compressNetworkUsingProjection - Compress neural network using projection - MATLAB (original) (raw)

Compress neural network using projection

Since R2022b

Syntax

Description

The compressNetworkUsingProjection function reduces the number of learnable parameters of layers by performing principal component analysis (PCA) of the neuron activations using a data set representative of the training data and then projects the learnable parameters into the subspace that maintains the highest variance in neuron activations. In some cases, this operation is equivalent to replacing layers with networks containing two or more layers with fewer learnable parameters.

Depending on the network, projection configuration, and code generation libraries used (including library-free code generation), forward passes of a projected deep neural network can be faster when you deploy the network to embedded hardware.

If you prune or quantize your network, then use compression using projection after pruning and before quantization.

[netProjected](#mw%5Ffbeb5bb1-cf3f-4b1f-984d-5d4cf33fd6cb) = compressNetworkUsingProjection([net](#mw%5Ff8176e6f-70cc-4e88-847e-08bd20f06a54%5Fsep%5Fmw%5Fb6163c1c-0e7a-4d18-81a3-6ef3212bec80),[mbq](#mw%5Ff8176e6f-70cc-4e88-847e-08bd20f06a54%5Fsep%5Fmw%5F17bf9c2a-ab93-4ae7-a174-34d6b0dcba2f)) compresses the dlnetwork object net by replacing layers with projected layers. The function compresses layers by performing principal component analysis (PCA) of the neuron activations using the data in theminibatchqueue object mbq and projects learnable parameters into the subspace that maintains the highest variance in neuron activations.This feature requires the Deep Learning Toolbox™ Model Compression Library support package. This support package is a free add-on that you can download using the Add-On Explorer. Alternatively, see Deep Learning Toolbox Model Compression Library.

example

[netProjected](#mw%5Ffbeb5bb1-cf3f-4b1f-984d-5d4cf33fd6cb) = compressNetworkUsingProjection([net](#mw%5Ff8176e6f-70cc-4e88-847e-08bd20f06a54%5Fsep%5Fmw%5Fb6163c1c-0e7a-4d18-81a3-6ef3212bec80),[X1,...,XN](#mw%5Ff8176e6f-70cc-4e88-847e-08bd20f06a54%5Fsep%5Fmw%5F654a70e9-ad70-4e71-8b8a-43a81c789f62)) compresses the network using the data in the dlarray objectsX1,...,XN, where N is the number of network inputs.

[netProjected](#mw%5Ffbeb5bb1-cf3f-4b1f-984d-5d4cf33fd6cb) = compressNetworkUsingProjection([net](#mw%5Ff8176e6f-70cc-4e88-847e-08bd20f06a54%5Fsep%5Fmw%5Fb6163c1c-0e7a-4d18-81a3-6ef3212bec80),[npca](#mw%5F6663b72b-f8f5-4348-8aed-6f6f13191dc8)) compresses the network using the neuronPCA objectnpca. The PCA step can be computationally intensive. If you expect to compress the same network multiple times (for example, when exploring different levels of compression), then you can perform the PCA step up front using a neuronPCA object.

[[netProjected](#mw%5Ffbeb5bb1-cf3f-4b1f-984d-5d4cf33fd6cb), [info](#mw%5F1a3da27d-4b9a-4242-ab2a-e57af4af4657)] = compressNetworkUsingProjection(___) also returns the structure info that contains information about the projected layers, the reduction of learnable parameters, and the explained variance achieved during compression.

[[netProjected](#mw%5Ffbeb5bb1-cf3f-4b1f-984d-5d4cf33fd6cb), [info](#mw%5F1a3da27d-4b9a-4242-ab2a-e57af4af4657)] = compressNetworkUsingProjection(___,[Name=Value](#namevaluepairarguments)) specifies additional options using one or more name-value arguments.

Examples

collapse all

Load the pretrained network in dlnetJapaneseVowels and the training data in JapaneseVowelsTrainData.

load dlnetJapaneseVowels load JapaneseVowelsTrainData

Create a mini-batch queue containing the training data. To create a mini-batch queue from in-memory data, convert the sequences to an array datastore.

adsXTrain = arrayDatastore(XTrain,OutputType="same");

Create the minibatchqueue object.

mbq = minibatchqueue(adsXTrain, ... MiniBatchSize=16, ... MiniBatchFcn=@preprocessMiniBatchPredictors, ... MiniBatchFormat="CTB");

Compress the network.

[netProjected,info] = compressNetworkUsingProjection(net,mbq);

Compressed network has 83.4% fewer learnable parameters. Projection compressed 2 layers: "lstm","fc"

View the network layers.

ans = 4×1 Layer array with layers:

 1   'sequenceinput'   Sequence Input    Sequence input with 12 dimensions
 2   'lstm'            Projected Layer   Projected LSTM with 100 hidden units
 3   'fc'              Projected Layer   Projected fully connected layer with output size 9
 4   'softmax'         Softmax           softmax

View the projected LSTM layer. The LearnablesReduction property shows the proportion of learnables removed in the layer. The Network property contains the neural network that represents the projection.

ans = ProjectedLayer with properties:

               Name: 'lstm'
      OriginalClass: 'nnet.cnn.layer.LSTMLayer'
LearnablesReduction: 0.8408
          InputSize: 12
         OutputSize: 100

Hyperparameters InputProjectorSize: 8 OutputProjectorSize: 7

Learnable Parameters Network: [1×1 dlnetwork]

State Parameters Network: [1×1 dlnetwork]

Network Learnable Parameters Network/lstm/InputWeights 400×8 dlarray Network/lstm/RecurrentWeights 400×7 dlarray Network/lstm/Bias 400×1 dlarray Network/lstm/InputProjector 12×8 dlarray Network/lstm/OutputProjector 100×7 dlarray

Network State Parameters Network/lstm/HiddenState 100×1 dlarray Network/lstm/CellState 100×1 dlarray

Show all properties

Mini-Batch Predictors Preprocessing Function

The preprocessMiniBatchPredictors function preprocesses a mini-batch of predictors by extracting the sequence data from the input cell array and truncating them along the second dimension so that they have the same length.

Note: Do not pad sequence data when doing the PCA step for projection as this can negatively impact the analysis. Instead, truncate mini-batches of data to have the same length or use mini-batches of size 1.

function X = preprocessMiniBatchPredictors(dataX)

X = padsequences(dataX,2,Length="shortest");

end

Input Arguments

collapse all

Neural network, specified as an initialized dlnetwork object.

Mini-batch queue that outputs data for each input of the network, specified as a minibatchqueue object.

The PCA step typically works best when using the full training set. However, any dataset that is representative of the training data distribution suffices. The input data must contain two or more observations and sequences must contain two or more time steps.

Note

Do not pad sequence as this can negatively impact the analysis. Instead, truncate mini-batches of data to have the same length or use mini-batches of size 1.

Input data, specified as a formatted dlarray.

For more information about dlarray formats, see the fmt input argument of dlarray.

The PCA step typically works best when using the full training set. However, any dataset that is representative of the training data distribution suffices. The input data must contain two or more observations and sequences must contain two or more time steps.

Note

Do not pad sequence as this can negatively impact the analysis. Instead, truncate mini-batches of data to have the same length or use mini-batches of size 1.

Neuron principal component analysis, specified as a neuronPCA object.

Name-Value Arguments

collapse all

Specify optional pairs of arguments asName1=Value1,...,NameN=ValueN, where Name is the argument name and Value is the corresponding value. Name-value arguments must appear after other arguments, but the order of the pairs does not matter.

Example: netProjected = compressNetworkUsingProjection(net,mbq,VerbosityLevel="off") compresses the network using projection and disables the command line display.

Data Types: string | cell

Target proportion of neuron activation variance explained by the remaining principal components of each projected layer, specified as a value between 0 (maximum compression) and 1 (project layers with minimal compression).

If you specify the ExplainedVarianceGoal option, then you must not specify the LearnablesReductionGoal option.

Data Types: single | double | int8 | int16 | int32 | int64 | uint8 | uint16 | uint32 | uint64

Target proportion of total number of network learnables to remove, specified as a nonnegative scalar less than or equal to 1.

If you specify the LearnablesReductionGoal option, then you must not specify the ExplainedVarianceGoal option. If you do not specify the LearnablesReductionGoal option, then the function compresses the network using the ExplainedVarianceGoal option.

If LearnablesReductionGoal is greater than the maximum possible reduction in learnables, then the function removes the maximum possible proportion of learnables. Use the neuronPCA function to determine the possible range of reduction in learnables.

If LearnablesReductionGoal is smaller than the maximum possible reduction in learnables, then the function removes at least the proportion of learnables specified by LearnablesReductionGoal. If removing a greater proportion of learnables does not reduce the explained variance, then the function automatically removes a higher proportion of learnables. For example, if you specify a learnables reduction goal of 0.2, and if the explained variance is the same for learnables reductions between 0.2 and0.5, then the function removes 50% of learnables.

Data Types: single | double | int8 | int16 | int32 | int64 | uint8 | uint16 | uint32 | uint64

Verbosity level, specified as one of these values:

Since R2023b

Flag to unpack projected layers, specified as one of these values:

Output Arguments

collapse all

Projected network, returned as a dlnetwork object.

After you compress the network using projection, you can fine-tune the network to help regain predictive accuracy lost by the compression process. For an example, seeCompress Neural Network Using Projection.

Projection information, returned as a structure with these fields:

Tips

Algorithms

collapse all

To compress a deep learning network, you can use projected layers. A projected layer is a type of deep learning layer that enables compression by reducing the number of stored learnable parameters. The layer introduces learnable projector matrices_Q_, replaces multiplications of the form Wx, where W is a learnable matrix, with the multiplication WQQ⊤x, and stores Q and W′=WQ instead of storing W. Projecting x into a lower dimensional space using Q typically requires less memory to store the learnable parameters and can have similarly strong prediction accuracy.

For some types of layers, you can represent a projected layer as a neural network containing two or more layers with fewer learnable parameters. For example, you can represent a projected convolution layer as three convolution layers that perform the input projection, convolution, and the output projection operations independently. When you compress a network using the compressNetworkUsingProjection function, the software replaces layers that support projection with ProjectedLayer objects that contain the equivalent neural network. To replace ProjectedLayer objects in a neural network with the equivalent neural network that represents the projection, use the unpackProjectedLayers function or set the UnpackProjectedLayers option of the compressNetworkUsingProjection function to 1 (true).

The compressNetworkUsingProjection function supports projecting these layers:

The compressNetworkUsingProjection function replaces projectable layers withProjectedLayer objects. A ProjectedLayer object contains information about the projection operation and contains the neural network that represents the projection.

The neural network that represents the projection depends on the type of layer:

References

Extended Capabilities

expand all

The compressNetworkUsingProjection function supports GPU array input with these usage notes and limitations:

For more information, see Run MATLAB Functions on a GPU (Parallel Computing Toolbox).

Version History

Introduced in R2022b

expand all

The compressNetworkUsingProjection function now supports projectingconvolution1dLayer objects.

The output info structure now has the fieldLayerNames that contains the names of the projected layers.

Starting in R2023b, the compressNetworkUsingProjection function replaces LSTM layers with ProjectedLayer objects with a network that contains a single lstmProjectedLayer object. In previous versions, the function replaces LSTM layers with lstmProjectedLayer objects directly.

To reproduce the previous behavior, replace the ProjectedLayer objects with their networks using the unpackProjectedLayers function or set theUnpackProjectedLayers option of thecompressNetworkUsingProjection function to 1 (true).