Code repository for the O'Reilly publication "Building Machine Learning Pipelines" by Hannes Hapke & Catherine Nelson
Download the initial dataset. From the root of this repository, execute
python3 utils/download_dataset.py
After this script runs, you should have a data
folder containing the file consumer_complaints_with_narrative.csv
.
The data that we use in this example project can be downloaded using the script above. The dataset is from a public dataset on customer complaints collected from the US Consumer Finance Protection Bureau. If you would like to reproduce our edited dataset, carry out the following steps:
- Download the dataset from https://www.consumerfinance.gov/data-research/consumer-complaints/#download-the-data
- Rename the columns to
[ "product", "sub_product", "issue", "sub_issue", "consumer_complaint_narrative", "company", "state", "zip_code", "company", "company_response", "timely_response", "consumer_disputed"]
- Filter the dataset to remove rows with missing data in the
consumer_complaint_narrative
column - In the
consumer_disputed
column, mapYes
to1
andNo
to0
Before building our TFX pipeline, we experimented with different feature engineering and model architectures. The notebooks in this folder preserve our experiments, and we then refactored our code into the interactive pipeline below.
The interactive-pipeline
folder contains a full interactive TFX pipeline for the consumer complaint data.
The pipelines
folder contains complete pipelines for the various orchestrators. See Chapters 11 and 12 for full details.
The following subfolders contain stand-alone code for individual chapters.
Chapter 7. Stand-alone code for TFMA, Fairness Indicators, What-If Tool. Note that these notebooks will not work in JupyterLab.
Chapter 10. Notebook outlining the implementation of custom TFX components from scratch and by inheriting existing functionality. Presented at the Apache Beam Summit 2020.
Chapter 14. Code for training a differentially private version of the demo project. Note that the TF-Privacy module only supports TF 1.x as of June 2020.
The code was written and tested for version 0.22.
-
As of 9/22/20, the interactive pipeline runs on TFX version 0.24.0rc1. Due to tiny TFX bugs, the pipelines currently don't work on the releases 0.23 and 0.24-rc0. Github issues have been filed with the TFX team specifically for the book pipelines (Issue 2500). We will update the repository once the issue is resolved.
-
As of 9/14/20, TFX only supports Python 3.8 with version >0.24.0rc0.