Applying Kedro Orchestration Pipelines to Deploy a Deep Learning Transformer Architecture for the Toxic Comment Classification Problem

Project Overview

This project leverages Kedro, a data and machine learning pipeline orchestration tool, to classify toxic comments using deep learning transformer architecture. It integrates Kedro-Viz for pipeline visualization and follows data engineering best practices for reproducibility and maintainability.

Key Features

• Pipeline Visualization: Integrates Kedro-Viz to provide an interactive visual representation of data workflows for better debugging and workflow comprehension.
• Multilingual Dataset Support: Handles multilingual datasets, including English, with preprocessing pipelines optimized for transformer models.
• Transformer-Based Modeling: Supports state-of-the-art deep learning models, such as BERT, for toxicity prediction.

Tech Stack

• Kedro v0.19.10
• TensorFlow for model training
• FastAPI for serving predictions
• MLflow for experiment tracking
• Docker for containerization
• Prometheus and Grafana for monitoring

For more information, see the Kedro documentation.

---

System Design

---

Kedro Pipeline Overview

---

Dataset Overview

This project utilizes datasets from the Jigsaw "Toxic Comment Classification" competition. The datasets include labeled comments sourced from platforms like Wikipedia and Civil Comments.

Description:

• Comment text: The primary data containing user-submitted comments;
• Toxic Column: A binary label where 1 indicates toxic comments, and 0 indicates non-toxic comments.

After that, you need to download the follow files:

• jigsaw-toxic-comment-train.csv;
• validation.csv;
• test.csv.

For all the files, add in the follow path: data/01_raw/

---

Rules and guidelines

To ensure consistency and reproducibility:

1. .gitignore Compliance: Do not remove any lines from the provided .gitignore file.
2. Data Engineering Convention: Follow Kedro's data engineering conventions.
3. Data Handling: Do not commit raw datasets to the repository.
4. Credential Security: Avoid committing credentials or local configurations to the repository. Store these in conf/local/.

---

Get started

Installation

Install the required dependencies:

pip install -r requirements.txt

Running the Pipeline

Execute the Kedro pipeline:

kedro run

Testing the Pipeline

Run tests using pytest:

pytest

Deploying the Model API

To serve the champions model, you can build the container toxic-comment-endpoint using the fastapi package for serving.

1. Build the Docker image:

   docker build -t toxic-comment-endpoint .

2. Run the Docker container:

   docker run -p 8001:8001 -e PORT=8001 toxic-comment-endpoint

3. Click on the link http://0.0.0.0:8001 and you see the follow web application

   

---

Monitoring with Prometheus and Grafana

Create a prometheus.yml file for monitoring and sending metrics to Grafana:

global:
  scrape_interval: 15s

scrape_configs:
  - job_name: "fastapi"
    metrics_path: "/metrics"
    static_configs:
      - targets: ["0.0.0.0:8001"]

remote_write:
  - url: "<YOUR_GRAFANA_URL>"
    remote_timeout: "30s"
    send_exemplars: false
    follow_redirects: true
    basic_auth:
      username: "<YOUR_USERNAME>"
      password: "<YOUR_API_TOKEN>"

Run Prometheus with the configuration file:

prometheus --config.file=prometheus.yml

Grafana Visualization

Additional Resources

• Kedro Documentation
• TensorFlow Documentation
• FastAPI Documentation
• MLflow Documentation
• Prometheus Documentation
• Grafana Documentation