MLOps: Real-World Machine Learning Projects for Professional

Load Reddit sentiment data from a GitHub CSV with clean comments and a 1, -1, 0 category, then perform data preprocessing and EDA with pandas.

Load and preprocess text data from Kaggle and GitHub in Google Colab, perform EDA, then apply stopword removal, lemmatization, and n-grams for a baseline model.

Learn to set up an MLflow server on AWS via EC2, configure an S3 bucket for artifacts, and run port 5000 to track experiments with a baseline model.

build a baseline text classifier with a random forest on bag-of-words features using countvectorizer (10k features), trained and logged to mlflow on aws with evaluation metrics.

Compare bag-of-words and tf-idf vectorization (unigram to trigram) for a random forest baseline, using mlflow to track experiments; find tf-idf with trigram improves recall and guides feature choices.

This lecture demonstrates tf-idf vectorization with two-grams and varying max features across experiments tracked in mlflow, with results stored in an S3 bucket and evaluated by accuracy, precision, and recall.

Explore handling imbalanced data by applying five techniques—class weights, oversampling (smart), undersampling, SMOTE, and random undersampling—on a tf-idf with trigram features.

Experiment with text vectorization from count vectorizer to tf-idf with trigram features and 1000 max features, apply oversampling, and compare models like xgboost and lightgbm with hyperparameter tuning and mlflow.

Compare stacking of logistic regression, k-nearest neighbors, and LightGBM with tf-idf preprocessing to a LightGBM model, tuned by Optuna, then plan to build an ml pipeline for a simpler model.

Build ML pipeline with DVC, covering data ingestion, data pre-processing, model building, evaluation, and MLflow-based model registration, then expose the model via a Flask API for a Chrome plugin.

Demonstrate a data ingestion pipeline by loading data from a URL, cleaning via preprocessing, and saving train and test CSV files in the data folder, integrated with DVC.

Demonstrates data preprocessing with nltk, including stripping lowercase, removing punctuation, and lemmatization, saving preprocessed data to data/interim for train and test sets, with a dvc pipeline tracking execution.

Build and save a text model by vectorizing data with tf-idf, training a Lightgbm model, and persisting both the model and the tf-idf vectorizer using DVC and the parameters YAML.

Explore mlflow-based model evaluation with metrics, confusion matrix, and classification reports, logging artifacts and run IDs to S3, and registering the latest model for API deployment.

Register a trained model with mlflow by loading run ID and model path from a JSON file and setting the tracking URI. Use mlflow.register_model to place the model in staging.

Learn to build a Flask application programming interface that serves a machine learning model for a Chrome extension on YouTube comments, including preprocessing, vectorization, loading models, and Postman testing.

Implement a Chrome plugin to fetch YouTube comments via Google Cloud API, perform sentiment analysis, and visualize results with charts and word clouds for an MLOps workflow.

Create a docker file using a python 3.8 slim buster image, dockerize the source into an app, install requirements, run app.py on Linux, and set up AWS.

Deploy a complete docker-based ml app on AWS by building a docker image, pushing to ECR, launching EC2, and setting up CI/CD with GitHub actions and IAM for port 8080.