返回
Transfer Learning using PyTorch — Part 2 – Vishnu Subramanian – Medium英文学习资料

Transfer Learning using PyTorch — Part 2 – Vishnu Subramanian – Medium英文学习资料

ID：38925596

大小：2.78 MB

页数：7页

时间：2019-06-21

Transfer Learning using PyTorch — Part 2 – Vishnu Subramanian – Medium英文学习资料_第1页
Transfer Learning using PyTorch — Part 2 – Vishnu Subramanian – Medium英文学习资料_第2页
Transfer Learning using PyTorch — Part 2 – Vishnu Subramanian – Medium英文学习资料_第3页
Transfer Learning using PyTorch — Part 2 – Vishnu Subramanian – Medium英文学习资料_第4页
Transfer Learning using PyTorch — Part 2 – Vishnu Subramanian – Medium英文学习资料_第5页
资源描述：

《Transfer Learning using PyTorch — Part 2 – Vishnu Subramanian – Medium英文学习资料》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、VishnuSubramanianFollowingLifelonglearner.Passionateaboutdeeplearning,distributedcomputing.Apr19·5minreadTransferLearningusingPyTorch—Part2InthepreviousblogwediscussedhowNeuralnetworksusetransferlearningforvariouscomputervisiontasks.Inthisblogwewilllookintothefollowing

2、.1.VGGArchitecture2.FinetuneVGGusingpre-convolutedfeatures3.Accuracy4.PerformancecomparisonbetweenPyTorchandKerasonTensorflowVGGArchitecture:OneofthemoststudiedDeeplearningmodelsfortransferlearningisVGG.WewillgothroughahighleveloverviewofVGGtounderstandhowitcanbeoptimal

3、lyusedintransferlearning.VGGmodelcanbesplitintotwokindsoflogicalblocks1.Convolutionblocks:Thepre-trainedVGGmodelistrainedonImagenetdatasetover1000categories.Theconvolutionalblockcontainsmultipleconvolutionlayers.Theinitiallayerscontainlowlevelfeatureslikelines,curves.T

4、helastconvolutionallayersinthisblockcontainmorecomplexkindoffeaturesofimageslikehand,leg,eyesandmanymore.Thebelowimagecaptureswhatkindoffeaturesarecapturedindifferentlayers.Asyoucanseefromtheaboveimages,thefeaturesbeingcapturedbytheconvolutionlayersofapre-trainedmodelca

5、nbeusedacrossmostkindofimageproblems.Theabovefeaturesmaynotworkforproblemslikecartoonanimations,medicalimagessincetheyneedcompletelydifferentfeatures.Theconvolutionlayersexhibit2importantproperties-1.Thenumberofparametersrequiredisfarlesscomparedtofullyconnectedlayer.Fo

6、rexampleaConvolutionlayerwith3*3*64sizefiltersneedonly576parameters.2.Convolutionlayersarecomputationallyexpensiveandtakelongertocomputetheoutput.2.FullyConnectedBlock:ThisblockcontainsDense(inKeras)/Linear(inPyTorch)layerswithdropouts.ThenumberofparameterstolearninFCla

7、yersarehugebuttakeswaylesstimetocompute.So,wegenerallyenduptakingpreconvolutedfeaturesfromConvolutionblockofVGGmodelasitisandtrainingonlythelastfewlayersoftheVGGmodelwhicharegenerallyfromfullyconnectedblock.FinetuneVGGusingpreconvolutedfeatures:Asweknowthatconvolutionl

8、ayersareexpensivetocalculate,itmakessensetocomputetheoutputoftheconvolutionlayersonceandusethemtotrainthefullyconnect

当前文档最多预览五页，下载文档查看全文

此文档下载收益归作者所有

当前文档最多预览五页，下载文档查看全文
温馨提示：
1. 部分包含数学公式或PPT动画的文件，查看预览时可能会显示错乱或异常，文件下载后无此问题，请放心下载。
2. 本文档由用户上传，版权归属用户，天天文库负责整理代发布。如果您对本文档版权有争议请及时联系客服。
3. 下载前请仔细阅读文档内容，确认文档内容符合您的需求后进行下载，若出现内容与标题不符可向本站投诉处理。
4. 下载文档时可能由于网络波动等原因无法下载或下载错误，付费完成后未能成功下载的用户请联系客服处理。