Pipeline Building Quickstart Tutorial

In this tutorial, you will learn how to perform the following tasks:

Create a pipeline for use with your data.
Run a pipeline and learn how to view the results.
Learn how to modify your pipeline so that you can modify the future runs easily.

Creating a Task

After you have successfully installed DBND, you can create a new task, run it, and change its inputs.

Step 1: Create a Task

Navigate to the root directory of your DBND project (the root directory is the one with the project.cfg file) via the command line.
In the root directory, create a new file called example.py and open it using your text editor/IDE. Let's add a simple Python function with a DBND @task decorator:

from dbnd import task

@task
def calculate_alpha(alpha: float = 0.5) -> float:
    alpha += 0.1
    return alpha

Save your file.

Step 2: Run a Task

Now that you have created a task let us run it. To run the task via CLI, use:

dbnd run example.calculate_alpha

You will see a similar snippet near the bottom of the command line output:

PARAMS:    : 
Name    Kind    Type    Format    Source    -= Value =-
alpha   param   float             default    0.5
result  output  float   .pickle              /Users/name/databand/data/dev/2021-08-02/calculate_alpha/calculate_alpha_d869b45637/result.txt :='0.6'

[Text omitted for brevity]

====================
= Your run has been successfully executed!

Step 3: Iterate Task Inputs

DBND allows you to change your task inputs from the CLI. Changing these values affects the output of your tasks.

In your task, the calculate_alpha function has a default value for the alpha parameter. You can override it using the CLI:

dbnd run example.calculate_alpha --set-root alpha=0.4

The results displayed should include the following snippet:

PARAMS:    :
Name    Kind    Type    Format    Source            -= Value =-
alpha   param   float             ctor               0.4
result  output  float   .pickle                      /Users/name/databand/data/dev/2021-08-02/calculate_alpha/calculate_alpha_d869b45637/result.txt :='0.5'
=
====================

Creating a Pipeline

In this section, you will create a pipeline using some example code.

Pipelines 101

A pipeline is a series of tasks wired together to run in some order.

In the following sections, you will work with a pipeline that produces an ML model that predicts different wine quality.

Step 1: Look at the Initial Code Workflow

The following is the initial workflow code before any modifications with DBND. The process includes the following tasks:

Task 1: Input Data - loading a data file containing the wines and their attributes
Task 2: Data Preparation - splitting the data set into distinct training and validation sets
Task 3: Model Training - creating an ElasticNet model based on the training data set
Task 4: Model Validation - testing the model with test data to create performance metrics

In addition to these four tasks, the workflow includes two input parameters that have been hardcoded.

import pandas as pd
import numpy as np
import logging
from sklearn.linear_model import ElasticNet
from sklearn.model_selection import train_test_split
from sklearn.metrics import mean_absolute_error, mean_squared_error, r2_score

logging.basicConfig(level=logging.INFO)

def training_script():
    # load data
    raw_data = pd.read_csv(data_repo.wines)

    # split data into training and validation sets
    train_df, validation_df = train_test_split(raw_data)

    # create hyperparameters and model
    alpha = 0.5
    l1_ratio = 0.2
    lr = ElasticNet(alpha=alpha, l1_ratio=l1_ratio)
    lr.fit(train_df.drop(["quality"], 1), train_df[["quality"]])

    # validation
    validation_x = validation_df.drop(["quality"], 1)
    validation_y = validation_df[["quality"]]
    prediction = lr.predict(validation_x)
    rmse = np.sqrt(mean_squared_error(validation_y, prediction))
    mae = mean_absolute_error(validation_y, prediction)
    r2 = r2_score(validation_y, prediction)

    logging.info("%s,%s,%s", rmse, mae, r2)

    return lr

training_script()

Step 2: Adapt the Workflow with DBND

Using DBND in the workflow involves the following steps:

Step 1. Functionalize the parts of the code that you want to modularize into tasks
Step 2. Assign @task decorators to define each function as a task
Step 3. Use the @pipeline decorator to define the structure of the pipeline

# python3.6
import numpy as np
import logging
from pandas import DataFrame
from dbnd import task, pipeline
from sklearn.linear_model import ElasticNet
from dbnd import log_dataframe, log_metric
from sklearn.metrics import mean_absolute_error, mean_squared_error, r2_score
from typing import Tuple
from sklearn.model_selection import train_test_split

logging.basicConfig(level=logging.INFO)

@task(result="training_set, validation_set")
def prepare_data(raw_data: DataFrame) -> Tuple[DataFrame, DataFrame]:
    train_df, validation_df = train_test_split(raw_data)

    return train_df, validation_df

@task
def train_model(
    training_set: DataFrame,
    alpha: float = 0.5,
    l1_ratio: float = 0.5,
) -> ElasticNet:
    lr = ElasticNet(alpha=alpha, l1_ratio=l1_ratio)
    lr.fit(training_set.drop(["quality"], 1), training_set[["quality"]])

    return lr

@task
def validate_model(model: ElasticNet, validation_dataset: DataFrame) -> str:
    log_dataframe("validation", validation_dataset)

    validation_x = validation_dataset.drop(["quality"], 1)
    validation_y = validation_dataset[["quality"]]

    prediction = model.predict(validation_x)
    rmse = np.sqrt(mean_squared_error(validation_y, prediction))
    mae = mean_absolute_error(validation_y, prediction)
    r2 = r2_score(validation_y, prediction)

    log_metric("rmse", rmse)
    log_metric("mae", rmse)
    log_metric("r2", r2)

    return "%s,%s,%s" % (rmse, mae, r2)

@pipeline(result=("model", "validation"))
def predict_wine_quality(
    raw_data: DataFrame,
    alpha: float = 0.5,
    l1_ratio: float = 0.5,
):
    training_set, validation_set = prepare_data(raw_data=raw_data)

    model = train_model(
        training_set=training_set, alpha=alpha, l1_ratio=l1_ratio
    )
    validation = validate_model(model=model, validation_dataset=validation_set)

    return model, validation

Pay attention to the following changes in the modified code above:

Each task has a @task decorator with input and output parameters.
The pipeline has a @pipeline decorator with defined input and output parameters.
The result keyword at the decorator is used for defining names for outputs of the task.
The input data (in this case, a CSV file) is no longer a part of the code. Instead, it is an attribute that's entered and is adjustable as a parameter.
All variable parameters have been defined as pipeline attributes.
Logging and metrics APIs have been added to the model_validation task to track performance.

Step 3: Create the Wine Quality Pipeline

Using the text editor/IDE of your choice, begin by creating a new Python module. You can call it with the name of your choice (e.g., predict_wine_quality).
Copy and save the DBND code provided on this page into your module.

Running a Pipeline with Data

Create a new file called wine.csv and copy in the sample data we've provided.
Once you've prepared your data file, you can run your pipeline with the CLI:

dbnd run predict_wine_quality.predict_wine_quality --set raw_data=wine.csv

Note that this command reflects the module and data file that is saved in your DBND directory; otherwise, you will need to provide full file names.

Running a Pipeline with Varying Inputs

You can run a pipeline with different input data and/or parameters by specifying a different data file or parameter as part of the dbnd run command.

Re-Run Your Pipeline with a Different Data File

To re-run your pipeline with a different data file, duplicate your original wine.csv file and call it wine2.csv.

In the command line, run:

dbnd run predict_wine_quality.predict_wine_quality --set raw_data=wine2.csv

DBND will run the pipeline using the new data file.

Re-Run Your Pipeline with Different Parameter Values

You can change the parameter values for a run by specifying new values as part of the dbnd run command. For example, if you wanted to set the values for alpha and l1_ratio to 0.7, you would run the following:

dbnd run predict_wine_quality.predict_wine_quality --set raw_data=wine.csv --set l1_ratio=0.7 --set alpha=0.7

DBND will run the pipeline using the values provided as the net input parameters.