Modern frameworks for machine learning

  • Published on
    12-Apr-2017

  • View
    281

  • Download
    1

Embed Size (px)

Transcript

<p> PowerPoint</p> <p>Machine Learning in practiceTools and frameworksSergii Nechuiviters.nechuiviter@gmail.comSamsung R&amp;D Institute Ukraine</p> <p>Sergii Nechuiviter. Samsung R&amp;D Institute Ukraine</p> <p>Machine Learning WorkflowWe need to go deeperFirst generation of NN frameworksModern NN frameworkSpeculations and summaryContent2A sample title</p> <p>Sergii Nechuiviter. Samsung R&amp;D Institute Ukraine</p> <p>A sample title3</p> <p>Machine LearningWorkflow</p> <p>Good beginning4Machine Learning Workflow</p> <p>Sergii Nechuiviter. Samsung R&amp;D Institute Ukraine</p> <p>Microsoft really does a lot to create the best Data scientist working environment.4</p> <p>Good Architecture</p> <p>Machine Learning Workflow5</p> <p>ExperimentsDatasetsTrained modelsNotebooksREST APIOther cognitiveservices</p> <p>A bit more advanced options</p> <p>Machine Learning Workflow6</p> <p>Wolfram Data Science platform (Expected soon)</p> <p>AmazoneMachine Learning</p> <p>ML FrameworkImplementationProsscikit-learnPythonThe most popular, standard de-factoAccord .Net Framework.Net C#Best for C#mlpackC++Fastest on CPUMLlib from SparkScalaDistributed out of the boxRapidMinerJavaVast pack of algorithmsGUI workflow designer</p> <p>Generic machine learning toolsMachine Learning Workflow7</p> <p>Sergii Nechuiviter. Samsung R&amp;D Institute Ukraine</p> <p>Too much data to analyze?8A sample title</p> <p>Sergii Nechuiviter. Samsung R&amp;D Institute Ukraine</p> <p>First generation ofNN frameworksA sample title9</p> <p>Sergii Nechuiviter. Samsung R&amp;D Institute Ukraine</p> <p>CaffeBerkeley Vision and Learning CenterSpecialized on convolutional NNConfigs, PythonNo cluster supportGPU support</p> <p>10A sample title</p> <p>Sergii Nechuiviter. Samsung R&amp;D Institute Ukraine</p> <p>Torch7FacebookImperativeLuaJIT + PythonNo cluster supportGPU supportEmbeddable, with ports to iOS, Android and FPGA backends</p> <p>11A sample title</p> <p>Sergii Nechuiviter. Samsung R&amp;D Institute Ukraine</p> <p>TheanoInitiated by LISAlab, now a vast communityComputational graph optimizations: generic and for RNNPythonNo cluster supportGPU supportPros: Easy NN with Keras,LasagneandBlocks; RNN optimizationsCons: Low-level, Error messages can be unhelpful</p> <p>12A sample title</p> <p>Sergii Nechuiviter. Samsung R&amp;D Institute Ukraine</p> <p>ModernNN frameworksModern NN frameworks13</p> <p>DL4J</p> <p>Sergii Nechuiviter. Samsung R&amp;D Institute Ukraine</p> <p>Deeplearning4JSkymindNN specific ? Fast tensor backendJava, Scala, ClojureSpark, Hadoop, Mesos, AkkaMulti-GPUEmbeddable due to JavaPros: Enterprise support</p> <p>DL4J14Modern NN frameworks</p> <p>Sergii Nechuiviter. Samsung R&amp;D Institute Ukraine</p> <p>TheanoInitiated by LISAlab, now a vast communityComputational graph optimizations: generic and for RNNPythonNo cluster supportGPU supportPros: Easy NN with Keras,LasagneandBlocks; RNN optimizationsCons: Low-level, Error messages can be unhelpful</p> <p>15Modern NN frameworks</p> <p>Sergii Nechuiviter. Samsung R&amp;D Institute Ukraine</p> <p>What is computational graph?Modern NN frameworks16</p> <p>Sergii Nechuiviter. Samsung R&amp;D Institute Ukraine</p> <p>Modern NN frameworks17Computational graph optimizations</p> <p>TensorFlowGoogleComputational graph optimizations: generic and for RNNC++, PythonFrom V0.7 cluster supportFrom V0.7 multi-GPULinux and ARM deployment; No WindowsGoogle Cloud Machine Learning, RescalePros: TensorBoard for visualizationCons: slow and bulk</p> <p>18Modern NN frameworks</p> <p>Sergii Nechuiviter. Samsung R&amp;D Institute Ukraine</p> <p>18</p> <p>CNTKMicrosoftComputational graph optimizations: generic and for RNNNetwork Description Language + C/C++Cluster supportMulti-GPU?Azure Machine LearningPro: fastest at ImageNetLSVR2015</p> <p>19Modern NN frameworks</p> <p>Sergii Nechuiviter. Samsung R&amp;D Institute Ukraine</p> <p>ImageNetLSVR2015</p> <p>20Modern NN frameworks</p> <p>Sergii Nechuiviter. Samsung R&amp;D Institute Ukraine</p> <p>NeonNervana?PythonClusterMulti-GPU?Nervana CloudPros: fastest convolutions</p> <p>21Modern NN frameworks</p> <p>Sergii Nechuiviter. Samsung R&amp;D Institute Ukraine</p> <p>convnet-benchmarksGoogleNet V1- Input 128x3x224x2246-core Intel Core i7-5930K CPU @ 3.50GHz+NVIDIA Titan X+Ubuntu 14.04 x86_64https://github.com/soumith/convnet-benchmarks22Modern NN frameworks</p> <p>Sergii Nechuiviter. Samsung R&amp;D Institute Ukraine</p> <p>MXNetDMLCComputational graph optimizations: generic and for RNN + ImperativePython/R/Julia/Go/C++Cluster supportMulti-GPU supportWindows, Linux, Mobile, AppleAWS recommendedPros: versatile, fast as Torch and Theano</p> <p>23Modern NN frameworks</p> <p>Sergii Nechuiviter. Samsung R&amp;D Institute Ukraine</p> <p>TheanoDL4JTensorFlowCNTKNeonMXNetCreatorSkymindGoogleMicrosoftNervanaDMLCComputational GraphV?VV?VCluster supportVVVVVGPUSupport1MultiMultiMultiMultiMultiMobilemay beV??VCloudVVVVVPerformanceOk?SlowFastFastOk</p> <p>Modern NN frameworksA sample title24</p> <p>Sergii Nechuiviter. Samsung R&amp;D Institute Ukraine</p> <p>Speculations and summarySpeculations and summary25</p> <p>Sergii Nechuiviter. Samsung R&amp;D Institute Ukraine</p> <p>Why deep learning happened?</p> <p>Rise of single CPU performance</p> <p>Distributed computingSingle core reach plateauof performance</p> <p>Price of computations is dropping26Speculations and summary</p> <p>Sergii Nechuiviter. Samsung R&amp;D Institute Ukraine</p> <p>Why deep learning happened?</p> <p>Rise of single CPU performance</p> <p>Distributed computingSingle core reach plateauof performance</p> <p>Price of computations is dropping??27Speculations and summaryHuman-Machine cooperation?</p> <p>Sergii Nechuiviter. Samsung R&amp;D Institute Ukraine</p> <p>Try Microsoft Azure Machine Learning FreeTrialTry scikit-learnKnow about computational graphsTry MXNet</p> <p>Be ready for changes!</p> <p>SummarySpeculations and summary28</p> <p>Sergii Nechuiviter. Samsung R&amp;D Institute Ukraine</p> <p>A sample title29</p>

Recommended

View more >