Classify Text Data Using Custom Training Loop

This example uses:

This example shows how to classify text data using a deep learning bidirectional long short-term memory (BiLSTM) network with a custom training loop.

When training a deep learning network using the trainNetwork function, if trainingOptions does not provide the options you need (for example, a custom learning rate schedule), then you can define your own custom training loop using automatic differentiation. For an example showing how to classify text data using the trainNetwork function, see Classify Text Data Using Deep Learning (Deep Learning Toolbox).

This example trains a network to classify text data with the time-based decay learning rate schedule: for each iteration, the solver uses the learning rate given by $ρ_{t} = \frac{ρ_{0}}{1 + k t}$ , where t is the iteration number, $ρ_{0}$ is the initial learning rate, and k is the decay.

Import Data

Import the factory reports data. This data contains labeled textual descriptions of factory events. To import the text data as strings, specify the text type to be 'string'.

filename = "factoryReports.csv";
data = readtable(filename,'TextType','string');
head(data)

ans=8×5 table
                                 Description                                       Category          Urgency          Resolution         Cost 
    _____________________________________________________________________    ____________________    ________    ____________________    _____

    "Items are occasionally getting stuck in the scanner spools."            "Mechanical Failure"    "Medium"    "Readjust Machine"         45
    "Loud rattling and banging sounds are coming from assembler pistons."    "Mechanical Failure"    "Medium"    "Readjust Machine"         35
    "There are cuts to the power when starting the plant."                   "Electronic Failure"    "High"      "Full Replacement"      16200
    "Fried capacitors in the assembler."                                     "Electronic Failure"    "High"      "Replace Components"      352
    "Mixer tripped the fuses."                                               "Electronic Failure"    "Low"       "Add to Watch List"        55
    "Burst pipe in the constructing agent is spraying coolant."              "Leak"                  "High"      "Replace Components"      371
    "A fuse is blown in the mixer."                                          "Electronic Failure"    "Low"       "Replace Components"      441
    "Things continue to tumble off of the belt."                             "Mechanical Failure"    "Low"       "Readjust Machine"         38

The goal of this example is to classify events by the label in the Category column. To divide the data into classes, convert these labels to categorical.

data.Category = categorical(data.Category);

View the distribution of the classes in the data using a histogram.

figure
histogram(data.Category);
xlabel("Class")
ylabel("Frequency")
title("Class Distribution")

The next step is to partition it into sets for training and validation. Partition the data into a training partition and a held-out partition for validation and testing. Specify the holdout percentage to be 20%.

cvp = cvpartition(data.Category,'Holdout',0.2);
dataTrain = data(training(cvp),:);
dataValidation = data(test(cvp),:);

Extract the text data and labels from the partitioned tables.

textDataTrain = dataTrain.Description;
textDataValidation = dataValidation.Description;
YTrain = dataTrain.Category;
YValidation = dataValidation.Category;

To check that you have imported the data correctly, visualize the training text data using a word cloud.

figure
wordcloud(textDataTrain);
title("Training Data")

View the number of classes.

classes = categories(YTrain);
numClasses = numel(classes)

numClasses = 4

Preprocess Text Data

Create a function that tokenizes and preprocesses the text data. The function preprocessText, listed at the end of the example, performs these steps:

Tokenize the text using tokenizedDocument.
Convert the text to lowercase using lower.
Erase the punctuation using erasePunctuation.

Preprocess the training data and the validation data using the preprocessText function.

documentsTrain = preprocessText(textDataTrain);
documentsValidation = preprocessText(textDataValidation);

View the first few preprocessed training documents.

documentsTrain(1:5)

ans = 
  5×1 tokenizedDocument:

     9 tokens: items are occasionally getting stuck in the scanner spools
    10 tokens: loud rattling and banging sounds are coming from assembler pistons
     5 tokens: fried capacitors in the assembler
     4 tokens: mixer tripped the fuses
     9 tokens: burst pipe in the constructing agent is spraying coolant

Create a single datastore that contains both the documents and the labels by creating arrayDatastore objects, then combining them using the combine function.

dsDocumentsTrain = arrayDatastore(documentsTrain,'OutputType','cell');
dsYTrain = arrayDatastore(YTrain,'OutputType','cell');
dsTrain = combine(dsDocumentsTrain,dsYTrain);

Create a datastore for the validation data using the same steps.

dsDocumentsValidation = arrayDatastore(documentsValidation,'OutputType','cell');
dsYValidation = arrayDatastore(YValidation,'OutputType','cell');
dsValidation = combine(dsDocumentsValidation,dsYValidation);

Create Word Encoding

To input the documents into a BiLSTM network, use a word encoding to convert the documents into sequences of numeric indices.

To create a word encoding, use the wordEncoding function.

enc = wordEncoding(documentsTrain)

enc = 
  wordEncoding with properties:

      NumWords: 421
    Vocabulary: [1×421 string]

Define Network

Define the BiLSTM network architecture. To input sequence data into the network, include a sequence input layer and set the input size to 1. Next, include a word embedding layer of dimension 25 and the same number of words as the word encoding. Next, include a BiLSTM layer and set the number of hidden units to 40. To use the BiLSTM layer for a sequence-to-label classification problem, set the output mode to 'last'. Finally, add a fully connected layer with the same size as the number of classes, and a softmax layer.

inputSize = 1;
embeddingDimension = 25;
numHiddenUnits = 40;

numWords = enc.NumWords;

layers = [
    sequenceInputLayer(inputSize,'Name','in')
    wordEmbeddingLayer(embeddingDimension,numWords,'Name','emb')
    bilstmLayer(numHiddenUnits,'OutputMode','last','Name','bilstm')
    fullyConnectedLayer(numClasses,'Name','fc')
    softmaxLayer('Name','sm')]

layers = 
  5×1 Layer array with layers:

     1   'in'       Sequence Input         Sequence input with 1 dimensions
     2   'emb'      Word Embedding Layer   Word embedding layer with 25 dimensions and 421 unique words
     3   'bilstm'   BiLSTM                 BiLSTM with 40 hidden units
     4   'fc'       Fully Connected        4 fully connected layer
     5   'sm'       Softmax                softmax

Convert the layer array to a layer graph and create a dlnetwork object.

lgraph = layerGraph(layers);
dlnet = dlnetwork(lgraph)

dlnet = 
  dlnetwork with properties:

         Layers: [5×1 nnet.cnn.layer.Layer]
    Connections: [4×2 table]
     Learnables: [6×3 table]
          State: [2×3 table]
     InputNames: {'in'}
    OutputNames: {'sm'}

Define Model Gradients Function

Create the function modelGradients, listed at the end of the example, that takes a dlnetwork object, a mini-batch of input data with corresponding labels, and returns the gradients of the loss with respect to the learnable parameters in the network and the corresponding loss.

Specify Training Options

Train for 30 epochs with a mini-batch size of 16.

numEpochs = 30;
miniBatchSize = 16;

Specify the options for Adam optimization. Specify an initial learn rate of 0.001 with a decay of 0.01, gradient decay factor 0.9, and squared gradient decay factor 0.999.

initialLearnRate = 0.001;
decay = 0.01;
gradientDecayFactor = 0.9;
squaredGradientDecayFactor = 0.999;

Train Model

Train the model using a custom training loop.

Initialize the training progress plot.

figure
lineLossTrain = animatedline('Color',[0.85 0.325 0.098]);

lineLossValidation = animatedline( ...
    'LineStyle','--', ...
    'Marker','o', ...
    'MarkerFaceColor','black');

ylim([0 inf])
xlabel("Iteration")
ylabel("Loss")
grid on

Initialize the parameters for Adam.

trailingAvg = [];
trailingAvgSq = [];

Create a minibatchqueue object that processes and manages the mini-batches of data. For each mini-batch:

Use the custom mini-batch preprocessing function preprocessMiniBatch (defined at the end of this example) to convert documents to sequences and one-hot encode the labels. To pass the word encoding to the mini-batch, create an anonymous function that takes two inputs.
Format the predictors with the dimension labels 'BTC' (batch, time, channel). The minibatchqueue object, by default, converts the data to dlarray objects with underlying type single.
Train on a GPU if one is available. The minibatchqueue object, by default, converts each output to gpuArray if a GPU is available. Using a GPU requires Parallel Computing Toolbox™ and a CUDA® enabled NVIDIA® GPU with compute capability 3.0 or higher.

mbq = minibatchqueue(dsTrain, ...
    'MiniBatchSize',miniBatchSize,...
    'MiniBatchFcn', @(X,Y) preprocessMiniBatch(X,Y,enc), ...
    'MiniBatchFormat',{'BTC',''});

Create a minibatchqueue object for the validation data using the same options and also specify to return partial mini-batches.

mbqValidation = minibatchqueue(dsValidation, ...
    'MiniBatchSize',miniBatchSize, ...
    'MiniBatchFcn', @(X,Y) preprocessMiniBatch(X,Y,enc), ...
    'MiniBatchFormat',{'BTC',''}, ...
    'PartialMiniBatch','return');

Train the network. For each epoch, shuffle the data and loop over mini-batches of data. At the end of each iteration, display the training progress. At the end of each epoch, validate the network using the validation data.

For each mini-batch:

Convert the documents to sequences of integers and one-hot encode the labels.
Convert the data to dlarray objects with underlying type single and specify the dimension labels 'BTC' (batch, time, channel).
For GPU training, convert to gpuArray objects.
Evaluate the model gradients, state, and loss using dlfeval and the modelGradients function and update the network state.
Determine the learning rate for the time-based decay learning rate schedule.
Update the network parameters using the adamupdate function.
Update the training plot.

iteration = 0;
start = tic;

% Loop over epochs.
for epoch = 1:numEpochs
    
    % Shuffle data.
    shuffle(mbq);
    
    % Loop over mini-batches.
    while hasdata(mbq)
        iteration = iteration + 1;
        
        % Read mini-batch of data.
        [dlX, dlY] = next(mbq);
        
        % Evaluate the model gradients, state, and loss using dlfeval and the
        % modelGradients function.
        [gradients,loss] = dlfeval(@modelGradients,dlnet,dlX,dlY);
        
        % Determine learning rate for time-based decay learning rate schedule.
        learnRate = initialLearnRate/(1 + decay*iteration);
        
        % Update the network parameters using the Adam optimizer.
        [dlnet,trailingAvg,trailingAvgSq] = adamupdate(dlnet, gradients, ...
            trailingAvg, trailingAvgSq, iteration, learnRate, ...
            gradientDecayFactor, squaredGradientDecayFactor);
        
        % Display the training progress.
        D = duration(0,0,toc(start),'Format','hh:mm:ss');
        addpoints(lineLossTrain,iteration,loss)
        title("Epoch: " + epoch + ", Elapsed: " + string(D))
        drawnow
        
        % Validate network.
        if iteration == 1 || ~hasdata(mbq)
            % Validation predictions.
            [~,lossValidation] = modelPredictions(dlnet,mbqValidation,classes);
            
            % Update plot.
            addpoints(lineLossValidation,iteration,lossValidation)
            drawnow
        end
    end
end

Test Model

Test the classification accuracy of the model by comparing the predictions on the validation set with the true labels.

Classify the validation data using modelPredictions function, listed at the end of the example.

dlYPred = modelPredictions(dlnet,mbqValidation,classes);
YPred = onehotdecode(dlYPred,classes,1)';

Evaluate the classification accuracy.

accuracy = mean(YPred == YValidation)

accuracy = 0.9167

Predict Using New Data

Classify the event type of three new reports. Create a string array containing the new reports.

reportsNew = [
    "Coolant is pooling underneath sorter."
    "Sorter blows fuses at start up."
    "There are some very loud rattling sounds coming from the assembler."];

Preprocess the text data using the preprocessing steps as the training documents.

documentsNew = preprocessText(reportsNew);
dsNew = arrayDatastore(documentsNew,'OutputType','cell');

Create a minibatchqueue object that processes and manages the mini-batches of data. For each mini-batch:

Use the custom mini-batch preprocessing function preprocessMiniBatchPredictors (defined at the end of this example) to convert documents to sequences. This preprocessing function does not require label data. To pass the word encoding to the mini-batch, create an anonymous function that takes one input only.
Format the predictors with the dimension labels 'BTC' (batch, time, channel). The minibatchqueue object, by default, converts the data to dlarray objects with underlying type single.
To make predictions for all observations, return any partial mini-batches.

mbqNew = minibatchqueue(dsNew, ...
    'MiniBatchSize',miniBatchSize, ...
    'MiniBatchFcn',@(X) preprocessMiniBatchPredictors(X,enc), ...
    'MiniBatchFormat','BTC', ...
    'PartialMiniBatch','return');

Classify the text data using modelPredictions function, listed at the end of the example and find the classes with the highest scores.

dlYPred = modelPredictions(dlnet,mbqNew,classes);
YPred = onehotdecode(dlYPred,classes,1)'

YPred = 3×1 categorical
     Leak 
     Electronic Failure 
     Mechanical Failure

Text Preprocessing Function

The function preprocessText performs these steps:

Tokenize the text using tokenizedDocument.
Convert the text to lowercase using lower.
Erase the punctuation using erasePunctuation.

function documents = preprocessText(textData)

% Tokenize the text.
documents = tokenizedDocument(textData);

% Convert to lowercase.
documents = lower(documents);

% Erase punctuation.
documents = erasePunctuation(documents);

end

Mini-Batch Preprocessing Function

The preprocessMiniBatch function converts a mini-batch of documents to sequences of integers and one-hot encodes label data.

function [X, Y] = preprocessMiniBatch(documentsCell,labelsCell,enc)

% Preprocess predictors.
X = preprocessMiniBatchPredictors(documentsCell,enc);

% Extract labels from cell and concatenate.
Y = cat(1,labelsCell{1:end});

% One-hot encode labels.
Y = onehotencode(Y,2);

% Transpose the encoded labels to match the network output.
Y = Y';

end

Mini-Batch Predictors Preprocessing Function

The preprocessMiniBatchPredictors function converts a mini-batch of documents to sequences of integers.

function X = preprocessMiniBatchPredictors(documentsCell,enc)

% Extract documents from cell and concatenate.
documents = cat(4,documentsCell{1:end});

% Convert documents to sequences of integers.
X = doc2sequence(enc,documents);
X = cat(1,X{:});

end

Model Gradients Function

The modelGradients function takes a dlnetwork object dlnet, a mini-batch of input data dlX with corresponding target labels T and returns the gradients of the loss with respect to the learnable parameters in dlnet, and the loss. To compute the gradients automatically, use the dlgradient function.

function [gradients,loss] = modelGradients(dlnet,dlX,T)

dlYPred = forward(dlnet,dlX);

loss = crossentropy(dlYPred,T);
gradients = dlgradient(loss,dlnet.Learnables);

loss = double(gather(extractdata(loss)));

end

Model Predictions Function

The modelPredictions function takes a dlnetwork object dlnet, a mini-batch queue, and outputs the model predictions by iterating over mini-batches in the queue. To evaluate validation data, this function optionally calcualtes the loss when given a mini-batch queue with two outputs.

function [dlYPred,loss] = modelPredictions(dlnet,mbq,classes)

% Initialize predictions.
numClasses = numel(classes);
outputCast = mbq.OutputCast{1};
dlYPred = dlarray(zeros(numClasses,0,outputCast),'CB');

% Reset mini-batch queue.
reset(mbq);

% For mini-batch queues with two ouputs, also compute the loss.
if mbq.NumOutputs == 1
    
    % Loop over mini-batches.
    while hasdata(mbq)
        
        % Make predictions.
        dlX = next(mbq);
        dlY = predict(dlnet,dlX);
        dlYPred = [dlYPred dlY];
    end
    
else
    % Initialize loss.
    numObservations = 0;
    loss = 0;
    
    % Loop over mini-batches.
    while hasdata(mbq)
        
        % Make predictions.
        [dlX,dlT] = next(mbq);
        dlY = predict(dlnet,dlX);
        dlYPred = [dlYPred dlY];
        
        % Calculate unnormalized loss.
        miniBatchSize = size(dlX,2);
        loss = loss + miniBatchSize * crossentropy(dlY, dlT);
        
        % Count observations.
        numObservations = numObservations + miniBatchSize;
    end
    
    % Normalize loss.
    loss = loss / numObservations;
    
    % Convert to double.
    loss = double(gather(extractdata(loss)));
end

end

Documentation