activations

Compute deep learning network layer activations

Description

You can extract features using a trained deep learning network on either a CPU or GPU. Using a GPU requires Parallel Computing Toolbox™ and a CUDA® enabled NVIDIA® GPU with compute capability 3.0 or higher. Specify the hardware requirements using the ExecutionEnvironment name-value pair argument.

features = activations(net,imds,layer) returns network activations for a specific layer using the trained network net and the image data in the image datastore imds.

features = activations(net,ds,layer) returns network activations using the data in the datastore ds. For networks with multiple inputs, use this syntax with a combined or transformed datastore object.

features = activations(net,X,layer) returns network activations using the image data in the numeric array X.

features = activations(net,sequences,layer) returns network activations for an LSTM or BiLSTM network, where sequences contains sequence or time series predictors.

features = activations(net,tbl,layer) returns network activations using the data in the table of images tbl.

example

features = activations(___,Name,Value) returns network activations with additional options specified by one or more name-value pair arguments. For example, 'OutputAs','rows' specifies the activation output format as 'rows'. Specify name-value pair arguments after all other input arguments.

Examples

collapse all

This example shows how to extract learned image features from a pretrained convolutional neural network, and use those features to train an image classifier. Feature extraction is the easiest and fastest way use the representational power of pretrained deep networks. For example, you can train a support vector machine (SVM) using fitcecoc (Statistics and Machine Learning Toolbox™) on the extracted features. Because feature extraction only requires a single pass through the data, it is a good starting point if you do not have a GPU to accelerate network training with.

Load Data

Unzip and load the sample images as an image datastore. imageDatastore automatically labels the images based on folder names and stores the data as an ImageDatastore object. An image datastore lets you store large image data, including data that does not fit in memory. Split the data into 70% training and 30% test data.

unzip('MerchData.zip');

imds = imageDatastore('MerchData', ...
    'IncludeSubfolders',true, ...
    'LabelSource','foldernames');

[imdsTrain,imdsTest] = splitEachLabel(imds,0.7,'randomized');

There are now 55 training images and 20 validation images in this very small data set. Display some sample images.

numImagesTrain = numel(imdsTrain.Labels);
idx = randperm(numImagesTrain,16);

for i = 1:16
    I{i} = readimage(imdsTrain,idx(i));
end

figure
imshow(imtile(I))

Load Pretrained Network

Load a pretrained AlexNet network. If the Deep Learning Toolbox Model for AlexNet Network support package is not installed, then the software provides a download link. AlexNet is trained on more than a million images and can classify images into 1000 object categories. For example, keyboard, mouse, pencil, and many animals. As a result, the model has learned rich feature representations for a wide range of images.

net = alexnet;

Display the network architecture. The network has five convolutional layers and three fully connected layers.

net.Layers
ans = 
  25x1 Layer array with layers:

     1   'data'     Image Input                   227x227x3 images with 'zerocenter' normalization
     2   'conv1'    Convolution                   96 11x11x3 convolutions with stride [4  4] and padding [0  0  0  0]
     3   'relu1'    ReLU                          ReLU
     4   'norm1'    Cross Channel Normalization   cross channel normalization with 5 channels per element
     5   'pool1'    Max Pooling                   3x3 max pooling with stride [2  2] and padding [0  0  0  0]
     6   'conv2'    Grouped Convolution           2 groups of 128 5x5x48 convolutions with stride [1  1] and padding [2  2  2  2]
     7   'relu2'    ReLU                          ReLU
     8   'norm2'    Cross Channel Normalization   cross channel normalization with 5 channels per element
     9   'pool2'    Max Pooling                   3x3 max pooling with stride [2  2] and padding [0  0  0  0]
    10   'conv3'    Convolution                   384 3x3x256 convolutions with stride [1  1] and padding [1  1  1  1]
    11   'relu3'    ReLU                          ReLU
    12   'conv4'    Grouped Convolution           2 groups of 192 3x3x192 convolutions with stride [1  1] and padding [1  1  1  1]
    13   'relu4'    ReLU                          ReLU
    14   'conv5'    Grouped Convolution           2 groups of 128 3x3x192 convolutions with stride [1  1] and padding [1  1  1  1]
    15   'relu5'    ReLU                          ReLU
    16   'pool5'    Max Pooling                   3x3 max pooling with stride [2  2] and padding [0  0  0  0]
    17   'fc6'      Fully Connected               4096 fully connected layer
    18   'relu6'    ReLU                          ReLU
    19   'drop6'    Dropout                       50% dropout
    20   'fc7'      Fully Connected               4096 fully connected layer
    21   'relu7'    ReLU                          ReLU
    22   'drop7'    Dropout                       50% dropout
    23   'fc8'      Fully Connected               1000 fully connected layer
    24   'prob'     Softmax                       softmax
    25   'output'   Classification Output         crossentropyex with 'tench' and 999 other classes

The first layer, the image input layer, requires input images of size 227-by-227-by-3, where 3 is the number of color channels.

inputSize = net.Layers(1).InputSize
inputSize = 1×3

   227   227     3

Extract Image Features

The network constructs a hierarchical representation of input images. Deeper layers contain higher-level features, constructed using the lower-level features of earlier layers. To get the feature representations of the training and test images, use activations on the fully connected layer 'fc7'. To get a lower-level representation of the images, use an earlier layer in the network.

The network requires input images of size 227-by-227-by-3, but the images in the image datastores have different sizes. To automatically resize the training and test images before they are input to the network, create augmented image datastores, specify the desired image size, and use these datastores as input arguments to activations.

augimdsTrain = augmentedImageDatastore(inputSize(1:2),imdsTrain);
augimdsTest = augmentedImageDatastore(inputSize(1:2),imdsTest);

layer = 'fc7';
featuresTrain = activations(net,augimdsTrain,layer,'OutputAs','rows');
featuresTest = activations(net,augimdsTest,layer,'OutputAs','rows');

Extract the class labels from the training and test data.

YTrain = imdsTrain.Labels;
YTest = imdsTest.Labels;

Fit Image Classifier

Use the features extracted from the training images as predictor variables and fit a multiclass support vector machine (SVM) using fitcecoc (Statistics and Machine Learning Toolbox).

mdl = fitcecoc(featuresTrain,YTrain);

Classify Test Images

Classify the test images using the trained SVM model and the features extracted from the test images.

YPred = predict(mdl,featuresTest);

Display four sample test images with their predicted labels.

idx = [1 5 10 15];
figure
for i = 1:numel(idx)
    subplot(2,2,i)
    I = readimage(imdsTest,idx(i));
    label = YPred(idx(i));
    
    imshow(I)
    title(label)
end

Calculate the classification accuracy on the test set. Accuracy is the fraction of labels that the network predicts correctly.

accuracy = mean(YPred == YTest)
accuracy = 1

This SVM has high accuracy. If the accuracy is not high enough using feature extraction, then try transfer learning instead.

Input Arguments

collapse all

Trained network, specified as a SeriesNetwork or a DAGNetwork object. You can get a trained network by importing a pretrained network (for example, by using the googlenet function) or by training your own network using trainNetwork.

Image datastore, specified as an ImageDatastore object.

ImageDatastore allows batch reading of JPG or PNG image files using prefetching. If you use a custom function for reading the images, then ImageDatastore does not prefetch.

Tip

Use augmentedImageDatastore for efficient preprocessing of images for deep learning including image resizing.

Do not use the readFcn option of imageDatastore as this option is usually significantly slower.

Datastore of input data.

For networks with a single input, ds can return either:

  • a single image

  • a cell array of images

  • a table, where the first column contains images

For networks with multiple inputs, ds must be a combined or transformed datastore that returns a cell array with numInputs columns containing the input data, where numInputs is the number of network inputs. The ith element of the cell array corresponds to the input net.InputNames(i).

For more information, see Datastores for Deep Learning.

Image data, specified as a numeric array. The size of the array depends on the type of image input:

InputDescription
2-D imagesA h-by-w-by-c-by-N numeric array, where h, w, and c are the height, width, and number of channels of the images, respectively, and N is the number of images.
3-D imagesA h-by-w-by-d-by-c-by-N numeric array, where h, w, d, and c are the height, width, depth, and number of channels of the images, respectively, and N is the number of images.

If the array contains NaNs, then they are propagated through the network.

For image input, if the 'OutputAs' option is 'channels', then the images in the input data X can be larger than the input size of the image input layer of the network. For other output formats, the images in X must have the same size as the input size of the image input layer of the network.

Sequence or time series data, specified as an N-by-1 cell array of numeric arrays, where N is the number of observations, a numeric array representing a single sequence, or a datastore.

For cell array or numeric array input, the dimensions of the numeric arrays containing the sequences depend on the type of data.

InputDescription
Vector sequencesc-by-s matrices, where c is the number of features of the sequences and s is the sequence length.
2-D image sequencesh-by-w-by-c-by-s arrays, where h, w, and c correspond to the height, width, and number of channels of the images, respectively, and s is the sequence length.
3-D image sequencesh-by-w-by-d-by-c-by-s, where h, w, d, and c correspond to the height, width, depth, and number of channels of the 3-D images, respectively, and s is the sequence length.

For datastore input, the datastore must return data as a cell array of sequences or a table whose first column contains sequences. The dimensions of the sequence data must correspond to the table above.

Table of images containing the input data in the first column. Each row in the table corresponds to an observation. The table contains absolute or relative file paths to an image, specified as a character vector or images specified as a numeric array.

Data Types: table

Layer to extract features from, specified as a numeric index or a character vector.

To compute the activations of a SeriesNetwork object, specify the layer using its numeric index, or as a character vector corresponding to the layer name.

To compute the activations of a DAGNetwork object, specify the layer as the character vector corresponding to the layer name. If the layer has multiple outputs, specify the layer and output as the layer name, followed by the character “/”, followed by the name of the layer output. That is, layer is of the form 'layerName/outputName'.

Example: 3

Example: 'conv1'

Example: 'mpool/out'

Name-Value Pair Arguments

Specify optional comma-separated pairs of Name,Value arguments. Name is the argument name and Value is the corresponding value. Name must appear inside quotes. You can specify several name and value pair arguments in any order as Name1,Value1,...,NameN,ValueN.

Example: activations(net,X,layer,'OutputAs','rows')

Format of output activations, specified as the comma-separated pair consisting of 'OutputAs' and either 'channels', 'rows', or 'columns'. For descriptions of the different output formats, see features.

For image input, if the 'OutputAs' option is 'channels', then the images in the input data X can be larger than the input size of the image input layer of the network. For other output formats, the images in X must have the same size as the input size of the image input layer of the network.

Example: 'OutputAs','rows'

Size of mini-batches to use for prediction, specified as a positive integer. Larger mini-batch sizes require more memory, but can lead to faster predictions.

Example: 'MiniBatchSize',256

Option to pad, truncate, or split input sequences, specified as one of the following:

  • 'longest' — Pad sequences in each mini-batch to have the same length as the longest sequence. This option does not discard any data, though padding can introduce noise to the network.

  • 'shortest' — Truncate sequences in each mini-batch to have the same length as the shortest sequence. This option ensures that no padding is added, at the cost of discarding data.

  • Positive integer — For each mini-batch, pad the sequences to the nearest multiple of the specified length that is greater than the longest sequence length in the mini-batch, and then split the sequences into smaller sequences of the specified length. If splitting occurs, then the software creates extra mini-batches. Use this option if the full sequences do not fit in memory. Alternatively, try reducing the number of sequences per mini-batch by setting the 'MiniBatchSize' option to a lower value.

To learn more about the effect of padding, truncating, and splitting the input sequences, see Sequence Padding, Truncation, and Splitting.

Example: 'SequenceLength','shortest'

Value by which to pad input sequences, specified as a scalar. The option is valid only when SequenceLength is 'longest' or a positive integer. Do not pad sequences with NaN, because doing so can propagate errors throughout the network.

Example: 'SequencePaddingValue',-1

Direction of padding or truncation, specified as one of the following:

  • 'right' — Pad or truncate sequences on the right. The sequences start at the same time step and the software truncates or adds padding to the end of the sequences.

  • 'left' — Pad or truncate sequences on the left. The software truncates or adds padding to the start of the sequences so that the sequences end at the same time step.

Because LSTM layers process sequence data one time step at a time, when the layer OutputMode property is 'last', any padding in the final time steps can negatively influence the layer output. To pad or truncate sequence data on the left, set the 'SequencePaddingDirection' option to 'left'.

For sequence-to-sequence networks (when the OutputMode property is 'sequence' for each LSTM layer), any padding in the first time steps can negatively influence the predictions for the earlier time steps. To pad or truncate sequence data on the right, set the 'SequencePaddingDirection' option to 'right'.

To learn more about the effect of padding, truncating, and splitting the input sequences, see Sequence Padding, Truncation, and Splitting.

Performance optimization, specified as the comma-separated pair consisting of 'Acceleration' and one of the following:

  • 'auto' — Automatically apply a number of optimizations suitable for the input network and hardware resource.

  • 'mex' — Compile and execute a MEX function. This option is available when using a GPU only. Using a GPU requires Parallel Computing Toolbox and a CUDA enabled NVIDIA GPU with compute capability 3.0 or higher. If Parallel Computing Toolbox or a suitable GPU is not available, then the software returns an error.

  • 'none' — Disable all acceleration.

The default option is 'auto'. If 'auto' is specified, MATLAB® will apply a number of compatible optimizations. If you use the 'auto' option, MATLAB does not ever generate a MEX function.

Using the 'Acceleration' options 'auto' and 'mex' can offer performance benefits, but at the expense of an increased initial run time. Subsequent calls with compatible parameters are faster. Use performance optimization when you plan to call the function multiple times using new input data.

The 'mex' option generates and executes a MEX function based on the network and parameters used in the function call. You can have several MEX functions associated with a single network at one time. Clearing the network variable also clears any MEX functions associated with that network.

The 'mex' option is only available when you are using a GPU. You must have a C/C++ compiler installed and the GPU Coder™ Interface for Deep Learning Libraries support package. Install the support package using the Add-On Explorer in MATLAB. For setup instructions, see MEX Setup (GPU Coder). GPU Coder is not required.

The 'mex' option does not support all layers. For a list of supported layers, see Supported Layers (GPU Coder). Recurrent neural networks (RNNs) containing a sequenceInputLayer are not supported.

You cannot use MATLAB Compiler™ to deploy your network when using the 'mex' option.

Example: 'Acceleration','mex'

Hardware resource, specified as the comma-separated pair consisting of 'ExecutionEnvironment' and one of the following:

  • 'auto' — Use a GPU if one is available; otherwise, use the CPU.

  • 'gpu' — Use the GPU. Using a GPU requires Parallel Computing Toolbox and a CUDA enabled NVIDIA GPU with compute capability 3.0 or higher. If Parallel Computing Toolbox or a suitable GPU is not available, then the software returns an error.

  • 'cpu' — Use the CPU.

Example: 'ExecutionEnvironment','cpu'

Output Arguments

collapse all

Activations from the network layer, returned as a numeric array or a cell array of numeric arrays. The format of features depends on the type of input data, the type of layer output, and the 'OutputAs' option.

Image Input

If net has an image input layer, then features is a numeric array.

'OutputAs'features
'channels'

For 2-D image data, features is an h-by-w-by-c-by-n array, where h, w, and c are the height, width, and number of channels for the output of the chosen layer, respectively, and n is the number of images. In this case, features(:,:,:,i) contains the activations for the ith image.

For 3-D image data, features is an h-by-w-by-d-by-c-by-n array, where h, w, d, and c are the height, width, depth, and number of channels for the output of the chosen layer, respectively, and n is the number of images. In this case, features(:,:,:,:,i) contains the activations for the ith image.

'rows'n-by-m matrix, where n is the number of observations, and m is the number of output elements from the chosen layer. In this case, features(i,:) contains the activations for the ith image.
'columns'm-by-n matrix, where m is the number of output elements from the chosen layer, and n is the number of observations. In this case, features(:,i) contains the activations for the ith image.

Sequence Input

If net has a sequence input layer, and layer has sequence output (for example, LSTM layers with output mode 'sequence'), then features is a cell array. In this case, the 'OutputAs' option must be 'channels'.

'OutputAs'features
'channels'

For vector sequence input, features is a n-by-1 cell array, of c-by-s matrices, where n is the number of sequences, c is the number of features in the sequence, and s is the sequence length.

For 2-D image sequence input, features is a n-by-1 cell array, of h-by-w-by-c-by-s matrices, where n is the number of sequences, h, w, and c are the height, width, and the number of channels of the images, respectively, and s is the sequence length.

For 3-D image sequence input, features is a n-by-1 cell array, of h-by-w-by-c-by-d-by-s matrices, where n is the number of sequences, h, w, d, and c are the height, width, depth, and the number of channels of the images, respectively, and s is the sequence length.

In these cases, features{i} contains the activations of the ith sequence.

If net has a sequence input layer and layer outputs non-sequence data (for example, an LSTM layer with output mode 'last'), then features is a numeric array.

'OutputAs'features
'channels'

For vector sequence input, features is a c-by-n matrix, where n is the number of sequences and c is the number of features in the sequence.

For 2-D image sequence input, features is a h-by-w-by-c-by-n array, where n is the number of sequences, h, w, and c are the height, width, and the number of channels of the images, respectively.

For 3-D image sequence input, features is a h-by-w-by-c-by-d-by-n array, where n is the number of sequences, h, w, d, and c are the height, width, depth, and the number of channels of the images, respectively.

In these cases, features{i} contains the activations of the ith sequence.

'rows'n-by-m matrix, where n is the number of observations, and m is the number of output elements from the chosen layer. In this case, features(i,:) contains the activations for the ith sequence.
'columns'm-by-n matrix, where m is the number of output elements from the chosen layer, and n is the number of observations. In this case, features(:,i) contains the activations for the ith image.

Algorithms

All functions for deep learning training, prediction, and validation in Deep Learning Toolbox™ perform computations using single-precision, floating-point arithmetic. Functions for deep learning include trainNetwork, predict, classify, and activations. The software uses single-precision arithmetic when you train networks using both CPUs and GPUs.

References

[1] M. Kudo, J. Toyama, and M. Shimbo. "Multidimensional Curve Classification Using Passing-Through Regions." Pattern Recognition Letters. Vol. 20, No. 11–13, pages 1103–1111.

[2] UCI Machine Learning Repository: Japanese Vowels Dataset. https://archive.ics.uci.edu/ml/datasets/Japanese+Vowels

Extended Capabilities

Introduced in R2016a