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 2.2.07Aug65431Jul30281514972130Jun2926232221191817161514131210987543228May252423211817161514131211108654330Apr292726242322212017161514131211109876532131Mar30292621201917161514121110987543228Feb272625242322212019171615141312118765431Jan30292827262221201413129876221Dec20191817161514131211improved printFixed min float valueAdded regressionDo not lock dicts when the pretraiend file is emptyAdded program arguments : scaleGrad and maxNormFocusedModule can now have pretraiend word embeddingsChanged the way prefix are handled in dictsCorrected bug where special columns where filled by extra TSV columnsallow multiple pretrained embeddings file for ContextModuleremoved word2vec from documentation<unk> in pretrained embeddgins is used for Dict::unknownValueStrAction EOS adds text metadataAdded Transformer MyModuleMoving nn to correct deviceMaking sure the nn are loaded to the correct deviceCorrected bugExplore different transitions only for parserSending torch object to CPU before saving them to diskADDCHARTOWORD can add multiple letters at onceCorrected bug in ContextualModuleFixed dynamic oracle for tokenization transitionsSenting neural network to correct deviceCorrected bug with multiple classifiersSupport for multiple classifiersCorrected a bug where a tensor was not send to the correct deviceShuffling batches Tensor in ConfigDatasetdifferent null value for each column in ContextModuleAdded cnnAdded seed program argumentBatched indexing for ContextualModuleAdding None to dict in cotnextual moduleAdded Contextual moduleCorrected bug on CustomHingeLoss where it would not run on gpuIn ContextModule : left-most and right-most child features are only taken into account if they are respectively to the left or to the right of their parentsRemoved debug printAdded custom hinge loss and added explorationThreshold program argumentAdded bce and mse lossesAdded gold parsing transitionsMake sure nothing is added to the dict mid trainingRemoved cycle checking in Action::attach and fixed index problem in some transition cost