public static Triple <Formatter, float, float> validate(LangDescriptor language, IList <InputDocument> documents, InputDocument testDoc, bool saveOutput, bool computeEditDistance)
{
// kNNClassifier.resetCache();
Corpus corpus = new Corpus(documents, language);
corpus.train();
// System.out.printf("%d feature vectors\n", corpus.featureVectors.size());
Formatter formatter = new Formatter(corpus, language.indentSize);
string output = formatter.format(testDoc, false);
float editDistance = 0;
if (computeEditDistance)
{
editDistance = Dbg.normalizedLevenshteinDistance(testDoc.content, output);
}
ClassificationAnalysis analysis = new ClassificationAnalysis(testDoc, formatter.AnalysisPerToken);
// System.out.println(testDoc.fileName+": edit distance = "+editDistance+", error rate = "+analysis.getErrorRate());
if (saveOutput)
{
File dir = new File(outputDir + "/" + language.name);
if (saveOutput)
{
dir = new File(outputDir + "/" + language.name);
dir.mkdir();
}
org.antlr.codebuff.misc.Utils.writeFile(dir.Path + "/" + System.IO.Path.GetFileName(testDoc.fileName), output);
}
return(new Triple <Formatter, float?, float?>(formatter, editDistance, analysis.ErrorRate));
}
public virtual Triple <Formatter, float, float> validate(LangDescriptor language, IList <InputDocument> documents, string fileToExclude, int k, FeatureMetaData[] injectWSFeatures, FeatureMetaData[] alignmentFeatures, string outputDir, bool computeEditDistance, bool collectAnalysis)
{
string path = System.IO.Path.GetFullPath(fileToExclude);
IList <InputDocument> others = BuffUtils.filter(documents, d => !d.fileName.Equals(path));
IList <InputDocument> excluded = BuffUtils.filter(documents, d => d.fileName.Equals(path));
Debug.Assert(others.Count == documents.Count - 1);
// kNNClassifier.resetCache();
if (excluded.Count == 0)
{
Console.Error.WriteLine("Doc not in corpus: " + path);
return(null);
}
InputDocument testDoc = excluded[0];
DateTime start = System.DateTime.Now;
Corpus corpus = new Corpus(others, language);
corpus.train();
DateTime stop = System.DateTime.Now;
Formatter formatter = new Formatter(corpus, language.indentSize, k, injectWSFeatures, alignmentFeatures);
InputDocument originalDoc = testDoc;
DateTime format_start = System.DateTime.Now;
string output = formatter.format(testDoc, collectAnalysis);
DateTime format_stop = System.DateTime.Now;
float editDistance = 0;
if (computeEditDistance)
{
editDistance = Dbg.normalizedLevenshteinDistance(testDoc.content, output);
}
ClassificationAnalysis analysis = new ClassificationAnalysis(originalDoc, formatter.AnalysisPerToken);
Console.WriteLine(testDoc.fileName + ": edit distance = " + editDistance + ", error rate = " + analysis.ErrorRate);
if (!string.ReferenceEquals(outputDir, null))
{
string dir = outputDir + "/" + language.name + "/" + Tool.version;
if (!System.IO.Directory.Exists(dir))
{
System.IO.Directory.CreateDirectory(dir);
}
org.antlr.codebuff.misc.Utils.writeFile(dir + "/" + System.IO.Path.GetFileName(testDoc.fileName), output);
}
var tms = (stop - start);
var fms = format_stop - format_start;
trainingTimes.Add((double)tms.Milliseconds);
float tokensPerMS = testDoc.tokens.Size / (float)fms.TotalMilliseconds;
formattingTokensPerMS.Add((double)tokensPerMS);
Console.Write("Training time = {0:D} ms, formatting {1:D} ms, {2,5:F3} tokens/ms ({3:D} tokens)\n", tms, fms, tokensPerMS, testDoc.tokens.Size);
// System.out.printf("classify calls %d, hits %d rate %f\n",
// kNNClassifier.nClassifyCalls, kNNClassifier.nClassifyCacheHits,
// kNNClassifier.nClassifyCacheHits/(float) kNNClassifier.nClassifyCalls);
// System.out.printf("kNN calls %d, hits %d rate %f\n",
// kNNClassifier.nNNCalls, kNNClassifier.nNNCacheHits,
// kNNClassifier.nNNCacheHits/(float) kNNClassifier.nNNCalls);
return(new Triple <Formatter, float, float>(formatter, editDistance, analysis.ErrorRate));
}
public static void runCaptureForOneLanguage(LangDescriptor language)
{
IList <string> filenames = Tool.getFilenames(language.corpusDir, language.fileRegex);
IList <InputDocument> documents = Tool.load(filenames, language);
foreach (string fileName in filenames)
{
// Examine info for this file in isolation
Corpus fileCorpus = new Corpus(fileName, language);
fileCorpus.train();
Console.WriteLine(fileName);
// examineCorpus(corpus);
ArrayListMultiMap <FeatureVectorAsObject, int> ws = getWSContextCategoryMap(fileCorpus);
ArrayListMultiMap <FeatureVectorAsObject, int> hpos = getHPosContextCategoryMap(fileCorpus);
// Compare with corpus minus this file
string path = fileName;
IList <InputDocument> others = BuffUtils.filter(documents, d => !d.fileName.Equals(path));
Corpus corpus = new Corpus(others, language);
corpus.train();
// examineCorpus(corpus);
ArrayListMultiMap <FeatureVectorAsObject, int> corpus_ws = getWSContextCategoryMap(corpus);
ArrayListMultiMap <FeatureVectorAsObject, int> corpus_hpos = getHPosContextCategoryMap(corpus);
foreach (FeatureVectorAsObject x in ws.Keys)
{
HashBag <int> fwsCats = getCategoriesBag(ws[x]);
IList <float> fwsRatios = getCategoryRatios(fwsCats.Values);
HashBag <int> wsCats = getCategoriesBag(corpus_ws[x]);
IList <float> wsRatios = getCategoryRatios(wsCats.Values);
// compare file predictions with corpus predictions
if (!fwsRatios.SequenceEqual(wsRatios))
{
Console.WriteLine(fwsRatios + " vs " + wsRatios);
}
HashBag <int> fhposCats = getCategoriesBag(hpos[x]);
HashBag <int> hposCats = getCategoriesBag(corpus_hpos[x]);
}
break;
}
}
public static org.antlr.codebuff.misc.Pair <int, int> test(LangDescriptor language, IList <InputDocument> others, InputDocument testDoc)
{
var train_start = System.DateTime.Now;
Corpus corpus = new Corpus(others, language);
corpus.train();
var train_stop = System.DateTime.Now;
var format_start = System.DateTime.Now;
Formatter formatter = new Formatter(corpus, language.indentSize, Formatter.DEFAULT_K, FEATURES_INJECT_WS, FEATURES_HPOS);
formatter.format(testDoc, false);
var format_stop = System.DateTime.Now;
var train_time = (train_stop - train_start) / 1000000;
var format_time = (format_stop - format_start) / 1000000;
Log.Write("{0} training of {1} = {2:D}ms formatting = {3:D}ms\n", language.name, testDoc.fileName, train_time, format_time);
return(new org.antlr.codebuff.misc.Pair <int, int>((int)train_time, (int)format_time));
}
public static void runCaptureForOneLanguage(LangDescriptor language)
{
IList <string> filenames = Tool.getFilenames(language.corpusDir, language.fileRegex);
IList <float> selfEditDistances = new List <float>();
foreach (string fileName in filenames)
{
Corpus corpus = new Corpus(fileName, language);
corpus.train();
InputDocument testDoc = Tool.parse(fileName, corpus.language);
Formatter formatter = new Formatter(corpus, language.indentSize);
string output = formatter.format(testDoc, false);
// System.out.println(output);
float editDistance = Dbg.normalizedLevenshteinDistance(testDoc.content, output);
Log.WriteLine(fileName + " edit distance " + editDistance);
selfEditDistances.Add(editDistance);
}
{
Corpus corpus = new Corpus(language.corpusDir, language);
corpus.train();
IList <float> corpusEditDistances = new List <float>();
foreach (string fileName in filenames)
{
InputDocument testDoc = Tool.parse(fileName, corpus.language);
Formatter formatter = new Formatter(corpus, language.indentSize);
string output = formatter.format(testDoc, false);
// System.out.println(output);
float editDistance = Dbg.normalizedLevenshteinDistance(testDoc.content, output);
Log.WriteLine(fileName + "+corpus edit distance " + editDistance);
corpusEditDistances.Add(editDistance);
}
// heh this gives info on within-corpus variability. i.e., how good/consistent is my corpus?
// those files with big difference are candidates for dropping from corpus or for cleanup.
IList <string> labels = BuffUtils.map(filenames, f => '"' + System.IO.Path.GetFileName(f) + '"');
string python = "#\n" + "# AUTO-GENERATED FILE. DO NOT EDIT\n" + "# CodeBuff <version> '<date>'\n" + "#\n" +
"import numpy as np\n" + "import matplotlib.pyplot as plt\n\n" + "fig = plt.figure()\n" +
"ax = plt.subplot(111)\n" + "labels = <labels>\n" + "N = len(labels)\n\n" +
"featureIndexes = range(0,N)\n" + "<lang>_self = <selfEditDistances>\n" +
"<lang>_corpus = <corpusEditDistances>\n" +
"<lang>_diff = np.abs(np.subtract(<lang>_self, <lang>_corpus))\n\n" +
"all = zip(<lang>_self, <lang>_corpus, <lang>_diff, labels)\n" +
"all = sorted(all, key=lambda x : x[2], reverse=True)\n" +
"<lang>_self, <lang>_corpus, <lang>_diff, labels = zip(*all)\n\n" +
"ax.plot(featureIndexes, <lang>_self, label=\"<lang>_self\")\n" +
"#ax.plot(featureIndexes, <lang>_corpus, label=\"<lang>_corpus\")\n" +
"ax.plot(featureIndexes, <lang>_diff, label=\"<lang>_diff\")\n" +
"ax.set_xticklabels(labels, rotation=60, fontsize=8)\n" +
"plt.xticks(featureIndexes, labels, rotation=60)\n" +
"ax.yaxis.grid(True, linestyle='-', which='major', color='lightgrey', alpha=0.5)\n\n" +
"ax.text(1, .25, 'median $f$ self distance = %5.3f, corpus+$f$ distance = %5.3f' %" +
" (np.median(<lang>_self),np.median(<lang>_corpus)))\n" + "ax.set_xlabel(\"File Name\")\n" +
"ax.set_ylabel(\"Edit Distance\")\n" +
"ax.set_title(\"Difference between Formatting File <lang> $f$\\nwith Training=$f$ and Training=$f$+Corpus\")\n" +
"plt.legend()\n" + "plt.tight_layout()\n" + "fig.savefig(\"images/" + language.name +
"_one_file_capture.pdf\", format='pdf')\n" + "plt.show()\n";
ST pythonST = new ST(python);
pythonST.add("lang", language.name);
pythonST.add("version", version);
pythonST.add("date", DateTime.Now);
pythonST.add("labels", labels.ToString());
pythonST.add("selfEditDistances", selfEditDistances.ToString());
pythonST.add("corpusEditDistances", corpusEditDistances.ToString());
string code = pythonST.render();
{
string fileName = "python/src/" + language.name + "_one_file_capture.py";
org.antlr.codebuff.misc.Utils.writeFile(fileName, code);
Log.WriteLine("wrote python code to " + fileName);
}
}
}
public static void computeConsistency(LangDescriptor language, bool report)
{
if (report)
{
Console.WriteLine("-----------------------------------");
Console.WriteLine(language.name);
Console.WriteLine("-----------------------------------");
}
Corpus corpus = new Corpus(language.corpusDir, language);
corpus.train();
// a map of feature vector to list of exemplar indexes of that feature
MyMultiMap <FeatureVectorAsObject, int> wsContextToIndex = new MyMultiMap <FeatureVectorAsObject, int>();
MyMultiMap <FeatureVectorAsObject, int> hposContextToIndex = new MyMultiMap <FeatureVectorAsObject, int>();
int n = corpus.featureVectors.Count;
for (int i = 0; i < n; i++)
{
int[] features = corpus.featureVectors[i];
wsContextToIndex.Map(new FeatureVectorAsObject(features, Trainer.FEATURES_INJECT_WS), i);
hposContextToIndex.Map(new FeatureVectorAsObject(features, Trainer.FEATURES_HPOS), i);
}
int num_ambiguous_ws_vectors = 0;
int num_ambiguous_hpos_vectors = 0;
// Dump output grouped by ws vs hpos then feature vector then category
if (report)
{
Console.WriteLine(" --- INJECT WS ---");
}
IList <double> ws_entropies = new List <double>();
foreach (FeatureVectorAsObject fo in wsContextToIndex.Keys)
{
var exemplarIndexes = wsContextToIndex[fo];
// we have group by feature vector, now group by cat with that set for ws
MyMultiMap <int, int> wsCatToIndexes = new MyMultiMap <int, int>();
foreach (int i in exemplarIndexes)
{
wsCatToIndexes.Map(corpus.injectWhitespace[i], i);
}
if (wsCatToIndexes.Count == 1)
{
continue;
}
if (report)
{
Console.WriteLine("Feature vector has " + exemplarIndexes.size() + " exemplars");
}
IList <int> catCounts = BuffUtils.map(wsCatToIndexes.Values, (x) => x.size());
double wsEntropy = Entropy.getNormalizedCategoryEntropy(Entropy.getCategoryRatios(catCounts));
if (report)
{
Console.Write("entropy={0,5:F4}\n", wsEntropy);
}
wsEntropy *= exemplarIndexes.size();
ws_entropies.Add(wsEntropy);
num_ambiguous_ws_vectors += exemplarIndexes.size();
if (report)
{
Console.Write(Trainer.featureNameHeader(Trainer.FEATURES_INJECT_WS));
}
if (report)
{
foreach (int cat in wsCatToIndexes.Keys)
{
var indexes = wsCatToIndexes[cat];
foreach (int i in indexes)
{
string display = getExemplarDisplay(Trainer.FEATURES_INJECT_WS, corpus, corpus.injectWhitespace, i);
Console.WriteLine(display);
}
Console.WriteLine();
}
}
}
if (report)
{
Console.WriteLine(" --- HPOS ---");
}
IList <double> hpos_entropies = new List <double>();
foreach (FeatureVectorAsObject fo in hposContextToIndex.Keys)
{
MyHashSet <int> exemplarIndexes = hposContextToIndex[fo];
// we have group by feature vector, now group by cat with that set for hpos
MyMultiMap <int, int> hposCatToIndexes = new MyMultiMap <int, int>();
foreach (int i in exemplarIndexes)
{
hposCatToIndexes.Map(corpus.hpos[i], i);
}
if (hposCatToIndexes.Count == 1)
{
continue;
}
if (report)
{
Console.WriteLine("Feature vector has " + exemplarIndexes.size() + " exemplars");
}
IList <int> catCounts = BuffUtils.map(hposCatToIndexes.Values, (x) => x.size());
double hposEntropy = Entropy.getNormalizedCategoryEntropy(Entropy.getCategoryRatios(catCounts));
if (report)
{
Console.Write("entropy={0,5:F4}\n", hposEntropy);
}
hposEntropy *= exemplarIndexes.size();
hpos_entropies.Add(hposEntropy);
num_ambiguous_hpos_vectors += exemplarIndexes.size();
if (report)
{
Console.Write(Trainer.featureNameHeader(Trainer.FEATURES_HPOS));
}
if (report)
{
foreach (int cat in hposCatToIndexes.Keys)
{
var indexes = hposCatToIndexes[cat];
foreach (int?i in indexes)
{
string display = getExemplarDisplay(Trainer.FEATURES_HPOS, corpus, corpus.hpos, i.Value);
Console.WriteLine(display);
}
Console.WriteLine();
}
}
}
Console.WriteLine();
Console.WriteLine(language.name);
Console.WriteLine("There are " + wsContextToIndex.Count + " unique ws feature vectors out of " + n + " = " + string.Format("{0,3:F1}%", 100.0 * wsContextToIndex.Count / n));
Console.WriteLine("There are " + hposContextToIndex.Count + " unique hpos feature vectors out of " + n + " = " + string.Format("{0,3:F1}%", 100.0 * hposContextToIndex.Count / n));
float prob_ws_ambiguous = num_ambiguous_ws_vectors / (float)n;
Console.Write("num_ambiguous_ws_vectors = {0,5:D}/{1,5:D} = {2,5:F3}\n", num_ambiguous_ws_vectors, n, prob_ws_ambiguous);
float prob_hpos_ambiguous = num_ambiguous_hpos_vectors / (float)n;
Console.Write("num_ambiguous_hpos_vectors = {0,5:D}/{1,5:D} = {2,5:F3}\n", num_ambiguous_hpos_vectors, n, prob_hpos_ambiguous);
// Collections.sort(ws_entropies);
// System.out.println("ws_entropies="+ws_entropies);
Console.WriteLine("ws median,mean = " + BuffUtils.median(ws_entropies) + "," + BuffUtils.mean(ws_entropies));
double expected_ws_entropy = (BuffUtils.sumDoubles(ws_entropies) / num_ambiguous_ws_vectors) * prob_ws_ambiguous;
Console.WriteLine("expected_ws_entropy=" + expected_ws_entropy);
Console.WriteLine("hpos median,mean = " + BuffUtils.median(hpos_entropies) + "," + BuffUtils.mean(hpos_entropies));
double expected_hpos_entropy = (BuffUtils.sumDoubles(hpos_entropies) / num_ambiguous_hpos_vectors) * prob_hpos_ambiguous;
Console.WriteLine("expected_hpos_entropy=" + expected_hpos_entropy);
}