108-2 Machine Learning (機器學習) - GitHub

李宏毅教授負責授課與作業規劃。

HW1 Linear Regression. 資料: 行政院環境環保署空氣品質監測網所下載的觀測資料目的: 本作業 ... Skiptocontent {{message}} IPINGCHOU / NTU_MachineLearning Public Notifications Fork 8 Star 22 108-2MachineLearning(機器學習) 22 stars 8 forks Star Notifications Code Issues 1 Pullrequests 0 Actions Projects 0 Wiki Security Insights More Code Issues Pullrequests Actions Projects Wiki Security Insights IPINGCHOU/NTU_MachineLearning Thiscommitdoesnotbelongtoanybranchonthisrepository,andmaybelongtoaforkoutsideoftherepository. master Branches Tags Couldnotloadbranches Nothingtoshow {{refName}} default Couldnotloadtags Nothingtoshow {{refName}} default 1 branch 4 tags Code Latestcommit   Gitstats 107 commits Files Permalink Failedtoloadlatestcommitinformation. Type Name Latestcommitmessage Committime HW10_Anomaly_Dection     HW11_GAN     HW12_Transfer_Learning     HW15_Reinforcement_Learning     HW1_LinearRegression     HW2_Classification     HW3_CNN     HW4_RNN     HW5_ExplainableAI     HW6_Adversarial_attack     HW8_Seq2Seq     HW9_Unsupervised_Learning     README.md     Viewcode NTU_MachineLearning HW1LinearRegression HW2BinaryClassification HW3CNN-FoodClassification HW4RecurrentNeuralNetworks-TextSentimentClassification HW5ExplainableAI-onCNNmodel(HW3) HW6Adversarialattack-onblackboxCNNmodel HW7NetworkCompression HW8Seq2seq-en2cn HW9UnsupervisedLearning HW10AnomalyDetection HW11GAN HW12TransferLearning HW13MetaLearning HW14Life-longLearning HW15ReinforcementLearning README.md NTU_MachineLearning 108-2MachineLearning(機器學習) 由李宏毅教授、吳沛遠教授和林宗男教授共同合授。

李宏毅教授負責授課與作業規劃。

HW1LinearRegression 資料:行政院環境環保署空氣品質監測網所下載的觀測資料 目的:本作業實作LinearRegression預測出PM2.5的數值 實作LinearRegression並手刻Adagrad以及Adamoptimizer實現梯度下降。

Simplebaseline√(ByLinearRegression) Strongbaseline√(ByLinearRegression) HW2BinaryClassification 資料:Census-Income(KDD)Dataset 目的:whethertheincomeofanindividualexceeds$50000ornot? 實作LogisticRegression以及ProbabilisticGenerativemodel，其中LogisticRegression以Cross-Entropy作為lossfunction並實現梯度下降 Simplebaseline√(BothLogisticRegressionandProbabilisticGenerativemodel) Strongbaseline√(ByasimpleNNmodel) HW3CNN-FoodClassification 資料：此次資料集為網路上蒐集到的食物照片，共有11類 Bread,Dairyproduct,Dessert,Egg,Friedfood,Meat,Noodles/Pasta,Rice,Seafood,Soup,andVegetable/Fruit. Trainingset:9866張 Validationset:3430張 Testingset:3347張 目的：食物分類 實作CNN模型 Simplebaseline√(ByasimpleCNNmodel) Strongbaseline√(ByaclassicVGG-16model) HW4RecurrentNeuralNetworks-TextSentimentClassification 資料：為Twitter上收集到的推文，每則推文都會被標注為正面或負面 除了labeleddata以外，還額外提供了120萬筆左右的unlabeleddata labeledtrainingdata：20萬 unlabeledtrainingdata：120萬 testingdata：20萬（10萬public，10萬private） 目的：語句分類 實作RNN模型，使用LSTM Simplebaseline√(ByasimpleLSTMmodel,withlabeledtrainingdata) Strongbaseline√(ByabidirectionalLSTMdeepmodelwithmorefclayers,finetuned,withbothlabelandunlableddata,over90wsentences.) HW5ExplainableAI-onCNNmodel(HW3) 資料：hw3CNNmodeloutput 目的：視覺化CNNmodeloutput Task1-SailencyMap Task2-FilterVisualization Task3-Lime Task4-Anyvisualization/explainingmethodyoulike-(SHAP) HW6Adversarialattack-onblackboxCNNmodel 資料：food-11dataset 目的：attackblackboxmodelbyFGSM-attackoranyotherattackforbetterresult Proxymodel:(pytorch-pretrained) VGG-16 VGG-19 ResNet-50 ResNet-101 DenseNet-121 DenseNet-169 實作：以FGSM通過simplebaseline，並以任意方法通過strongbaseline Simplebaseline√(ByFGSMattack,eps=0.3,acc=0.07,L-inf=16.7250) Strongbaseline√(BybasiciterativemethodFGSM(I-FGSM),eps=0.03,alpha=0.005,acc=0,L-inf=8.4000) AlsotriedOne-Pixel-Attack,butwithslowcomputationspeedandbadresult:( HW7NetworkCompression pass HW8Seq2seq-en2cn 資料：manythings之cmn-eng train:18000sents valid:500sents test:2636sents 目的：英翻中，每次輸入單一句子 實作：利用GRU建立Seq2Seq模型，分別實現以下task並比較。

TeacherForcing AttentionMechanism BeamSearch ScheduleSampling 此次並無baseline，BeamSearch以heapq實現，Schedulesampling選擇使用Linear,Exponential,InverseSigmoid實現。

可參見:https://arxiv.org/abs/1506.03099 HW9UnsupervisedLearning 資料：? train:8500pics,32323 valX:500pics,32323 valY:labels,500 目的:將照片分為風景與非風景照 實作:建立autoencoder，並對其降維後分群 Simplebaseline√(Bytutorial,simpleCNNlayers,acc=0.74918) Strongbaseline√(BydeeperCNNmodleandbmforeachlayer,acc=0.78941) U-Net,simpleU-Netwithbm,betterreconstructionbutloweracc:((acc=0.73789) HW10AnomalyDetection HW11GAN 資料：facesbyCrypko,https://crypko.ai/#/ train:10000pics,64643 目的：訓練出GAN 實作：DCGAN,WGAN 此次作業無Kaggle，實作DCGAN觀察modecollapse以及實作WGAN避免modecollapse。

詳細列於report.pdf之中 HW12TransferLearning 資料：? train:5000pics,32323,10labels,realRGBimages target:100,000pics,28281,nolabels,graffitiimages 目的：只訓練在trainlabel上，在不需要targetlabel的情況下提升targetlabel的預測準確度。

實作：DaNN,MSDA Simplebaseline√(BytutorialDaNN,Source:Cannytransfer,acc=0.57270) Strongbaseline√(ByMSDA,Source:Canny,Sobel,Laplacian,Graytransfer,acc=0.75790) MSDAreference:https://github.com/VisionLearningGroup/VisionLearningGroup.github.io/tree/master/M3SDA/code_MSDA_digit HW13MetaLearning pass HW14Life-longLearning pass HW15ReinforcementLearning 資料:gym-LunarLander 目的：以PGbasedalgorithm讓登陸艇成功登陸月球~ 實作： PG discountrewardwithPG A2C discountrewardwithnstepinfoA2C About 108-2MachineLearning(機器學習) Resources Readme Stars 22 stars Watchers 3 watching Forks 8 forks Releases 4 hw9models Latest May14,2020 +3releases Packages0 Nopackagespublished Languages Python 99.7% Shell 0.3% Youcan’tperformthatactionatthistime. Yousignedinwithanothertaborwindow.Reloadtorefreshyoursession. Yousignedoutinanothertaborwindow.Reloadtorefreshyoursession.