void fitFeaturized(org.nd4j.linalg.dataset.MultiDataSet input)

fitFeaturized

public void fitFeaturized(org.nd4j.linalg.dataset.DataSet input)

fitFeaturized

public void fitFeaturized(org.nd4j.linalg.dataset.api.iterator.DataSetIterator iter)