Manage your ML models with Weights and Biases and Airflow

Weights and Biases (W&B) is a machine learning platform for model management that includes features like experiment tracking, dataset versioning, and model performance evaluation and visualization. Using W&B with Airflow gives you a powerful ML orchestration stack with first-class features for building, training, and managing your models. In this tutorial, you’ll learn how to create an Airflow DAG that completes feature engineering, model training, and predictions with the Astro Python SDK and scikit-learn, and registers the model with W&B for evaluation and visualization.

​

Time to complete

This tutorial takes approximately one hour to complete.

​

Assumed knowledge

To get the most out of this tutorial, you should be familiar with:

• Airflow operators. See Operators 101.
• The Astro Python SDK. See Astro Python SDK documentation
• Weights and Biases. See What is Weights and Biases?.

​

Prerequisites

• The Astro CLI.
• A Weights and Biases account. Personal accounts are available for free.

​

Quickstart

If you have a GitHub account, you can get started quickly by cloning the demo repository. For more detailed instructions for setting up the project, start with Step 1.

1. Clone the demo repository:
   git clone https://github.com/astronomer/airflow-wandb-demo
   cd airflow-wandb-demo
   
2. Update the .env file with your WANDB_API_KEY.
3. Start Airflow by running:
   astro dev start
   
4. Continue with Step 7 below.

​

Step 1: Configure your Astro project

Use the Astro CLI to create and run an Airflow project locally.

1. Create a new Astro project:
   $ mkdir astro-wandb-tutorial && cd astro-wandb-tutorial
   $ astro dev init
   
2. Add the following line to the requirements.txt file of your Astro project:
   astro-sdk-python[postgres]==1.5.3
   wandb==0.14.0
   pandas==1.5.3
   numpy==1.24.2
   scikit-learn==1.2.2
   
   This installs the packages needed to transform the data and run feature engineering, model training, and predictions.

​

Step 2: Prepare the data

This tutorial creates a model that classifies churn risk based on customer data.

1. Create a subfolder called data in your Astro project include folder.
2. Download the demo CSV files from this GitHub directory.
3. Save the downloaded CSV files in the include/data folder. You should have 5 files in total.

​

Step 3: Create your SQL transformation scripts

Before feature engineering and training, the data needs to be transformed. This tutorial uses the Astro Python SDK transform_file function to complete several transformations using SQL.

1. Create a file in your include folder called customer_churn_month.sql and copy the following code into the file.
   with subscription_periods as (
       select subscription_id, 
           customer_id, 
           cast(start_date as date) as start_date, 
           cast(end_date as date) as end_date, 
           monthly_amount 
           from {{subscription_periods}}
   ),
   
   months as (
       select cast(date_month as date) as date_month from {{util_months}}
   ),
   
   customers as (
       select
           customer_id,
           date_trunc('month', min(start_date)) as date_month_start,
           date_trunc('month', max(end_date)) as date_month_end
       from subscription_periods
       group by 1
   ),
   
   customer_months as (
       select
           customers.customer_id,
           months.date_month
       from customers
       inner join months
           on  months.date_month >= customers.date_month_start
           and months.date_month < customers.date_month_end
   ),
   
   joined as (
       select
           customer_months.date_month,
           customer_months.customer_id,
           coalesce(subscription_periods.monthly_amount, 0) as mrr
       from customer_months
       left join subscription_periods
           on customer_months.customer_id = subscription_periods.customer_id
           and customer_months.date_month >= subscription_periods.start_date
           and (customer_months.date_month < subscription_periods.end_date
                   or subscription_periods.end_date is null)
   ),
   
   customer_revenue_by_month as (
       select
           date_month,
           customer_id,
           mrr,
           mrr > 0 as is_active,
           min(case when mrr > 0 then date_month end) over (
               partition by customer_id
           ) as first_active_month,
   
           max(case when mrr > 0 then date_month end) over (
               partition by customer_id
           ) as last_active_month,
   
           case
           when min(case when mrr > 0 then date_month end) over (
               partition by customer_id
           ) = date_month then true
           else false end as is_first_month,
           case
           when max(case when mrr > 0 then date_month end) over (
               partition by customer_id
           ) = date_month then true
           else false end as is_last_month
       from joined
   ),
   
   joined1 as (
   
       select
           date_month + interval '1 month' as date_month,
           customer_id,
           0::float as mrr,
           false as is_active,
           first_active_month,
           last_active_month,
           false as is_first_month,
           false as is_last_month
   
       from customer_revenue_by_month
   
       where is_last_month
   
   )
   
   select * from joined1;
   

   See all 91 lines
2. Create another file in your include folder called customers.sql and copy the following code into the file.
   with
   customers as (
   
       select *
       from {{customers_table}}
   
   ),
   
   orders as (
   
       select *
       from {{orders_table}}
   
   ),
   
   payments as (
   
       select *
       from {{payments_table}}
   
   ),
   
   customer_orders as (
   
       select
       customer_id,
       cast(min(order_date) as date) as first_order,
       cast(max(order_date) as date) as most_recent_order,
       count(order_id) as number_of_orders
       from orders
   
       group by customer_id
   
   ),
   
   customer_payments as (
   
       select
       orders.customer_id,
       sum(amount / 100) as total_amount
   
       from payments
   
       left join orders on payments.order_id = orders.order_id
   
       group by orders.customer_id
   
   ),
   
   final as (
   
       select
       customers.customer_id,
       customers.first_name,
       customers.last_name,
       customer_orders.first_order,
       customer_orders.most_recent_order,
       customer_orders.number_of_orders,
       customer_payments.total_amount as customer_lifetime_value
   
       from customers
   
       left join customer_orders on customers.customer_id = customer_orders.customer_id
   
       left join customer_payments on customers.customer_id = customer_payments.customer_id
   
   )
   
   select
   *
   from final
   

   See all 71 lines

​

Step 4: Create a W&B API key

In your W&B account, create an API key that you will use to connect Airflow to W&B. You can create a key by going to the Authorize page or your user settings.

​

Step 5: Set up your connections and environment variables

You’ll use environment variables to create Airflow connections to Snowflake and W&B, as well as to configure the Astro Python SDK.

1. Open the .env file in your Astro project and paste the following code.
   WANDB_API_KEY='<your-wandb-api-key>'
   AIRFLOW_CONN_POSTGRES_DEFAULT='postgresql://postgres:postgres@host.docker.internal:5432/postgres?options=-csearch_path%3Dtmp_astro'
   
2. Replace <your-wandb-api-key> with the API key you created in Step 4. No changes are needed for the AIRFLOW_CONN_POSTGRES_DEFAULT environment variable.

​

Step 6: Create your DAG

1. Create a file in your Astro project dags folder called customer_analytics.py and copy the following code into the file:
   from datetime import datetime
   import os
   
   from astro import sql as aql
   from astro.files import File
   from astro.sql.table import Table
   from airflow.decorators import dag, task_group
   
   import pandas as pd
   import numpy as np
   from sklearn.model_selection import train_test_split
   from sklearn.ensemble import RandomForestClassifier
   import tempfile
   import pickle
   from pathlib import Path
   
   import wandb
   from wandb.sklearn import plot_precision_recall, plot_feature_importances
   from wandb.sklearn import plot_class_proportions, plot_learning_curve, plot_roc
   
   _POSTGRES_CONN = "postgres_default"
   wandb_project = "demo"
   wandb_team = "astro-demos"
   local_data_dir = "include/data"
   sources = ["subscription_periods", "util_months", "customers", "orders", "payments"]
   
   
   @dag(schedule=None, start_date=datetime(2023, 1, 1), catchup=False)
   def customer_analytics():
       @task_group()
       def extract_and_load(sources: list) -> dict:
           for source in sources:
               aql.load_file(
                   task_id=f"load_{source}",
                   input_file=File(f"{local_data_dir}/{source}.csv"),
                   output_table=Table(
                       name=f"STG_{source.upper()}", conn_id=_POSTGRES_CONN
                   ),
                   if_exists="replace",
               )
   
       @task_group()
       def transform():
           aql.transform_file(
               task_id="transform_churn",
               file_path=f"{Path(__file__).parent.as_posix()}/../include/customer_churn_month.sql",
               parameters={
                   "subscription_periods": Table(
                       name="STG_SUBSCRIPTION_PERIODS", conn_id=_POSTGRES_CONN
                   ),
                   "util_months": Table(name="STG_UTIL_MONTHS", conn_id=_POSTGRES_CONN),
               },
               op_kwargs={
                   "output_table": Table(
                       name="CUSTOMER_CHURN_MONTH", conn_id=_POSTGRES_CONN
                   )
               },
           )
   
           aql.transform_file(
               task_id="transform_customers",
               file_path=f"{Path(__file__).parent.as_posix()}/../include/customers.sql",
               parameters={
                   "customers_table": Table(name="STG_CUSTOMERS", conn_id=_POSTGRES_CONN),
                   "orders_table": Table(name="STG_ORDERS", conn_id=_POSTGRES_CONN),
                   "payments_table": Table(name="STG_PAYMENTS", conn_id=_POSTGRES_CONN),
               },
               op_kwargs={"output_table": Table(name="CUSTOMERS", conn_id=_POSTGRES_CONN)},
           )
   
       @aql.dataframe()
       def features(customer_df: pd.DataFrame, churned_df: pd.DataFrame) -> pd.DataFrame:
           customer_df["customer_id"] = customer_df["customer_id"].apply(str)
           customer_df.set_index("customer_id", inplace=True)
   
           churned_df["customer_id"] = churned_df["customer_id"].apply(str)
           churned_df.set_index("customer_id", inplace=True)
           churned_df["is_active"] = churned_df["is_active"].astype(int).replace(0, 1)
   
           df = (
               customer_df[["number_of_orders", "customer_lifetime_value"]]
               .join(churned_df[["is_active"]], how="left")
               .fillna(0)
               .reset_index()
           )  # inplace=True)
   
           return df
   
       @aql.dataframe()
       def train(df: pd.DataFrame) -> dict:
           features = ["number_of_orders", "customer_lifetime_value"]
           target = ["is_active"]
   
           test_size = 0.3
           X_train, X_test, y_train, y_test = train_test_split(
               df[features], df[target], test_size=test_size, random_state=1883
           )
           X_train = np.array(X_train.values.tolist())
           y_train = np.array(y_train.values.tolist()).reshape(
               len(y_train),
           )
           y_train = y_train.reshape(
               len(y_train),
           )
           X_test = np.array(X_test.values.tolist())
           y_test = np.array(y_test.values.tolist())
           y_test = y_test.reshape(
               len(y_test),
           )
   
           model = RandomForestClassifier()
           _ = model.fit(X_train, y_train)
           model_params = model.get_params()
   
           y_pred = model.predict(X_test)
           y_probas = model.predict_proba(X_test)
           importances = model.feature_importances_
           indices = np.argsort(importances)[::-1]
   
           wandb.login()
           run = wandb.init(
               project=wandb_project,
               config=model_params,
               entity=wandb_team,
               group="wandb-demo",
               name="jaffle_churn",
               dir="include",
               mode="online",
           )
   
           wandb.config.update(
               {"test_size": test_size, "train_len": len(X_train), "test_len": len(X_test)}
           )
           plot_class_proportions(y_train, y_test, ["not_churned", "churned"])
           plot_learning_curve(model, X_train, y_train)
           plot_roc(y_test, y_probas, ["not_churned", "churned"])
           plot_precision_recall(y_test, y_probas, ["not_churned", "churned"])
           plot_feature_importances(model)
   
           model_artifact_name = "churn_classifier"
   
           with tempfile.NamedTemporaryFile(delete=False) as tf:
               pickle.dump(model, tf)
               tf.close()
               artifact = wandb.Artifact(model_artifact_name, type="model")
               artifact.add_file(local_path=tf.name, name=model_artifact_name)
               wandb.log_artifact(artifact)
               os.remove(tf.name)
   
           wandb.finish()
   
           return {"run_id": run.id, "artifact_name": model_artifact_name}
   
       @aql.dataframe()
       def predict(model_info: dict, customer_df: pd.DataFrame) -> pd.DataFrame:
           wandb.login()
           run = wandb.init(
               project=wandb_project,
               entity=wandb_team,
               group="wandb-demo",
               name="jaffle_churn",
               dir="include",
               resume="must",
               id=model_info["run_id"],
           )
   
           customer_df.fillna(0, inplace=True)
   
           features = ["number_of_orders", "customer_lifetime_value"]
   
           artifact = run.use_artifact(
               f"{model_info['artifact_name']}:latest", type="model"
           )
   
           with tempfile.TemporaryDirectory() as td:
               with open(artifact.file(td), "rb") as mf:
                   model = pickle.load(mf)
                   customer_df["PRED"] = model.predict_proba(
                       np.array(customer_df[features].values.tolist())
                   )[:, 0]
   
           wandb.finish()
   
           customer_df.reset_index(inplace=True)
   
           return customer_df
   
       _extract_and_load = extract_and_load(sources)
   
       _transformed = transform()
   
       _features = features(
           customer_df=Table(name="customers", conn_id=_POSTGRES_CONN),
           churned_df=Table(name="customer_churn_month", conn_id=_POSTGRES_CONN),
       )
   
       _model_info = train(df=_features)
   
       _predict_churn = predict(
           model_info=_model_info,
           customer_df=Table(name="customers", conn_id=_POSTGRES_CONN),
           output_table=Table(name=f"pred_churn", conn_id=_POSTGRES_CONN),
       )
   
       _extract_and_load >> _transformed >> _features
   
   
   customer_analytics()
   

   See all 208 lines
   This DAG completes the following steps:
   • The extract_and_load task group contains one task for each CSV in your include/data folder that uses the Astro Python SDK load_file function to load the data to Postgres.
   • The transform task group contains two tasks that transform the data using the Astro Python SDK transform_file function and the SQL scripts in your include folder.
   • The features task is a Python function implemented with the Astro Python SDK @dataframe decorator that uses Pandas to create the features needed for the model.
   • The train task is a Python function implemented with the Astro Python SDK @dataframe decorator that uses scikit-learn to train a Random Forest classifier model and push the results to W&B.
   • The predict task pulls the model from W&B in order to make predictions and stores them in Postgres.
2. Run the following command to start your project in a local environment:
   astro dev start
   

​

Step 7: Run your DAG and view results

1. Open the Airflow UI, unpause the customer_analytics DAG, and trigger the DAG.
2. The logs in the train and predict tasks will contain a link to your W&B project which shows plotted results from the training and prediction.
   
   Go to one of the links to view the results in W&B.