Skip to content
Snippets Groups Projects

Repository graph

You can move around the graph by using the arrow keys.
Select Git revision
  • loss
  • master default protected
  • producer
3 results
Created with Raphaël 2.2.025Nov12109710Oct931Aug18222Jul2130Apr2818161514130Mar25242321201612765432122Feb201917161211109876520Jan191715847Dec17Nov1630Oct2623221614910Sep916Aug11765431Jul30281514972130Jun2926232221191817161514131210987543228May252423211817161514131211108654330Apr292726242322212017161514131211109876532131Mar30292621201917161514121110987543228Feb272625242322212019171615141312118765431Jan30292827262221201413129876221Dec20191817161514131211Added learnable coefs for the losslosslossHaving separate wordEmbeddings for special valuesCorrected bug where wordIndex would overflowmoveWordIndexRelaxed is skipping multiwordsIgnoring multiwords while moving word indexTried to improve pretrainedConcat works with multiple columnsAdded dimmension reduction in Concat moduleSpecial embeddings can be trained even with lockPretrainedDisable parralelismAllow spaces in w2vAllow space in w2v formatAdded different types of entropiesSet finished status of rawInput on true at constructionproducerproducercorrected bug where apply was processed when there was nothing to produceWorking first versionAdded output sesAdded and integred producerAdded option memcheck to trainAdded date in timestampsShuffle dataset on CPU to avoid CUDA sync errorAdded module HistoryMineSpeed up oracleIncreased max nb examples to 20 millionsChanged the encoding of features in certain modulesFixed 'outputAll' option for LSTM moduleAdded sanity check when loading pretrained word embeddingsWhen constructing WordEmbedding, only use max_norm when necessaryImproved forcing EOS transition, usefull for lineByLine modeAllow blank line to be commented in tsvWhen loading optimizer, do not silently failRemoved debug printRemoved warningAdded functions getDevice and setDeviceAdded test for overflowFixed discrete features valuesAdded fmt as a submoduleUsing bare std::string instead of flyweight because it caused datarace in some computersChanged execution policy from par_unseq to parParallel extractExamples