Applied Machine learning using H2O, python and R Workshop
3 likes1,810 views
The document provides an overview of H2O.ai, including its company background, products, and features related to machine learning. It outlines installation procedures, usage with various languages, data munging techniques, and model training with examples from machine learning problems like price prediction and classification. Key highlights include the platform's analytics capabilities, collaborative resources, and community engagement aspects.
![What we covered Data munging in H2O with python
• H2O and Python
• Jupyter notebook Demo
• Data import
• Row, column, data frames, slicing, binding, exporting, factoring
• Using functions
Q
Reference:
[1] http://localhost:8888/notebooks/H2O%20Data%20Ingest%20Demo.ipynb
[2] http://localhost:8888/notebooks/H2O%20Frame%20manipulation.ipynb](https://image.slidesharecdn.com/h2o-workshop-170227171429/85/Applied-Machine-learning-using-H2O-python-and-R-Workshop-28-320.jpg)
![What we covered
• Response Column - Price
• buildModel 'glm', {"model_id":"glm-2fcdf58a-da59-4b3f-8a68-
1177bce9531c","training_frame":"kc90_house_data.hex","nfolds":"10","seed":-
1,"response_column":"price","ignored_columns":["id"],"ignore_const_cols":true,"family":"gaussian","solver":"AUTO","alpha":[],"lam
bda":[],"lambda_search":false,"standardize":true,"non_negative":false,"fold_assignment":"AUTO","score_each_iteration":false,"co
mpute_p_values":false,"remove_collinear_columns":false,"max_iterations":-
1,"link":"family_default","max_runtime_secs":0,"keep_cross_validation_predictions":false,"keep_cross_validation_fold_assignment":
false,"missing_values_handling":"MeanImputation","intercept":true,"objective_epsilon":-
1,"beta_epsilon":0.0001,"gradient_epsilon":-1,"prior":-1,"max_active_predictors":-1}
o r2 0.017224
• buildModel 'glm', {"model_id":"glm-2fcdf58a-da59-4b3f-8a68-
1177bce9531c","training_frame":"kc90_house_data.hex","nfolds":"10","seed":-
1,"response_column":"price","ignored_columns":["id"],"ignore_const_cols":true,"family":"gaussian","solver":"AUTO","alpha":[0.001
],"lambda":[0.1],"lambda_search":false,"standardize":false,"non_negative":false,"fold_assignment":"AUTO","score_each_iteration":
false,"compute_p_values":false,"remove_collinear_columns":false,"max_iterations":-
1,"link":"family_default","max_runtime_secs":0,"keep_cross_validation_predictions":false,"keep_cross_validation_fold_assignment":
false,"missing_values_handling":"MeanImputation","intercept":true,"objective_epsilon":-
1,"beta_epsilon":0.0001,"gradient_epsilon":-1,"prior":-1,"max_active_predictors":-1}
o r2 0.619503](https://image.slidesharecdn.com/h2o-workshop-170227171429/85/Applied-Machine-learning-using-H2O-python-and-R-Workshop-32-320.jpg)
Applied Machine learning using H2O, python and R Workshop
- 1. Avkash Chauhan ([email protected]) VP, Enterprise Customers
- 2. Agenda • H2O Intro • Installation • Using H2O from FLOW, R & Python • Data munging in H2O with Python • 2 examples of machine learning problems o GBM, GLM, DRF o Understanding Models, improvements, • Machine learning production pipeline H2O.ai is a Visionary in the Gartner Magic Quadrant for Data Science Platforms
- 4. H2O.ai Company Overview Founded 2011 Venture-backed, debuted in 2012 Products • H2O Open Source In-Memory AI Prediction Engine • Sparkling Water • STEAM • DEEP WATER Mission Operationalize Data Science, and provide a platform for users to build beautiful data products Team 60+ employees worldwide • CA, NY, UT, Japan, UK • Distributed Systems Engineers doing Machine Learning • World-class visualization designers Headquarters Mountain View, CA
- 6. Customers and Use Cases Financial Insurance MarketingTelecom Healthcare
- 7. Open Source Users & Community H2O Users List: http://www.h2o.ai/user-list/
- 8. About Myself • VP – Enterprise products & customers – Handling paid enterprise customer’s requirements – building product(s) – Helping community • LinkedIn: https://www.linkedin.com/in/avkashchauhan/ • Blog: https://aichamp.wordpress.com/ • Community: https://community.h2o.ai/index.html • Twitter: @avkashchauhan
- 11. Key features • Open Source (Apache 2.0) • All supported ML algorithms are coded by our engineers • Designed for speed, scalability and for super large data-sets • Same distribution for open source community & enterprise • Very active production, every other week release • Vibrant open source community o https://community.h2o.ai • Enterprise Support portal o https://support.h2o.ai • We have 70,000 users, 8,000 organizations and growing daily
- 12. Usage: Simple Solution o Single Deployable compiled Java code (jar) o Ready to use point and click FLOW Interface o Connection from R and Python after specific packages are installed o Use Java, Scala natively and any other language through RESTful API o Deployable models - Binary & Java (POJO & MOJO) o One click prediction/scoring engine
- 13. Usage: Complex Solution o Multi-node Deployment o Spark and Hadoop distributed environment • Sparkling Water (Spark + H2O) o Data ingested from various inputs • S3, HDFS, NFS, JDBC, Object store etc. • Streaming support in Spark (through Sparking Water) o Distributed machine learning for every algorithm in platform o Prediction service deployment on several machines
- 14. Current Algorithm Overview Statistical Analysis • Linear Models (GLM) • Naïve Bayes Ensembles • Random Forest • Distributed Trees • Gradient Boosting Machine • R Package - Stacking / Super Learner Deep Neural Networks • Multi-layer Feed-Forward Neural Network • Auto-encoder • Anomaly Detection Clustering • K-Means Dimension Reduction • Principal Component Analysis • Generalized Low Rank Models Solvers & Optimization • Generalized ADMM Solver • L-BFGS (Quasi Newton Method) • Ordinary Least-Square Solver • Stochastic Gradient Descent Data Munging • Scalable Data Frames • Sort, Slice, Log Transform • Data.table (1B rows groupBy record)
- 16. JobFluid Vector Frame MRTaskDistributed K/V Store Distributed Fork/JoinNon-Blocking Hash Table Distributed In-Memory Processing Core H2O: Architecture REST / JSON Parse Exploratory Analysis Feature Engineering ML Algorithms Model Evaluation Scoring Data/Model Export SQL NFS Local S3 HDFS POJO Production Environments
- 17. Sparkling Water - High Level Architecture
- 18. Deep Water : Architecture Node 1 Node N Scala Spark H2O Java Execution Engine TensorFlow/mxnet/Caffe C++ GPU CPU TensorFlow/mxnet/Caffe C++ GPU CPU RPC R/Py/Flow/Scala client REST API Web server H2O Java Execution Engine grpc/MPI/RDMA Scala Spark
- 20. What we covered with H2O Installation • Installation, using H2O with R & Python o Installation H2O • Help – help(h2o.init), h2o.cluster_status() o H2O Github repo • Source code - glance o R package installation o Python Package Installation o Connecting H2O from R o Connecting H2O from Python Q
- 21. H2O FLOW DEMO Q
- 22. What we covered in FLOW DEMO • FLOW Intro • Running Examples • Generating Data • Working with UI, Cell, Running FLOW Script • Importing Data o Chunk Distribution o Feature analysis • Building models from imported data • Understanding models o Binary Model, POJO, MOJO • Listing all Jobs • Using HELP • Understanding RESTful Interface • Reading Logs, Water Meter (CPU analysis), Stack Trace etc. Q
- 23. Data manipulation between H2O, R & python Q
- 24. Data manipulation between H2O, R, python • Import data in between H2O, python, R and others.. Q H2O R, python, Java, Scala, etc.. 3600 View
- 25. Data manipulation between H2O, R, python • Import data in python o import pandas as pd • datasetpath = "/Users/avkashchauhan/tools/datasets/kaggle-imdb-word2vec” • train_data = pd.read_csv(datasetpath + "/labeledTrainData.tsv", header=0, delimiter="t", quoting=3) o h2o.ls() o train_data.describe() o train_data_h2o = h2o.H2OFrame(train_data, destination_frame = "train_data_h2o") o h2o.ls() o train_data_h2o.describe() o Now Look the same frame in the FLOW - getFrames • In R • > h2o.ls() • > rdf = h2o.getFrame("train_data_h2o") • > rdf • > summary(rdf) Q Part 1 Reference: http://localhost:8888/notebooks/H2O-start-test-and-data-switch-python-r.ipynb
- 26. Data manipulation between H2O, R, python • Import data in R o > iris o > mydf = as.h2o(iris) • This frame will be imported as iris original frame name o > summary(mydf) o > summary(iris) o > h2o.ls() • You will see the iris entry as h2o frames list • Check FLOW as well and you will see iris there too o > mydf = as.h2o(iris, destination_frame = “mydf”) o > h2o.ls() o In Python • h2o.ls() • my_python_df= h2o.get_frame("mydf") • my_python_df • h2o.ls() Q Part 2
- 27. Data munging in H2O with Python Q
- 28. What we covered Data munging in H2O with python • H2O and Python • Jupyter notebook Demo • Data import • Row, column, data frames, slicing, binding, exporting, factoring • Using functions Q Reference: [1] http://localhost:8888/notebooks/H2O%20Data%20Ingest%20Demo.ipynb [2] http://localhost:8888/notebooks/H2O%20Frame%20manipulation.ipynb
- 29. Price Prediction using GLM, GBM, DRF Q
- 30. Problem Description • Kaggle o https://www.kaggle.com/harlfoxem/housesalesprediction • Local Datasets o /Users/avkashchauhan/learn/seattle-workshop/kc_house_data.csv o /Users/avkashchauhan/learn/seattle-workshop/kc_house_orig.csv • Documentation o http://docs.h2o.ai/h2o/latest-stable/h2o-docs/index.html
- 31. What we covered • Data Import • Understood data, frame distribution, chunk, compression etc. • Understood all features, through histograms, enums, etc. • Split data frames for train and test • Converting features to Factors/Enum • Imputation of values • Training o with Training frame only o With cross-validation o with validation frame • Understanding Model details • Documentation o http://docs.h2o.ai/h2o/latest-stable/h2o-docs/index.html
- 32. What we covered • Response Column - Price • buildModel 'glm', {"model_id":"glm-2fcdf58a-da59-4b3f-8a68- 1177bce9531c","training_frame":"kc90_house_data.hex","nfolds":"10","seed":- 1,"response_column":"price","ignored_columns":["id"],"ignore_const_cols":true,"family":"gaussian","solver":"AUTO","alpha":[],"lam bda":[],"lambda_search":false,"standardize":true,"non_negative":false,"fold_assignment":"AUTO","score_each_iteration":false,"co mpute_p_values":false,"remove_collinear_columns":false,"max_iterations":- 1,"link":"family_default","max_runtime_secs":0,"keep_cross_validation_predictions":false,"keep_cross_validation_fold_assignment": false,"missing_values_handling":"MeanImputation","intercept":true,"objective_epsilon":- 1,"beta_epsilon":0.0001,"gradient_epsilon":-1,"prior":-1,"max_active_predictors":-1} o r2 0.017224 • buildModel 'glm', {"model_id":"glm-2fcdf58a-da59-4b3f-8a68- 1177bce9531c","training_frame":"kc90_house_data.hex","nfolds":"10","seed":- 1,"response_column":"price","ignored_columns":["id"],"ignore_const_cols":true,"family":"gaussian","solver":"AUTO","alpha":[0.001 ],"lambda":[0.1],"lambda_search":false,"standardize":false,"non_negative":false,"fold_assignment":"AUTO","score_each_iteration": false,"compute_p_values":false,"remove_collinear_columns":false,"max_iterations":- 1,"link":"family_default","max_runtime_secs":0,"keep_cross_validation_predictions":false,"keep_cross_validation_fold_assignment": false,"missing_values_handling":"MeanImputation","intercept":true,"objective_epsilon":- 1,"beta_epsilon":0.0001,"gradient_epsilon":-1,"prior":-1,"max_active_predictors":-1} o r2 0.619503
- 33. GLM - Regularization • L1 Lasso and L2 Ridge: o Regularization is used solve problems with overfitting in GLM. o Penalties are introduced to avoid overfitting, o To reduce variance of the prediction error o To handle correlated predictors. • H2O - Elastic Net o Alpha (0 – 1)/ Lambda(0 – 1, into very small fraction i.e. 0.0001) o Alpha = 0 –> Ridge o Alpha = 1 -> LASSO o Lambda – 0.0 > No regularization • Lambda Search o Enable Lambda_Search o lambda_min_ratio o Nlambdas o max_active_predictors • Look for Intercept, and p-values in the docs too • Documentation: http://docs.h2o.ai/h2o/latest-stable/h2o-docs/data-science/glm.html#regularization
- 34. Linear Regression - Prediction • Import data to predict • Feature Engineering o Proper encoding for numerical data o Hide – ID and other features • Experimentation with Alpha and Lambda o 0.001 & 0.1 – 0.61 – Did Prediction o 0.0001 & 0.1 - 0.61 o 0.0001 & 0.5 – 0.58 o 0.001 & 0.01 - 0.67 – Did Prediction o 0.004 & 0.01 – 0.70 – Did prediction • Understanding Training and validation – r2 values
- 35. Price Prediction - GBM • Feature Engineering o Proper encoding for numerical data o Hide – ID and other features - All default - Training r2 = 0.95 & Validation r2 = 0.76 – Try prediction - Learning Rate – 0.5 + all default = 0.98/0.71 - Learning Rate – 0.01 + all default = 0.50/0.40 - Setting – stopping rounds
- 36. Understanding GLM, GBM, DRF - Validation and Cross Validation - Scoring History, Validation History (Validation Frame/CV) - Training and Validation Metrics - Using Stopping metrics into Tree based algorithms – DRF/GBM - How adding tree depth changes the results? - Variable Importance - Gains and Lift Chart - Confusion Matrix (TPR/FPR) - Solvers Note: Using 3 FLOW interface build – GLM, GBM, DRF Model
- 37. Improving Overall Results - Feature Engineering - Adding proper categories - Year, Waterfront, view, condition, grade, zipcode – Factors - How r2 values helps better prediction - GBM Improvements with CV - 0.3 Learning, all default – 0.795 >> Perform Prediction - Ntree=60, depth=5,l-rate=0.21,row-sr=0.8, col-sr=0.8> 0.80 - GBM with Validation frame - Ntree=60, depth=5,l-rate=0.21,row-sr=0.8, col-sr=0.8> 0.82 - DRF with CV and default settings > 0.80 - Finally now – Ignore Date column - DRF - 0.86 - GBM – 0.86 - With Update Feature Engineering - GLM – Aplha:0.004, lambda:0.01 > r2 = 0.77
- 38. R - Demo - Start R - > h2o.init() - > h2o.clusterStatus() - ?h2o.gbm > Use the sample - > h2o.varimp(gbm) - > h2o.varimp_plot(gbm) - Performance - > perf <- h2o.performance(gbm, australia.hex) - > h2o.mse(perf) - > h2o.r2(perf) - Add CV – Rebuild the model with nfolds = 5 - > h2o.cross_validation_models(gbm) - Kmeans - ?h2o.kmeans - Run the example - Want to eval the K? - kmodel = h2o.kmeans(training_frame = prostate.hex, k = 10, x = c("AGE", "RACE", "VOL", "GLEASON")) - kmodel = h2o.kmeans(training_frame = prostate.hex, k = 100, x = c("AGE", "RACE", "VOL", "GLEASON"), estimate_k = T)
- 39. Supplement Information - https://www.kaggle.com/harlfoxem/d/harlfoxem/housesalespre diction/house-price-prediction-part-1/discussion - https://www.kaggle.com/auygur/d/harlfoxem/housesalespredict ion/step-by-step-house-price-prediction-r-2-0-77/code - https://rpubs.com/MagicSea/property_price
- 40. Solving a Binomial Classification problem Q
- 41. Binomial Classification Problem Description • Titanic Survivors list o Passengers on board: 1317, others were crew of total 2224 • Kaggle o https://www.kaggle.com/c/titanic • Download Dataset: o http://biostat.mc.vanderbilt.edu/wiki/pub/Main/DataSets/titanic3.xls • Local Datasets o /Users/avkashchauhan/learn/seattle-workshop/titanic_list.csv • Documentation o http://docs.h2o.ai/h2o/latest-stable/h2o-docs/index.html
- 42. Predicting Titanic Survival Rate • Reference: o Using scikit-learn • http://localhost:8888/notebooks/Titanic%20Survival%20Dem o%20in%20Python.ipynb o Using H2O python estimators • http://localhost:8888/notebooks/Titanic%20Survival%20Dem o%20in%20H2O%20and%20Python.ipynb
- 43. What we learn– ROC Curve AUC of 0.5 is random and 1 is perfect
- 44. What we did • GLM/GBM/RF in Python + scikit-learn • GLM in H2O from Python • Grid Search • Working in FLOW o Ingest, split frame o GLM in H2O from FLOW • Response must be number – Understanding r2 o GBM in FLOW • Response numeric– Regression – r2 • Response enum – classification – AUC (Area under the curve) – Confusion matrix – ROC Curve – NEXT PAGE • AUC o AUC of 0.5 is random and 1 is perfect o Improve AUC
- 45. What we covered • Data Import • Understood data, frame distribution, chunk, compression etc. • Understood all features, through histograms, enums, etc. • Split data frames for train and test • Converting features to Factors/Enum • Imputation of values
- 47. Supported Model Types in H2O • Binary • POJO • MOJO exportModel "word2vec-0dfb7bfd-7a5d-42a9-9ebf-82706304a4fe", "/Users/avkashchauhan/Downloads/word2vec-0dfb7bfd-7a5d-42a9-9ebf- 82706304a4fe", overwrite: true
- 48. POJO Demo • Run “GBM_Airlines_Classification” demo from FLOW examples • Download POJO from FLOW • Create a temp folder/Be in temp folder o Move gbm_pojo_test.java • Get Model from RESTful API – $ curl -X GET "http://localhost:54321/3/Models.java/gbm_pojo_test" >> gbm_pojo_test_1.java o Get H2O gen-model • curl http://localhost:54321/3/h2o-genmodel.jar > h2o-genmodel.jar o Create main.java o Compile • $ javac -cp h2o-genmodel.jar -J-Xmx2g -J-XX:MaxPermSize=128m gbm_pojo_test.java main.java • Verify with compiled class files o Run • $ java -cp .:h2o-genmodel.jar main • See Results • Docs: https://github.com/h2oai/h2o-3/blob/master/h2o-docs/src/product/howto/POJO_QuickStart.md
- 49. MOJO Demo • Run “GBM_Airlines_Classification” demo from FLOW examples • Download POJO from FLOW • Create a temp folder/Be in temp folder o Move gbm_pojo_test.zip • Get Model from RESTful API – $ curl -X GET "http://localhost:54321/3/Models.java/gbm_pojo_test.zip" >> gbm_pojo_test.zip o Get H2O gen-model • curl http://localhost:54321/3/h2o-genmodel.jar > h2o-genmodel.jar o Create main_mojo.java o Compile • $ javac -cp h2o-genmodel.jar -J-Xmx2g -J-XX:MaxPermSize=128m gbm_pojo_test.java main_mojo.java – Make sure main method name is main_mojo • Verify with compiled class files o Run • $ java -cp .:h2o-genmodel.jar main • See Results • Docs: https://github.com/h2oai/h2o-3/blob/master/h2o-docs/src/product/howto/MOJO_QuickStart.md
- 50. Thank you so much