Theano and LSTM for Sentiment Analysis

《Theano and LSTM for Sentiment Analysis》由会员上传分享，免费在线阅读，更多相关内容在学术论文-天天文库。

1、TheanoandLSTMforSentimentAnalysisFrédéricBastienDépartementd'InformatiqueetdeRechercheOpérationnelleUniversitédeMontréalMontréal,Canadabastienf@iro.umontreal.caPresentationpreparedwithPierreLucCarrier,KyungHyunChoandÇa§larGülçehreNext.ML2015IntroductionTheanoRNNLSTMExercicesTaskThisisac

2、lassicationtaskwhere,givenareviewofamovie,weneedtoestablishwhetherthemoviereviewispositiveornegative.WeusetheIMDBdataset.2/60IntroductionTheanoRNNLSTMExercicesIntroductionTheanoCompiling/RunningModifyingexpressionsDebuggingScanRNNLSTMExercices3/60IntroductionTheanoRNNLSTMExercicesHighl

3、evelPython<-{NumPy/SciPy/libgpuarray}<-Theano<-Pylearn2IPython:OOcodinglanguageINumpy:n-dimensionalarrayobjectandscienticcomputingtoolboxISciPy:sparsematrixobjectsandmorescienticcomputingfunctionalityIlibgpuarray:GPUn-dimensionalarrayobjectinCforCUDAandOpenCLITheano:compiler/symbolicg

4、raphmanipulationIPylearn2:machinelearningframeworkforresearchers1/60IntroductionTheanoRNNLSTMExercicesPythonIGeneral-purposehigh-levelOOinterpretedlanguageIEmphasizescodereadabilityIComprehensivestandardlibraryIDynamictypeandmemorymanagementISlowexecutionIEasilyextensiblewithCIPopularin

5、webdevelopmentandscienticcommunities2/60IntroductionTheanoRNNLSTMExercicesNumPy/SciPyIPythonoatsarefull-edgedobjectsontheheapINotsuitableforhigh-performancecomputing!INumPyprovidesann-dimensionalnumericarrayinPythonIPerfectforhigh-performancecomputingISlicesofarraysareviews(nocopying

6、)INumPyprovidesIElementwisecomputationsILinearalgebra,FouriertransformsIPseudorandomnumbergenerators(manydistributions)ISciPyprovideslotsmore,includingISparsematricesIMorelinearalgebraISolversandoptimizationalgorithmsIMatlab-compatibleI/OII/Oandsignalprocessingforimagesandaudio3/60Intro

7、ductionTheanoRNNLSTMExercicesWhat'smissing?INon-lazyevaluation(requiredbyPython)hurtsperformanceIBoundtotheCPUILackssymbolicorautomaticdierentiationINoautomaticspeedandstabilityoptimization4/60IntroductionTheanoRNNLSTMExercicesGoalofthestackFasttodevelopFasttorun5/60Introducti