Refractor (for Wine Quality Prediction)

This is the CML port of the Refractor prototype which is part of the Interpretability report from Cloudera Fast Forward Labs.

CML Applications: Train and inspect a new model locally

This project uses the Applications feature of CML (>=1.2) and CDSW (>=1.7) to instantiate a UI frontend for visual interpretability and decision management.

Train a predictor model

A model has been pre-trained and placed in the models directory.
Start a Python 3 Session with at least 8GB of memory and run the utils/setup.py code. This will create the minimum setup to use existing, pretrained models.

If you want to retrain the model run the train.py code to train a new model.

The model artifact will be saved in the models directory named after the datestamp, dataset and algorithm (ie. 20191120T161757_wine_gb). The default settings will create a gradiant boosted classifier model against the wine dataset. However, the code is vary modular and can train multiple model types against essentially any tabular dataset (see below for details).

Deploy Predictor and Explainer models

Go to the Models section and create a new predictor model. The sample features below should predict an "Poor" quality wine.

• Name: Predictor
• Description: Predict wine quality
• File: deploy_model.py
• Function: predict
• Input: {"fixedAcidity":7.4, "volatileAcidity":0.7, "citricAcid":0, "residualSugar":1.9, "chlorides":0.076, "freeSulfurDioxide":11, "totalSulfurDioxide":34, "density":0.9978, "pH":3.51, "sulphates":0.56, "Alcohol":9.4}
• Kernel: Python 3

If you created your own model (see above)

• Click on "Set Environment Variables" and add:
  • Name: MODEL_NAME
  • Value: 20191120T161757_wine_gb your model name from above Click "Add" and "Deploy Model"

Create a new Explainer model. The sample features below should predict an "Excellent" quality wine.

• Name: Explainer
• Description: Explain wine quality prediction
• File: deploy_model.py
• Function: explain
• Input: {"fixedAcidity":5.9, "volatileAcidity":0.550, "citricAcid":0.10, "residualSugar":2.2, "chlorides":0.062, "freeSulfurDioxide":39.0, "totalSulfurDioxide":51.0, "density":0.99512, "pH":3.52, "sulphates":0.76, "Alcohol":11.2}
• Kernel: Python 3

If you created your own model (see above)

• Click on "Set Environment Variables" and add:
  • Name: MODEL_NAME
  • Value: 20191120T161757_wine_gb your model name from above Click "Add" and "Deploy Model"

In the deployed Explainer model -> Settings note (copy) the "Access Key" (ie. mukd9sit7tacnfq2phhn3whc4unq1f38)

Instatiate the flask UI application

From the Project level click on "Open Workbench" (note you don't actually have to Launch a session) in order to edit a file. Select the flask/single_view.html file and paste the Access Key in at line 19. Save and go back to the Project.

Go to the Applications section and select "New Application" with the following:

• Name: Visual Wine Analysis
• Subdomain: wine-prediction
• Script: flask_app.py
• Kernel: Python 3
• Engine Profile: 1vCPU / 2 GiB Memory

If you created your own model (see above)

• Add Environment Variables:
  • Name: MODEL_NAME
  • Value: 20191120T161757_wine_gb your model name from above
    Click "Add" and "Deploy Model"

After the Application deploys, click on the blue-arrow next to the name. The initial view is a table of rows selected at random from the dataset. This shows a global view of which features are most important for the prediction made.

Clicking on any single row will show a "local" interpretabilty of a particular instance. Here you can see how adjusting any one of the features will change the instance's prediction.

Additional options

By default this code trains a gradiant boosted classifier for the wine dataset.
There are other datasets and other model types as well. Look at run_experiment.py for examples or set the Project environment variables to try other datasets and models:
Name Value
DATASET ibm (default) | breastcancer | iris
MODEL_TYPE linear (default) | gb | nonlinear | voting

NOTE that not all of these options have been fully tested so your mileage may vary.