Skip to content

Latest commit

 

History

History
91 lines (67 loc) · 7.02 KB

tutorial2-create-your-component.md

File metadata and controls

91 lines (67 loc) · 7.02 KB

Tutorial 2: Create your own component

In tutorial 1, you have created a pipeline which is constructed by built-in modules and existing components in this gallery. In this tutorial, you will continue this story to introduce how to create your own component.

Component example

In tutorial 1, you registered XGBRegressorTraining component from gallery to your workspace. Let's use this component as an example to understand the structure of a component.

In general, Azure Machine Learning component is constituted by 2 files:

  1. Component specification in Yaml format. XGBRegressorTraining is the Yaml file in the example . This file defines the component with 4 sections:
    • Metadata: e.g. name, display_name, type, tags.
    • Interface: inputs and outputs of the component.
    • Implementation: For CommandComponent, command is the 1-line command which will be executed.
    • Environment: environment section defines the docker image and other dependencies.
  2. Relative Python code which includes your own logic. XGBRegressorTraining is the python file for our example.

The correlation between component UI, Yaml spec and Py code is illustrated as below: create-component

Useful tips about the Yaml spec:

  • Component name constraints. Name is the unique identifier across all components. Currently, name only accepts letters, numbers and '-._'.
  • Component type & schema. Currently, Azure Machine Learning component only supports CommandComponent. The schema and type settings are fixed values right now to support CommandComponent.
  • Data interface. DataFrameDirectory is the common data interface when uses Azure Machine Learning datasets across different components. This type could be easily transformed to Pandas DataFrame. It is suggested to use DataFrameDirectory as the data inputs/outputs when you define your components.

  • Port & Parameters between UI and Yaml Spec. From component UI, there are 2 types of inputs/outputs: Ports and Parameters. create-component

    When you use DataFrameDirectory or AnyDirectory as an input/output type, it will show as port. When you use basic type, like integer, float or string, it will show in parameters right panel.

You can refer to 'component spec definition' for more details about the Yaml Spec settings and examples.

Build XGBRegressorEvaluation component

Let's start creating a new component! After you train the XGBoost regressor model for predicting automobile price, you need to have a new component, XGBRegressorEvaluation, to evaluate the performance of this model. We can define this new evaluation component interface as below: create-component

Yaml spec

Now, you can try to create Yaml file for this component to fit XGBRegressorEvaluation definition. Copy the Yaml template to start and then refer to XGBRegressorTraining Yaml example to help you fill in the content. You can also check the finalized yaml spec XGBRegressorEvaluation when you finish your file or need some suggestions.

Py code

You will use 'Root-Mean-Square Error (RMSE)' from sklearn.metrics. Now, the core pseudo code to evaluate a given model is illustrated as below:

## Load model
import xgboost as xgb
xg_reg = xgb.XGBRegressor()
xg_reg.load_model("xgboost_modelfile.json")

## Evaluation
import numpy as np
from sklearn.metrics import mean_squared_error
preds = xg_reg.predict(evaluationData_features)
rmse = np.sqrt(mean_squared_error(evaluationData_lable, preds))
print("RMSE: %f" % (rmse))

Follow 2 steps below to upgrade these pseudo code to fit XGBRegressorEvaluation defination:

  1. Parse the inputs from args. Generally, use 'argparse.ArgumentParser' and 'add_argument' to get all input values from command.
  2. Prepare your data for prediction and evaluation. Leverage 'load_data_frame_from_directory' to convert input 'DataFrameDirectory' data to Pandas DataFrame.

You can try to follow the XGBRegressorTraining py code for an example. We also prepared a finalized XGBRegressorEvaluation py code for your reference.

Register component from local

When the Yaml Spec and Py code are ready, you can either edit them from example or download them from this folder directly. Follow the registration guidance in tutorial 1 to register the XGBRegressorEvaluation component on Azure Machine Learning Modules page. The only difference is that you need to register the component from 'Local files' and choose the right path which contains the Yaml Spec and Py code:

create-component-from-localfiles

Then click 'Next' and 'Create' to finish the creation of XGBRegressorEvaluation component. You will see the new component is listed in Modules page.

component-description

Update pipeline

Now let's use the new component to evaluate the output model from the pipeline we created in tutorial 1.

Go to the 'Designer' page, select Tutorial 1 pipeline, and find the XGBRegressorEvaluation component under 'Custom Module' collection: component-description

Drag this component to the canvas and link it to precedent components and define its parameter values:

  • The second output of 'Split Data' component is the data input as evaluation data set.
  • 'XGBRegressorTraining' output the model directory. This is the input model directory for XGBRegressorEvaluation component.
  • Put 'price' as the value of 'Lable_Col'.
  • Put 'xgb_modelfile.json' as the value of 'Model_FileName'. This should be aligned with XGBRegressorTraining model file defination.

component-description

Now you can submit the new pipeline with new component to evaluate the model performance.

Contribute to the gallery

Azure Machine Learning Gallery always welcomes contributor to enrich the gallery with more components or pipelines. Once your component is ready to serve your scenario, please consider sharing it into the gallery. Refer to this link to start sharing your component into gallery.