How to use an existing DL4J trained model to classify new input - dl4j

Question

I have a DL4J LSTM model that generates a binary classification of sequential input. i have trained and tested the model and am happy with the precision/recall. Now I want to use this model to predict the binary classification of new inputs. How do I do this? i.e. how do I give the trained neural network a single Input (file containing the sequence of feature rows) and get the binary classification of this input file.
Here is my original training data set iterator:
SequenceRecordReader trainFeatures = new CSVSequenceRecordReader(0, ","); //skip no header lines
try {
trainFeatures.initialize( new NumberedFileInputSplit(featureBaseDir + "/s_%d.csv", 0,this._modelDefinition.getNB_TRAIN_EXAMPLES()-1));
} catch (IOException e) {
trainFeatures.close();
throw new IOException(String.format("IO error %s. during trainFeatures", e.getMessage()));
} catch (InterruptedException e) {
trainFeatures.close();
throw new IOException(String.format("Interrupted exception error %s. during trainFeatures", e.getMessage()));
}
SequenceRecordReader trainLabels = new CSVSequenceRecordReader();
try {
trainLabels.initialize(new NumberedFileInputSplit(labelBaseDir + "/s_%d.csv", 0,this._modelDefinition.getNB_TRAIN_EXAMPLES()-1));
} catch (InterruptedException e) {
trainLabels.close();
trainFeatures.close();
throw new IOException(String.format("Interrupted exception error %s. during trainLabels initialise", e.getMessage()));
}
DataSetIterator trainData = new SequenceRecordReaderDataSetIterator(trainFeatures, trainLabels,
this._modelDefinition.getBATCH_SIZE(),this._modelDefinition.getNUM_LABEL_CLASSES(), false, SequenceRecordReaderDataSetIterator.AlignmentMode.ALIGN_END);
Here is my model:
MultiLayerConfiguration conf = new NeuralNetConfiguration.Builder()
.seed(this._modelDefinition.getRANDOM_SEED()) //Random number generator seed for improved repeatability. Optional.
.weightInit(WeightInit.XAVIER)
.updater(new Nesterovs(this._modelDefinition.getLEARNING_RATE()))
.gradientNormalization(GradientNormalization.ClipElementWiseAbsoluteValue) //Not always required, but helps with this data set
.gradientNormalizationThreshold(0.5)
.list()
.layer(0, new LSTM.Builder().activation(Activation.TANH).nIn(this._modelDefinition.getNB_INPUTS()).nOut(this._modelDefinition.getLSTM_LAYER_SIZE()).build())
.layer(1, new LSTM.Builder().activation(Activation.TANH).nIn(this._modelDefinition.getLSTM_LAYER_SIZE()).nOut(this._modelDefinition.getLSTM_LAYER_SIZE()).build())
.layer(2,new DenseLayer.Builder().nIn(this._modelDefinition.getLSTM_LAYER_SIZE()).nOut(this._modelDefinition.getLSTM_LAYER_SIZE())
.weightInit(WeightInit.XAVIER)
.build())
.layer(3, new RnnOutputLayer.Builder(LossFunctions.LossFunction.MCXENT)
.activation(Activation.SOFTMAX).nIn(this._modelDefinition.getLSTM_LAYER_SIZE()).nOut(this._modelDefinition.getNUM_LABEL_CLASSES()).build())
.pretrain(false).backprop(true).build();
I train the model over N epochs to get my optimal scores. I save the model, now I want to open the model and get classifications for new sequential feature files.
If there is an example of this - please let me know where.
thanks
anton

Answer 1

The answer is to feed the model the exact same input as we trained with, except set the labels to -1. The output will be an INDarray containing the probability of 0 in one array and the probability of 1 in the other array, showing up in the last sequence line.
Here is the code:
public void getOutputsForTheseInputsUsingThisNet(String netFilePath,String inputFileDir) throws Exception {
//open the network file
File locationToSave = new File(netFilePath);
MultiLayerNetwork nNet = null;
logger.info("Trying to open the model");
try {
nNet = ModelSerializer.restoreMultiLayerNetwork(locationToSave);
logger.info("Success: Model opened");
} catch (IOException e) {
throw new Exception(String.format("Unable to open model from %s because of error %s", locationToSave.getAbsolutePath(),e.getMessage()));
}
logger.info("Loading test data");
SequenceRecordReader testFeatures = new CSVSequenceRecordReader(0, ","); //skip no lines at the top - i.e. no header
try {
testFeatures.initialize(new NumberedFileInputSplit(inputFileDir + "/features/s_4180%d.csv", 0, 4));
} catch (InterruptedException e) {
testFeatures.close();
throw new Exception(String.format("IO error %s. during testFeatures", e.getMessage()));
}
logger.info("Loading label data");
SequenceRecordReader testLabels = new CSVSequenceRecordReader();
try {
testLabels.initialize(new NumberedFileInputSplit(inputFileDir + "/labels/s_4180%d.csv", 0,4));
} catch (InterruptedException e) {
testLabels.close();
testFeatures.close();
throw new IOException(String.format("Interrupted exception error %s. during testLabels initialise", e.getMessage()));
}
//DataSetIterator inputData = new Seque
logger.info("creating iterator");
DataSetIterator testData = new SequenceRecordReaderDataSetIterator(testFeatures, testLabels,
this._modelDefinition.getBATCH_SIZE(),this._modelDefinition.getNUM_LABEL_CLASSES(), false, SequenceRecordReaderDataSetIterator.AlignmentMode.ALIGN_END);
//now use it to classify some data
logger.info("classifying examples");
INDArray output = nNet.output(testData);
logger.info("outputing the classifications");
if(output==null||output.isEmpty())
throw new Exception("There is no output");
System.out.println(output);
//sample output
// [[[ 0, 0, 0, 0, 0.9882, 0, 0, 0, 0],
// [ 0, 0, 0, 0, 0.0118, 0, 0, 0, 0]],
//
// [[ 0, 0.1443, 0, 0, 0, 0, 0, 0, 0],
// [ 0, 0.8557, 0, 0, 0, 0, 0, 0, 0]],
//
// [[ 0, 0, 0, 0, 0, 0, 0, 0, 0.9975],
// [ 0, 0, 0, 0, 0, 0, 0, 0, 0.0025]],
//
// [[ 0, 0, 0, 0, 0, 0, 0.8482, 0, 0],
// [ 0, 0, 0, 0, 0, 0, 0.1518, 0, 0]],
//
// [[ 0, 0, 0, 0.8760, 0, 0, 0, 0, 0],
// [ 0, 0, 0, 0.1240, 0, 0, 0, 0, 0]]]
}