/UKY-MSP-Azure-ML

University of Kentucky MSP - Azure Machine Learning Workshop

This readme contains a tutorial for attendees of the University of Kentucky MSP Azure Machine Learning Workshop. To register, visit this link: https://bit.ly/AzureMLUKY

Step 1: Register for Azure for Students

Azure for Students gives K-12 and higher education students across the globe access to hundreds of Azure resources, including those needed for Azure Machine Learning Studio.

Step 2: Download and Import "train.csv" Dataset

Follow the link: ---------- to download the "train.csv" dataset. This is the dataset which we will use to train and test the model. Select "New" then select Dataset and follow the on-screen instructions to upload the downloaded dataset from your computer. Then under the "Select Datasets" experiment item, drag your "train.csv" tab into the workspace.

Step 3: Exclude the unused Columns

DELETE: All the ML Steps (but not pretty)

  1. Create a new experiment

  2. Add "select columns in dataset" -> "Exclude: PassengerId, Name, Ticket, Cabin"

  3. Add "clean missing data" -> Cleaning mode: Remove entire row

  4. Add "Edit metadata" -> Pclass, Sex, Survived, Embarked -> Categorical: Make categorical

  5. Add "Edit metadata" -> Survived -> Fields: label

  6. Add "Split Data" -> Fraction of rows = 0.7 -> Stratified split: True, on "Survived"

  7. Add 2 "Train model" -> select column: Survived

  8. Add two-class neural network and two-class SVM, wire to train model

  9. Add 2 "Score model"

  10. Add "evaluate model"

  11. Run and visualize output -> show which model is more accurate

  12. Set up web service -> Predictive web service [Recommended]

  13. Run predictive experiment

  14. Deploy web service

  15. Next to "Request/Response" , click "Test"

  16. Submit a random sample of data

  17. Look at the last part of the JSON that's returned: integer 0=died, 1=survived, decimal is probability

Steps on New Platform:

  1. Visit aka.ms/AzureForStudents
  2. Click Activate Now and sign in with your LinkBlue credentials
  3. Enter your email address and phone number when prompted - you may have to verify one of them
  4. Open the azure portal
  5. Click the menu on the left-hand side, go to resource groups
  6. Add a new resource group, we'll call it "MSPAprilML"
  7. Review+Create, then Create
  8. Now, in portal search bar, go to "Machine Learning"
  9. Give it a name, choose your resource group you just made, choose pricing "ENTERPRISE" Note: You won't be billed - this draws from your Azure for Students credit
  10. Review + Create, then Create [ ~ 3 minutes ]

Let's go get the dataset

  1. Go to https://bit.ly/TitanicDataML
  2. From Excel online, go to File -> Save As -> Download a Copy : Let's look at the CSV and see what data we have
  3. Go back to Azure portal (portal.azure.com), search for Machine Learning and open that panel
  4. Open your ML workspace resource
  5. Click "Launch Now" inside the New Machine Learning Studio dialog box
  6. You may have to sign back in
  7. Manage Compute
  8. Training Clusters -> New -> VM Size: STANDARD_D11_V2 -> Min number: 1 -> Max number: 2 -> CREATE
  9. Inference Clusters -> New -> Central US -> DevTest -> Create

Dataset Upload

  1. Designer -> Prebuilt -> give it a name
  2. First, we need to upload our dataset. From the left sidebar, go to "Dataset" this is a GUI that allows us to build interactive ML. Create Dataset -> From local files. Give it a name Leave tabular Leave datastore option default, browse to downloaded file and upload Column headers -> choose "use headers from first files" Next all the way through
  3. Back to designer, open previous pipeline
  4. Select compute target -> choose what you created earlier
  5. Drop down datasets, drag in in titanic
  6. Drag in "select column in dataset", wire titanic output to input of this module, then edit "select columns" for the following: ALL EXCEPT:
  • PassengerId, Name, Ticket, Cabin
  1. Submit, select compute target, create new, predefined (give it a name), submit
  2. Submit, create new experiment, give name, submit On the backend, Azure is dynamically assigning a cluster of virtual machines to run your experiment. So, this may take a LOT of time.
  3. Add "clean missing data" -> Select columns: edit columns (By Name: Add all) -> Cleaning mode: Remove entire row, submit
  4. Add "Edit metadata" -> Pclass, Sex, Survived, Embarked -> Categorical: Make categorical :: SUBMIT
  5. Add "Edit metadata" -> Survived -> Fields: label :: SUBMIT
  6. Add "split data" -> stratified split: true, on "Survived" -> 0.7 :: SUBMIT
  7. Add 2 "Train Model" blocks -> Edit label column: survived
  8. Add "two-class decision forest" and "2-class SVM", wire to train model blocks
  9. Add 2 "Score Model" blocks and wire
  10. Add an "Evaluate Model" block and wire :: SUBMIT
  11. Click on "Evaluate Model" -> Output -> We can visualize and see evaluation results
  12. Click on higher scored "train model" -> create real-time inference pipeline
  13. Once in real-time inference pipeline, delete Evaluate Model (since we don't have two models to compare) :: SUBMIT
  14. Once run is finished, DEPLOY
  15. Choose compute target, click deploy
  16. Once done, "view live endpoint" -> Test
  17. Enter some data, see results!