From workflows to pipelines
1. From workflows to pipelines
In the first chapter, you learned what a typical machine learning workflow looks like. A properly structured workflow is very similar to the software concept of a pipeline: a series of data processing elements processed in sequence. In this lesson you will learn to turn your workflows into solid, super-efficient pipelines.2. Revisiting our workflow
Consider the following script. First, we import a random forest and split the data into training and test. We then use grid search cross-validation on the training set to tune the maximum depth of the random forest. Using the best depth, we select the best three features and estimate accuracy on the test data. But what if we also wanted to tune the number of estimators used by the random forest, as well as the number of features selected? Should we split the data further? Does the order in which we optimize the parameters matter?3. The power of grid search
Well, you could start by optimizing tree depth. As shown in the figure on the right, you would check three possible values for depth and determine 10 is best. Meanwhile, you kept the number of estimators fixed at its default value, namely 10.4. The power of grid search
To tune the number of estimators, you can now hold depth fixed at 10 instead. This determines 30 estimators to perform best. As shown in the Figure, this sequential workflow has only checked 5 out of 9 possible combinations of values for depth and number of estimators.5. The power of grid search
Instead, by simply expanding the grid dictionary to include both parameters, you can optimize across all possible combinations.6. Pipelines
Grid search can jointly optimize any classifier parameters, but what about other parameters in your workflow, like the number of features, k? This is not as easy because these parameters live in different objects.7. Pipelines
The Pipeline module solves this problem by wrapping an entire workflow in a single object. This means you can now use GridSearchCV() to optimize the hyperparameters from all the steps in your pipeline in one go!8. Pipelines
Start by creating a pipeline object with named steps: in this case, feature selection comes first, followed by the classifier. Next, you have to create a parameter grid just like when tuning a single classifier. You will have to name the keys to match the way a pipeline object internally represents its parameters: use the name of the step in question, followed by the name of the parameter in question, separated by a double underscore. For example, classifier-double underscore-max_depth refers to the parameter max_depth in the classifier step of the pipeline. The value of each key is a list of candidate values for that parameter. You can then input the pipeline object and this parameter grid to GridSearchCV() as usual.9. Customizing your pipeline
A pipeline with grid search CV is an extremely flexible and bug-free workflow. You can also use performance metrics other than accuracy, by overriding the default value of the optional parameter "scoring" in grid search CV. Let's say you want to use AUC. You first need to wrap it inside a scorer object, so that it can accept a classifier and data as its input, rather than the predicted labels. You can do this using make_scorer(). You then input the resulting object as the scoring argument in grid search CV.10. Don't overdo it
Just remember that grid search will fit one classifier for each value combination. Checking 3 values for each of 3 hyperparameters using 10-fold cross-validation will fit the classifier 270 times!11. Supercharged workflows
Time to practice supercharging your workflows with pipelines.Create Your Free Account
or
By continuing, you accept our Terms of Use, our Privacy Policy and that your data is stored in the USA.