org.deeplearning4j.nn.transferlearning

Class TransferLearning.Builder

java.lang.Object

org.deeplearning4j.nn.transferlearning.TransferLearning.Builder

Enclosing class:

TransferLearning

public static class TransferLearning.Builder
extends java.lang.Object

Constructor Summary

Constructors

Constructor and Description
Builder(MultiLayerNetwork origModel) Multilayer Network to tweak for transfer learning

Method Summary

Modifier and Type	Method and Description
TransferLearning.Builder	addLayer(Layer layer) Add layers to the net Required if layers are removed.
MultiLayerNetwork	build() Returns a model with the fine tune configuration and specified architecture changes.
TransferLearning.Builder	fineTuneConfiguration(FineTuneConfiguration finetuneConfiguration) Fine tune configurations specified will overwrite the existing configuration if any Usage example: specify a learning rate will set specified learning rate on all layers Refer to the fineTuneConfiguration class for more details
TransferLearning.Builder	nOutReplace(int layerNum, int nOut, Distribution dist) Modify the architecture of a layer by changing nOut Note this will also affect the layer that follows the layer specified, unless it is the output layer
TransferLearning.Builder	nOutReplace(int layerNum, int nOut, Distribution dist, Distribution distNext) Modify the architecture of a layer by changing nOut Note this will also affect the layer that follows the layer specified, unless it is the output layer Can specify different weight init schemes for the specified layer and the layer that follows it.
TransferLearning.Builder	nOutReplace(int layerNum, int nOut, Distribution dist, WeightInit schemeNext) Modify the architecture of a layer by changing nOut Note this will also affect the layer that follows the layer specified, unless it is the output layer Can specify different weight init schemes for the specified layer and the layer that follows it.
TransferLearning.Builder	nOutReplace(int layerNum, int nOut, WeightInit scheme) Modify the architecture of a layer by changing nOut Note this will also affect the layer that follows the layer specified, unless it is the output layer
TransferLearning.Builder	nOutReplace(int layerNum, int nOut, WeightInit scheme, Distribution distNext) Modify the architecture of a layer by changing nOut Note this will also affect the layer that follows the layer specified, unless it is the output layer Can specify different weight init schemes for the specified layer and the layer that follows it.
TransferLearning.Builder	nOutReplace(int layerNum, int nOut, WeightInit scheme, WeightInit schemeNext) Modify the architecture of a layer by changing nOut Note this will also affect the layer that follows the layer specified, unless it is the output layer Can specify different weight init schemes for the specified layer and the layer that follows it.
TransferLearning.Builder	removeLayersFromOutput(int layerNum) Remove last "n" layers of the net At least an output layer must be added back in
TransferLearning.Builder	removeOutputLayer() Helper method to remove the outputLayer of the net.
TransferLearning.Builder	setFeatureExtractor(int layerNum) Specify a layer to set as a "feature extractor" The specified layer and the layers preceding it will be "frozen" with parameters staying constant
TransferLearning.Builder	setInputPreProcessor(int layer, InputPreProcessor processor) Specify the preprocessor for the added layers for cases where they cannot be inferred automatically.

Methods inherited from class java.lang.Object

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

Constructor Detail

Builder

public Builder(MultiLayerNetwork origModel)

Multilayer Network to tweak for transfer learning

Parameters:

origModel -

Method Detail

fineTuneConfiguration

public TransferLearning.Builder fineTuneConfiguration(FineTuneConfiguration finetuneConfiguration)

Fine tune configurations specified will overwrite the existing configuration if any Usage example: specify a learning rate will set specified learning rate on all layers Refer to the fineTuneConfiguration class for more details

Parameters:

finetuneConfiguration -

Returns:

Builder

setFeatureExtractor

public TransferLearning.Builder setFeatureExtractor(int layerNum)

Specify a layer to set as a "feature extractor" The specified layer and the layers preceding it will be "frozen" with parameters staying constant

Parameters:

layerNum -

Returns:

Builder

nOutReplace

public TransferLearning.Builder nOutReplace(int layerNum,
                                            int nOut,
                                            WeightInit scheme)

Modify the architecture of a layer by changing nOut Note this will also affect the layer that follows the layer specified, unless it is the output layer

Parameters:

layerNum - The index of the layer to change nOut of

nOut - Value of nOut to change to

scheme - Weight Init scheme to use for params in layernum and layernum+1

Returns:

Builder

nOutReplace

public TransferLearning.Builder nOutReplace(int layerNum,
                                            int nOut,
                                            Distribution dist)

Modify the architecture of a layer by changing nOut Note this will also affect the layer that follows the layer specified, unless it is the output layer

Parameters:

layerNum - The index of the layer to change nOut of

nOut - Value of nOut to change to

dist - Distribution to use in conjunction with weight init DISTRIBUTION for params in layernum and layernum+1

Returns:

Builder

See Also:

DISTRIBUTION

nOutReplace

public TransferLearning.Builder nOutReplace(int layerNum,
                                            int nOut,
                                            WeightInit scheme,
                                            WeightInit schemeNext)

Modify the architecture of a layer by changing nOut Note this will also affect the layer that follows the layer specified, unless it is the output layer Can specify different weight init schemes for the specified layer and the layer that follows it.

Parameters:

layerNum - The index of the layer to change nOut of

nOut - Value of nOut to change to

scheme - Weight Init scheme to use for params in the layerNum

schemeNext - Weight Init scheme to use for params in the layerNum+1

Returns:

Builder

nOutReplace

public TransferLearning.Builder nOutReplace(int layerNum,
                                            int nOut,
                                            Distribution dist,
                                            Distribution distNext)

Modify the architecture of a layer by changing nOut Note this will also affect the layer that follows the layer specified, unless it is the output layer Can specify different weight init schemes for the specified layer and the layer that follows it.

Parameters:

layerNum - The index of the layer to change nOut of

nOut - Value of nOut to change to

dist - Distribution to use for params in the layerNum

distNext - Distribution to use for parmas in layerNum+1

Returns:

Builder

See Also:

DISTRIBUTION

nOutReplace

public TransferLearning.Builder nOutReplace(int layerNum,
                                            int nOut,
                                            WeightInit scheme,
                                            Distribution distNext)

Modify the architecture of a layer by changing nOut Note this will also affect the layer that follows the layer specified, unless it is the output layer Can specify different weight init schemes for the specified layer and the layer that follows it.

Parameters:

layerNum - The index of the layer to change nOut of

nOut - Value of nOut to change to

scheme - Weight init scheme to use for params in layerNum

distNext - Distribution to use for parmas in layerNum+1

Returns:

Builder

See Also:

DISTRIBUTION

nOutReplace

public TransferLearning.Builder nOutReplace(int layerNum,
                                            int nOut,
                                            Distribution dist,
                                            WeightInit schemeNext)

Modify the architecture of a layer by changing nOut Note this will also affect the layer that follows the layer specified, unless it is the output layer Can specify different weight init schemes for the specified layer and the layer that follows it.

Parameters:

layerNum - The index of the layer to change nOut of

nOut - Value of nOut to change to

dist - Distribution to use for parmas in layerNum

schemeNext - Weight init scheme to use for params in layerNum+1

Returns:

Builder

See Also:

DISTRIBUTION

removeOutputLayer

public TransferLearning.Builder removeOutputLayer()

Helper method to remove the outputLayer of the net. Only one of the two - removeOutputLayer() or removeLayersFromOutput(layerNum) - can be specified When removing layers at the very least an output layer should be added with .addLayer(...)

Returns:

Builder

removeLayersFromOutput

public TransferLearning.Builder removeLayersFromOutput(int layerNum)

Remove last "n" layers of the net At least an output layer must be added back in

Parameters:

layerNum - number of layers to remove

Returns:

Builder

addLayer

public TransferLearning.Builder addLayer(Layer layer)

Add layers to the net Required if layers are removed. Can be called multiple times and layers will be added in the order with which they were called. At the very least an outputLayer must be added (output layer should be added last - as per the note on order) Learning configs (like updaters, learning rate etc) specified with the layer here will be honored

Parameters:

layer - layer conf to add (similar to the NeuralNetConfiguration .list().layer(...)

Returns:

Builder

setInputPreProcessor

public TransferLearning.Builder setInputPreProcessor(int layer,
                                                     InputPreProcessor processor)

Specify the preprocessor for the added layers for cases where they cannot be inferred automatically.

Parameters:

processor - to be used on the data

Returns:

Builder

build

public MultiLayerNetwork build()

Returns a model with the fine tune configuration and specified architecture changes. .init() need not be called. Can be directly fit.

Returns:

MultiLayerNetwork