public AbstractEvaluate(RNNOptions options)
{
// the matrix will be [gold][predicted]
// TODO: make this an option
this.op = options;
this.Reset();
}
/// <summary>
/// Expected arguments are
/// <c>-gold gold -predicted predicted</c>
/// For example <br />
/// <c>java edu.stanford.nlp.sentiment.ExternalEvaluate annotatedTrees.txt predictedTrees.txt</c>
/// </summary>
public static void Main(string[] args)
{
RNNOptions curOptions = new RNNOptions();
string goldPath = null;
string predictedPath = null;
for (int argIndex = 0; argIndex < args.Length;)
{
if (Sharpen.Runtime.EqualsIgnoreCase(args[argIndex], "-gold"))
{
goldPath = args[argIndex + 1];
argIndex += 2;
}
else
{
if (Sharpen.Runtime.EqualsIgnoreCase(args[argIndex], "-predicted"))
{
predictedPath = args[argIndex + 1];
argIndex += 2;
}
else
{
int newArgIndex = curOptions.SetOption(args, argIndex);
if (newArgIndex == argIndex)
{
throw new ArgumentException("Unknown argument " + args[argIndex]);
}
argIndex = newArgIndex;
}
}
}
if (goldPath == null)
{
log.Info("goldPath not set. Exit.");
System.Environment.Exit(-1);
}
if (predictedPath == null)
{
log.Info("predictedPath not set. Exit.");
System.Environment.Exit(-1);
}
// filterUnknown not supported because I'd need to know which sentences
// are removed to remove them from predicted
IList <Tree> goldTrees = SentimentUtils.ReadTreesWithGoldLabels(goldPath);
IList <Tree> predictedTrees = SentimentUtils.ReadTreesWithPredictedLabels(predictedPath);
Edu.Stanford.Nlp.Sentiment.ExternalEvaluate evaluator = new Edu.Stanford.Nlp.Sentiment.ExternalEvaluate(curOptions, predictedTrees);
evaluator.Eval(goldTrees);
evaluator.PrintSummary();
}
private SentimentModel(TwoDimensionalMap <string, string, SimpleMatrix> binaryTransform, TwoDimensionalMap <string, string, SimpleTensor> binaryTensors, TwoDimensionalMap <string, string, SimpleMatrix> binaryClassification, IDictionary <string,
SimpleMatrix> unaryClassification, IDictionary <string, SimpleMatrix> wordVectors, RNNOptions op)
{
this.op = op;
this.binaryTransform = binaryTransform;
this.binaryTensors = binaryTensors;
this.binaryClassification = binaryClassification;
this.unaryClassification = unaryClassification;
this.wordVectors = wordVectors;
this.numClasses = op.numClasses;
if (op.numHid <= 0)
{
int nh = 0;
foreach (SimpleMatrix wv in wordVectors.Values)
{
nh = wv.GetNumElements();
}
this.numHid = nh;
}
else
{
this.numHid = op.numHid;
}
this.numBinaryMatrices = binaryTransform.Size();
binaryTransformSize = numHid * (2 * numHid + 1);
if (op.useTensors)
{
binaryTensorSize = numHid * numHid * numHid * 4;
}
else
{
binaryTensorSize = 0;
}
binaryClassificationSize = (op.combineClassification) ? 0 : numClasses * (numHid + 1);
numUnaryMatrices = unaryClassification.Count;
unaryClassificationSize = numClasses * (numHid + 1);
rand = new Random(op.randomSeed);
identity = SimpleMatrix.Identity(numHid);
}
/// <summary>The traditional way of initializing an empty model suitable for training.</summary>
public SentimentModel(RNNOptions op, IList <Tree> trainingTrees)
{
this.op = op;
rand = new Random(op.randomSeed);
if (op.randomWordVectors)
{
InitRandomWordVectors(trainingTrees);
}
else
{
ReadWordVectors();
}
if (op.numHid > 0)
{
this.numHid = op.numHid;
}
else
{
int size = 0;
foreach (SimpleMatrix vector in wordVectors.Values)
{
size = vector.GetNumElements();
break;
}
this.numHid = size;
}
TwoDimensionalSet <string, string> binaryProductions = TwoDimensionalSet.HashSet();
if (op.simplifiedModel)
{
binaryProductions.Add(string.Empty, string.Empty);
}
else
{
// TODO
// figure out what binary productions we have in these trees
// Note: the current sentiment training data does not actually
// have any constituent labels
throw new NotSupportedException("Not yet implemented");
}
ICollection <string> unaryProductions = Generics.NewHashSet();
if (op.simplifiedModel)
{
unaryProductions.Add(string.Empty);
}
else
{
// TODO
// figure out what unary productions we have in these trees (preterminals only, after the collapsing)
throw new NotSupportedException("Not yet implemented");
}
this.numClasses = op.numClasses;
identity = SimpleMatrix.Identity(numHid);
binaryTransform = TwoDimensionalMap.TreeMap();
binaryTensors = TwoDimensionalMap.TreeMap();
binaryClassification = TwoDimensionalMap.TreeMap();
// When making a flat model (no symantic untying) the
// basicCategory function will return the same basic category for
// all labels, so all entries will map to the same matrix
foreach (Pair <string, string> binary in binaryProductions)
{
string left = BasicCategory(binary.first);
string right = BasicCategory(binary.second);
if (binaryTransform.Contains(left, right))
{
continue;
}
binaryTransform.Put(left, right, RandomTransformMatrix());
if (op.useTensors)
{
binaryTensors.Put(left, right, RandomBinaryTensor());
}
if (!op.combineClassification)
{
binaryClassification.Put(left, right, RandomClassificationMatrix());
}
}
numBinaryMatrices = binaryTransform.Size();
binaryTransformSize = numHid * (2 * numHid + 1);
if (op.useTensors)
{
binaryTensorSize = numHid * numHid * numHid * 4;
}
else
{
binaryTensorSize = 0;
}
binaryClassificationSize = (op.combineClassification) ? 0 : numClasses * (numHid + 1);
unaryClassification = Generics.NewTreeMap();
// When making a flat model (no symantic untying) the
// basicCategory function will return the same basic category for
// all labels, so all entries will map to the same matrix
foreach (string unary in unaryProductions)
{
unary = BasicCategory(unary);
if (unaryClassification.Contains(unary))
{
continue;
}
unaryClassification[unary] = RandomClassificationMatrix();
}
numUnaryMatrices = unaryClassification.Count;
unaryClassificationSize = numClasses * (numHid + 1);
}
/*
* // An example of how you could read in old models with readObject to fix the serialization
* // You would first read in the old model, then reserialize it
* private void readObject(ObjectInputStream in)
* throws IOException, ClassNotFoundException
* {
* ObjectInputStream.GetField fields = in.readFields();
* binaryTransform = ErasureUtils.uncheckedCast(fields.get("binaryTransform", null));
*
* // transform binaryTensors
* binaryTensors = TwoDimensionalMap.treeMap();
* TwoDimensionalMap<String, String, edu.stanford.nlp.rnn.SimpleTensor> oldTensors = ErasureUtils.uncheckedCast(fields.get("binaryTensors", null));
* for (String first : oldTensors.firstKeySet()) {
* for (String second : oldTensors.get(first).keySet()) {
* binaryTensors.put(first, second, new SimpleTensor(oldTensors.get(first, second).slices));
* }
* }
*
* binaryClassification = ErasureUtils.uncheckedCast(fields.get("binaryClassification", null));
* unaryClassification = ErasureUtils.uncheckedCast(fields.get("unaryClassification", null));
* wordVectors = ErasureUtils.uncheckedCast(fields.get("wordVectors", null));
*
* if (fields.defaulted("numClasses")) {
* throw new RuntimeException();
* }
* numClasses = fields.get("numClasses", 0);
*
* if (fields.defaulted("numHid")) {
* throw new RuntimeException();
* }
* numHid = fields.get("numHid", 0);
*
* if (fields.defaulted("numBinaryMatrices")) {
* throw new RuntimeException();
* }
* numBinaryMatrices = fields.get("numBinaryMatrices", 0);
*
* if (fields.defaulted("binaryTransformSize")) {
* throw new RuntimeException();
* }
* binaryTransformSize = fields.get("binaryTransformSize", 0);
*
* if (fields.defaulted("binaryTensorSize")) {
* throw new RuntimeException();
* }
* binaryTensorSize = fields.get("binaryTensorSize", 0);
*
* if (fields.defaulted("binaryClassificationSize")) {
* throw new RuntimeException();
* }
* binaryClassificationSize = fields.get("binaryClassificationSize", 0);
*
* if (fields.defaulted("numUnaryMatrices")) {
* throw new RuntimeException();
* }
* numUnaryMatrices = fields.get("numUnaryMatrices", 0);
*
* if (fields.defaulted("unaryClassificationSize")) {
* throw new RuntimeException();
* }
* unaryClassificationSize = fields.get("unaryClassificationSize", 0);
*
* rand = ErasureUtils.uncheckedCast(fields.get("rand", null));
* op = ErasureUtils.uncheckedCast(fields.get("op", null));
* op.classNames = op.DEFAULT_CLASS_NAMES;
* op.equivalenceClasses = op.APPROXIMATE_EQUIVALENCE_CLASSES;
* op.equivalenceClassNames = op.DEFAULT_EQUIVALENCE_CLASS_NAMES;
* }
*/
/// <summary>
/// Given single matrices and sets of options, create the
/// corresponding SentimentModel.
/// </summary>
/// <remarks>
/// Given single matrices and sets of options, create the
/// corresponding SentimentModel. Useful for creating a Java version
/// of a model trained in some other manner, such as using the
/// original Matlab code.
/// </remarks>
internal static Edu.Stanford.Nlp.Sentiment.SentimentModel ModelFromMatrices(SimpleMatrix W, SimpleMatrix Wcat, SimpleTensor Wt, IDictionary <string, SimpleMatrix> wordVectors, RNNOptions op)
{
if (!op.combineClassification || !op.simplifiedModel)
{
throw new ArgumentException("Can only create a model using this method if combineClassification and simplifiedModel are turned on");
}
TwoDimensionalMap <string, string, SimpleMatrix> binaryTransform = TwoDimensionalMap.TreeMap();
binaryTransform.Put(string.Empty, string.Empty, W);
TwoDimensionalMap <string, string, SimpleTensor> binaryTensors = TwoDimensionalMap.TreeMap();
binaryTensors.Put(string.Empty, string.Empty, Wt);
TwoDimensionalMap <string, string, SimpleMatrix> binaryClassification = TwoDimensionalMap.TreeMap();
IDictionary <string, SimpleMatrix> unaryClassification = Generics.NewTreeMap();
unaryClassification[string.Empty] = Wcat;
return(new Edu.Stanford.Nlp.Sentiment.SentimentModel(binaryTransform, binaryTensors, binaryClassification, unaryClassification, wordVectors, op));
}
public ExternalEvaluate(RNNOptions op, IList <Tree> predictedTrees)
: base(op)
{
this.predicted = predictedTrees;
}
/// <exception cref="System.IO.IOException"/>
public static void Main(string[] args)
{
string basePath = "/user/socherr/scr/projects/semComp/RNTN/src/params/";
int numSlices = 25;
bool useEscapedParens = false;
for (int argIndex = 0; argIndex < args.Length;)
{
if (Sharpen.Runtime.EqualsIgnoreCase(args[argIndex], "-slices"))
{
numSlices = System.Convert.ToInt32(args[argIndex + 1]);
argIndex += 2;
}
else
{
if (Sharpen.Runtime.EqualsIgnoreCase(args[argIndex], "-path"))
{
basePath = args[argIndex + 1];
argIndex += 2;
}
else
{
if (Sharpen.Runtime.EqualsIgnoreCase(args[argIndex], "-useEscapedParens"))
{
useEscapedParens = true;
argIndex += 1;
}
else
{
log.Info("Unknown argument " + args[argIndex]);
System.Environment.Exit(2);
}
}
}
}
SimpleMatrix[] slices = new SimpleMatrix[numSlices];
for (int i = 0; i < numSlices; ++i)
{
slices[i] = LoadMatrix(basePath + "bin/Wt_" + (i + 1) + ".bin", basePath + "Wt_" + (i + 1) + ".txt");
}
SimpleTensor tensor = new SimpleTensor(slices);
log.Info("W tensor size: " + tensor.NumRows() + "x" + tensor.NumCols() + "x" + tensor.NumSlices());
SimpleMatrix W = LoadMatrix(basePath + "bin/W.bin", basePath + "W.txt");
log.Info("W matrix size: " + W.NumRows() + "x" + W.NumCols());
SimpleMatrix Wcat = LoadMatrix(basePath + "bin/Wcat.bin", basePath + "Wcat.txt");
log.Info("W cat size: " + Wcat.NumRows() + "x" + Wcat.NumCols());
SimpleMatrix combinedWV = LoadMatrix(basePath + "bin/Wv.bin", basePath + "Wv.txt");
log.Info("Word matrix size: " + combinedWV.NumRows() + "x" + combinedWV.NumCols());
File vocabFile = new File(basePath + "vocab_1.txt");
if (!vocabFile.Exists())
{
vocabFile = new File(basePath + "words.txt");
}
IList <string> lines = Generics.NewArrayList();
foreach (string line in IOUtils.ReadLines(vocabFile))
{
lines.Add(line.Trim());
}
log.Info("Lines in vocab file: " + lines.Count);
IDictionary <string, SimpleMatrix> wordVectors = Generics.NewTreeMap();
for (int i_1 = 0; i_1 < lines.Count && i_1 < combinedWV.NumCols(); ++i_1)
{
string[] pieces = lines[i_1].Split(" +");
if (pieces.Length == 0 || pieces.Length > 1)
{
continue;
}
wordVectors[pieces[0]] = combinedWV.ExtractMatrix(0, numSlices, i_1, i_1 + 1);
if (pieces[0].Equals("UNK"))
{
wordVectors[SentimentModel.UnknownWord] = wordVectors["UNK"];
}
}
// If there is no ",", we first try to look for an HTML escaping,
// then fall back to "." as better than just a random word vector.
// Same for "``" and ";"
CopyWordVector(wordVectors, ",", ",");
CopyWordVector(wordVectors, ".", ",");
CopyWordVector(wordVectors, ";", ";");
CopyWordVector(wordVectors, ".", ";");
CopyWordVector(wordVectors, "``", "``");
CopyWordVector(wordVectors, "''", "``");
if (useEscapedParens)
{
ReplaceWordVector(wordVectors, "(", "-LRB-");
ReplaceWordVector(wordVectors, ")", "-RRB-");
}
RNNOptions op = new RNNOptions();
op.numHid = numSlices;
op.lowercaseWordVectors = false;
if (Wcat.NumRows() == 2)
{
op.classNames = new string[] { "Negative", "Positive" };
op.equivalenceClasses = new int[][] { new int[] { 0 }, new int[] { 1 } };
// TODO: set to null once old models are updated
op.numClasses = 2;
}
if (!wordVectors.Contains(SentimentModel.UnknownWord))
{
wordVectors[SentimentModel.UnknownWord] = SimpleMatrix.Random(numSlices, 1, -0.00001, 0.00001, new Random());
}
SentimentModel model = SentimentModel.ModelFromMatrices(W, Wcat, tensor, wordVectors, op);
model.SaveSerialized("matlab.ser.gz");
}
/// <summary>Trains a sentiment model.</summary>
/// <remarks>
/// Trains a sentiment model.
/// The -trainPath argument points to a labeled sentiment treebank.
/// The trees in this data will be used to train the model parameters (also to seed the model vocabulary).
/// The -devPath argument points to a second labeled sentiment treebank.
/// The trees in this data will be used to periodically evaluate the performance of the model.
/// We won't train on this data; it will only be used to test how well the model generalizes to unseen data.
/// The -model argument specifies where to save the learned sentiment model.
/// </remarks>
/// <param name="args">Command line arguments</param>
public static void Main(string[] args)
{
RNNOptions op = new RNNOptions();
string trainPath = "sentimentTreesDebug.txt";
string devPath = null;
bool runGradientCheck = false;
bool runTraining = false;
bool filterUnknown = false;
string modelPath = null;
for (int argIndex = 0; argIndex < args.Length;)
{
if (Sharpen.Runtime.EqualsIgnoreCase(args[argIndex], "-train"))
{
runTraining = true;
argIndex++;
}
else
{
if (Sharpen.Runtime.EqualsIgnoreCase(args[argIndex], "-gradientcheck"))
{
runGradientCheck = true;
argIndex++;
}
else
{
if (Sharpen.Runtime.EqualsIgnoreCase(args[argIndex], "-trainpath"))
{
trainPath = args[argIndex + 1];
argIndex += 2;
}
else
{
if (Sharpen.Runtime.EqualsIgnoreCase(args[argIndex], "-devpath"))
{
devPath = args[argIndex + 1];
argIndex += 2;
}
else
{
if (Sharpen.Runtime.EqualsIgnoreCase(args[argIndex], "-model"))
{
modelPath = args[argIndex + 1];
argIndex += 2;
}
else
{
if (Sharpen.Runtime.EqualsIgnoreCase(args[argIndex], "-filterUnknown"))
{
filterUnknown = true;
argIndex++;
}
else
{
int newArgIndex = op.SetOption(args, argIndex);
if (newArgIndex == argIndex)
{
throw new ArgumentException("Unknown argument " + args[argIndex]);
}
argIndex = newArgIndex;
}
}
}
}
}
}
}
// read in the trees
IList <Tree> trainingTrees = SentimentUtils.ReadTreesWithGoldLabels(trainPath);
log.Info("Read in " + trainingTrees.Count + " training trees");
if (filterUnknown)
{
trainingTrees = SentimentUtils.FilterUnknownRoots(trainingTrees);
log.Info("Filtered training trees: " + trainingTrees.Count);
}
IList <Tree> devTrees = null;
if (devPath != null)
{
devTrees = SentimentUtils.ReadTreesWithGoldLabels(devPath);
log.Info("Read in " + devTrees.Count + " dev trees");
if (filterUnknown)
{
devTrees = SentimentUtils.FilterUnknownRoots(devTrees);
log.Info("Filtered dev trees: " + devTrees.Count);
}
}
// TODO: binarize the trees, then collapse the unary chains.
// Collapsed unary chains always have the label of the top node in
// the chain
// Note: the sentiment training data already has this done.
// However, when we handle trees given to us from the Stanford Parser,
// we will have to perform this step
// build an uninitialized SentimentModel from the binary productions
log.Info("Sentiment model options:\n" + op);
SentimentModel model = new SentimentModel(op, trainingTrees);
if (op.trainOptions.initialMatrixLogPath != null)
{
StringUtils.PrintToFile(new File(op.trainOptions.initialMatrixLogPath), model.ToString(), false, false, "utf-8");
}
// TODO: need to handle unk rules somehow... at test time the tree
// structures might have something that we never saw at training
// time. for example, we could put a threshold on all of the
// rules at training time and anything that doesn't meet that
// threshold goes into the unk. perhaps we could also use some
// component of the accepted training rules to build up the "unk"
// parameter in case there are no rules that don't meet the
// threshold
if (runGradientCheck)
{
RunGradientCheck(model, trainingTrees);
}
if (runTraining)
{
Train(model, modelPath, trainingTrees, devTrees);
model.SaveSerialized(modelPath);
}
}