public static bool Learn(LDAConfig ldaConfig, bool copyFeaturizedDoc)
{
if (File.Exists(ldaConfig.Model))
{
StatusMessage.Write("Skipping, model already exists. " + ldaConfig.Model);
return(false);
}
var featurizedDocFile = ldaConfig.FeaturizedDocuments;
if (copyFeaturizedDoc)
{
featurizedDocFile = string.Format(@"{0}\{1}", Path.GetDirectoryName(ldaConfig.WordTopicAllocations), Path.GetFileName(ldaConfig.FeaturizedDocuments));
File.Copy(ldaConfig.FeaturizedDocuments, featurizedDocFile);
}
StatusMessage.Write("Running VW to learn LDA...");
var command = AppDomain.CurrentDomain.BaseDirectory + "vw.exe";
var args =
" " + featurizedDocFile +
" --hash strings" +
" --lda " + ldaConfig.LDAParameters.NumTopics +
" --lda_alpha " + ldaConfig.LDAParameters.Alpha +
" --lda_rho " + ldaConfig.LDAParameters.Rho +
" --lda_D " + ldaConfig.ModelStatistics.DocumentCount +
" --minibatch " + ldaConfig.LDAParameters.Minibatch +
" --power_t " + ldaConfig.LDAParameters.PowerT +
" --initial_t " + ldaConfig.LDAParameters.InitialT +
" -b " + (int)Math.Ceiling(Math.Log(ldaConfig.ModelStatistics.VocabularySize, 2.0)) + // Gets size of the hash table used to store the topic allocations for each word.
" --passes " + ldaConfig.LDAParameters.Passes +
" -c " +
" --readable_model " + ldaConfig.WordTopicAllocations +
" -p " + ldaConfig.DocumentTopicAllocations +
" -f " + ldaConfig.Model;
Console.RunCommand(command, args);
if (copyFeaturizedDoc)
{
ConsoleColor color;
FileManager.DeleteFile(featurizedDocFile, out color);
FileManager.DeleteFile(featurizedDocFile + ".cache", out color);
}
return(true);
}
// Load existing LDA config files (*.LDAConfig.json)
public static List <string> LoadLDAConfigFiles(string modelRepositoryPath,
string trainingSampleName,
LDAConfig defaultModelConfig,
string SourceFolderOfLDAConfigFiles,
string learningConfigFilesFolder,
ref List <string> listOfLDAConfigFilesForFeaturization,
ref List <string> listOfLDAConfigFilesForTest)
{
// Get all ldaconfig.json files.
List <string> listOfLDAConfigFilesForTraining = new List <string>();
var ldaConfigFiles = Directory.GetFiles(SourceFolderOfLDAConfigFiles, "*LDAConfig.json", SearchOption.AllDirectories);
foreach (string ldaConfigFile in ldaConfigFiles)
{
// Load ldaconfig
var config = Program.LoadLDAConfig(ldaConfigFile);
var newConfig = new LDAConfig()
{
LDAParameters = config.LDAParameters,
FeaturizationParameters = config.FeaturizationParameters,
SampleName = trainingSampleName,
ModelRepositoryPath = modelRepositoryPath,
Locale = defaultModelConfig.Locale,
Corpus = defaultModelConfig.Corpus,
ModelStatistics = new ModelStatistics()
};
bool minMaxAdded = false;
UpdateConfigFileLists(ref listOfLDAConfigFilesForTraining, ref listOfLDAConfigFilesForTest, ref listOfLDAConfigFilesForFeaturization, learningConfigFilesFolder, newConfig, ref minMaxAdded);
}
listOfLDAConfigFilesForFeaturization.AddRange(listOfLDAConfigFilesForTraining.GroupBy(
config => new
{
Program.LoadLDAConfig(config).FeaturizationParameters.MinWordDocumentFrequency,
Program.LoadLDAConfig(config).FeaturizationParameters.MaxRalativeWordDocumentFrequency
})
.Select(g => g.First()));
return(listOfLDAConfigFilesForTraining);
}
/// <summary>
/// Load individual parameter range files and generate List<LDAConfig>
/// </summary>
/// <param name="modelRepositoryPath">The top folder where you want to save all the models that will be learned</param>
/// <param name="folderOfParamRangeFiles">The folder where the parameter range files are located</param>
/// <param name="trainingSampleName">the name of the training sample</param>
/// <param name="defaultModelConfig">the template of LDAConfig file</param>
/// <param name="configFilesFolder">The folder where LDAConfig files (for training) will be saved</param>
/// <param name="listOfLDAConfigFilesForFeaturization">absolute paths for LDAConfig files (for featurization)</param>
/// <param name="listOfLDAConfigFilesForTest">absolute paths for LDAConfig files (for metrics computation)</param>
/// <returns>absolute paths for LDAConfig files (for training)</returns>
public static List <string> GenerateLDAConfigFiles(string modelRepositoryPath,
string folderOfParamRangeFiles,
string trainingSampleName,
LDAConfig defaultModelConfig,
string configFilesFolder,
ref List <string> listOfLDAConfigFilesForFeaturization,
ref List <string> listOfLDAConfigFilesForTest)
{
var alphaRange = new List <double>();
var rhoRange = new List <double>();
var numOfTopicsRange = new List <int>();
var minibatchRange = new List <int>();
var powerTRange = new List <double>();
var initialTRange = new List <double>();
var passesRange = new List <int>();
var minWordDocFreqRange = new List <int>();
var maxRelWordDocFreqRange = new List <float>();
// Load LDA parameters from a single file or individual files.
int parameterIndex = -1;
if (File.Exists(Path.Combine(folderOfParamRangeFiles, "LDAParameters.tsv")))
{
foreach (var line in File.ReadLines(Path.Combine(folderOfParamRangeFiles, "LDAParameters.tsv")))
{
string lineContent = line.Trim().ToLower();
if (PARAMETER_INDEX_DICTIONARY.ContainsKey(lineContent))
{
parameterIndex = PARAMETER_INDEX_DICTIONARY[lineContent];
}
else
{
switch (parameterIndex)
{
case 0:
alphaRange = Helper.ParseListOfValues <double>(lineContent);
break;
case 1:
rhoRange = Helper.ParseListOfValues <double>(lineContent);
break;
case 2:
minWordDocFreqRange = Helper.ParseListOfValues <int>(lineContent);
break;
case 3:
maxRelWordDocFreqRange = Helper.ParseListOfValues <float>(lineContent);
break;
case 4:
numOfTopicsRange = Helper.ParseListOfValues <int>(lineContent);
break;
case 5:
passesRange = Helper.ParseListOfValues <int>(lineContent);
break;
case 6:
minibatchRange = Helper.ParseListOfValues <int>(lineContent);
break;
case 7:
initialTRange = Helper.ParseListOfValues <double>(lineContent);
break;
case 8:
powerTRange = Helper.ParseListOfValues <double>(lineContent);
break;
default:
break;
}
parameterIndex = -1;
}
}
}
else
{
alphaRange = File.ReadLines(Path.Combine(folderOfParamRangeFiles, "Alpha.txt")).Select(Double.Parse).ToList();
rhoRange = File.ReadLines(Path.Combine(folderOfParamRangeFiles, "Rho.txt")).Select(Double.Parse).ToList();
numOfTopicsRange = File.ReadLines(Path.Combine(folderOfParamRangeFiles, "NumTopics.txt")).Select(int.Parse).ToList();
minibatchRange = File.ReadLines(Path.Combine(folderOfParamRangeFiles, "Minibatch.txt")).Select(int.Parse).ToList();
powerTRange = File.ReadLines(Path.Combine(folderOfParamRangeFiles, "PowerT.txt")).Select(Double.Parse).ToList();
initialTRange = File.ReadLines(Path.Combine(folderOfParamRangeFiles, "InitialT.txt")).Select(Double.Parse).ToList();
passesRange = File.ReadLines(Path.Combine(folderOfParamRangeFiles, "Passes.txt")).Select(int.Parse).ToList();
minWordDocFreqRange = File.ReadLines(Path.Combine(folderOfParamRangeFiles, "MinWordDocumentFrequency.txt")).Select(int.Parse).ToList();
maxRelWordDocFreqRange = File.ReadLines(Path.Combine(folderOfParamRangeFiles, "MaxRalativeWordDocumentFrequency.txt")).Select(Single.Parse).ToList();
}
List <string> listOfLDAConfigFilesForTraining = new List <string>();
foreach (var min in minWordDocFreqRange)
{
foreach (var max in maxRelWordDocFreqRange)
{
var f = new FeaturizationParameters()
{
MinWordDocumentFrequency = min,
MaxRalativeWordDocumentFrequency = max
};
// a boolean flag ensuring each combination of min/max to be added into the list only once.
bool minMaxAdded = false;
foreach (var alpha in alphaRange)
{
foreach (var rho in rhoRange)
{
foreach (var numOfTopics in numOfTopicsRange)
{
foreach (var miniBatch in minibatchRange)
{
foreach (var powerT in powerTRange)
{
foreach (var initialT in initialTRange)
{
foreach (var passes in passesRange)
{
var p = new LDAParameters()
{
Alpha = alpha,
Rho = rho,
NumTopics = numOfTopics,
Minibatch = miniBatch,
PowerT = powerT,
InitialT = initialT,
Passes = passes
};
var config = new LDAConfig()
{
LDAParameters = p,
FeaturizationParameters = f,
SampleName = trainingSampleName,
ModelRepositoryPath = modelRepositoryPath,
Locale = defaultModelConfig.Locale,
Corpus = defaultModelConfig.Corpus,
ModelStatistics = new ModelStatistics()
};
UpdateConfigFileLists(ref listOfLDAConfigFilesForTraining, ref listOfLDAConfigFilesForTest, ref listOfLDAConfigFilesForFeaturization, configFilesFolder, config, ref minMaxAdded);
}
}
}
}
}
}
}
minMaxAdded = false;
} // foreach max
}
return(listOfLDAConfigFilesForTraining);
}
/// <summary>
/// Update the following lists of LDA config files, given an instance of LDAConfig:
/// listOfLDAConfigFilesForTraining, listOfLDAConfigFilesForTest and listOfLDAConfigFilesForFeaturization.
/// </summary>
private static void UpdateConfigFileLists(ref List <string> listOfLDAConfigFilesForTraining, ref List <string> listOfLDAConfigFilesForTest, ref List <string> listOfLDAConfigFilesForFeaturization, string learningConfigFilesFolder, LDAConfig config, ref bool minMaxAdded)
{
string json = JsonConvert.SerializeObject(config);
string ldaConfigFilePath = string.Format(@"{0}\{1}.LDAConfig.json", learningConfigFilesFolder, config.modelName);
File.WriteAllText(ldaConfigFilePath, json);
// Add config files into lists.
listOfLDAConfigFilesForTraining.Add(ldaConfigFilePath);
listOfLDAConfigFilesForTest.Add(config.LDAConfigFile);
if (!minMaxAdded)
{
listOfLDAConfigFilesForFeaturization.Add(ldaConfigFilePath);
minMaxAdded = true;
}
}
public static List <string> LoadLDAParameterTable(string modelRepositoryPath,
string trainingSampleName,
LDAConfig defaultModelConfig,
string ldaParameterTablePath,
string learningConfigFilesFolder,
ref List <string> listOfLDAConfigFilesForFeaturization,
ref List <string> listOfLDAConfigFilesForTest)
{
// Get all ldaconfig.json files.
List <string> listOfLDAConfigFilesForTraining = new List <string>();
List <List <string> > lists = File.ReadLines(ldaParameterTablePath)
.Skip(1) // skip column header
.Select(line => line.Split(new char[] { '\t' }, StringSplitOptions.RemoveEmptyEntries).ToList())
.ToList();
// Group all rows by <min, max>.
var minMaxGroups = lists.GroupBy(list => new
{
min = Helper.GetValue <int>(list[0]),
max = Helper.GetValue <float>(list[1])
});
foreach (var group in minMaxGroups)
{
var f = new FeaturizationParameters()
{
MinWordDocumentFrequency = group.Key.min,
MaxRalativeWordDocumentFrequency = group.Key.max
};
// a boolean flag ensuring each combination of min/max to be added into the list only once.
bool minMaxAdded = false;
foreach (var row in group)
{
// The default (LDA parameter) table has the following columns.
// min max topicCount alpha rho miniBatch passes initialT powerT
// E.g. \\ICE-Recommender\ModelRepository\TuneLDAParameters\RangesOfParams\LDAParameterTable.tsv
int numOfTopics = Helper.GetValue <int>(row[2]);
double alpha = Helper.GetValue <double>(row[3]);
double rho = Helper.GetValue <double>(row[4]);
int miniBatch = Helper.GetValue <int>(row[5]);
int passes = Helper.GetValue <int>(row[6]);
double initialT = Helper.GetValue <double>(row[7]);
double powerT = Helper.GetValue <double>(row[8]);
var p = new LDAParameters()
{
Alpha = alpha,
Rho = rho,
NumTopics = numOfTopics,
Minibatch = miniBatch,
PowerT = powerT,
InitialT = initialT,
Passes = passes
};
var config = new LDAConfig()
{
LDAParameters = p,
FeaturizationParameters = f,
SampleName = trainingSampleName,
ModelRepositoryPath = modelRepositoryPath,
Locale = defaultModelConfig.Locale,
Corpus = defaultModelConfig.Corpus,
ModelStatistics = new ModelStatistics()
};
UpdateConfigFileLists(ref listOfLDAConfigFilesForTraining, ref listOfLDAConfigFilesForTest, ref listOfLDAConfigFilesForFeaturization, learningConfigFilesFolder, config, ref minMaxAdded);
}
minMaxAdded = false;
}
return(listOfLDAConfigFilesForTraining);
}
/// <summary>
/// Initialize module and generate data files if not exist
/// </summary>
/// <param name="modelConfig">model config file</param>
/// <returns>true if successful</returns>
public bool Initialize(LDAConfig modelConfig)
{
// initialize vocabularies
StatusMessage.Write("Loading corpus vocabulary file...");
this.corpusVocabulary = CorpusVocabulary.Load(modelConfig.CorpusVocabulary);
if (this.corpusVocabulary == null)
{
return(false);
}
this.NumTopics = modelConfig.LDAParameters.NumTopics;
this.NumWords = modelConfig.ModelStatistics.VocabularySize;
this.documentTopicAllocationsFile = modelConfig.DocumentTopicAllocations;
this.numDocs = modelConfig.ModelStatistics.DocumentCount;
this.numPasses = modelConfig.LDAParameters.Passes;
// Initialize array of topic metrics: Coherence, Specificity, Distinctiveness
this.topicMetrics = new double[this.NumTopics, sizeof(MetricType)];
// sanity check
if (this.NumWords != this.corpusVocabulary.Count())
{
StatusMessage.Write("Number of vocabularies mismatch. Check your parameters.");
return(false);
}
// initialize topic words allocation
this.wordTopicAllocationsFileName = modelConfig.WordTopicAllocations;
string topicWordsAllocBinFile = wordTopicAllocationsFileName + BinFileExt;
StatusMessage.Write("Loading topic words allocation bin file...");
if (!this.LoadTopicWords(topicWordsAllocBinFile))
{
StatusMessage.Write("Generating topic words allocation bin file...");
if (!this.BuildTopicWordsAllocation(this.wordTopicAllocationsFileName, topicWordsAllocBinFile))
{
return(false);
}
}
// initialize words documents list map
string wordDocsListMapFile = Path.GetDirectoryName(modelConfig.FeaturizedDocuments) + @"\" + WordDocsListMapFileName;
StatusMessage.Write("Loading word documents list bin file...");
if (!this.LoadWordDocsListMap(wordDocsListMapFile))
{
StatusMessage.Write("Generating word documents list bin file...");
if (!this.BuildWordDocsListMap(modelConfig.FeaturizedDocuments, wordDocsListMapFile))
{
return(false);
}
}
// init topics info
var topicsInfoFile = Path.GetDirectoryName(modelConfig.DocumentTopicAllocations) + @"\" + TopicsInfoFileName;
if (!this.LoadTopicsInfo(topicsInfoFile))
{
if (!this.BuildTopicsInfo(topicsInfoFile))
{
return(false);
}
}
return(true);
}
