Scalable Machine Learning and Deep Learning - Lab1: Titanic Dataset Classifier

Github repository:

https://github.com/daniel-rdt/serverless_ml_titanic_dr

File running sequence for titanic survival prediction:

1. run: python titanic-feature-pipeline.py (once)
2. run: python titanic-training-pipeline.py (once)
3. upload huggingface-spaces-titanic folder files to HuggingFace

Hugging Face Spaces public URL: https://huggingface.co/spaces/daniel-rdt/titanic

File running sequence for titanic survival prediction for most recent passenger added:

4. run: python titanic-feature-pipeline-daily.py (set LOCAL=FALSE to start scheduled daily run on MODAL)
5. run: python titanic-batch-inference-pipeline.py (set LOCAL=FALSE to start scheduled daily run on MODAL)
6. upload huggingface-spaces-titanic-monitor folder files to HuggingFace

Hugging Face Spaces public URL: https://huggingface.co/spaces/daniel-rdt/titanic_monitor

File explanation:

titanic_feature_prep.py:

• contains the function "titanic_prep()"
• downloading data from GitHub repository;
• feature engineering:
• 1. removed "PassengerId","Name","Ticket" and "Cabin" columns;
• 2. all ages rounded up to next integer, so age less than 1 is rounded up to 1;
• 3. filled missing values for "Age" with created random age values in range from min. age to max. age;
• 4. filled missing values for "Embarked" with new "Unknown" category;
• 5. turned categorical variables ("Embarked" and "Sex") into binary numerical variables;
• 6. dropped old categorical columns "Sex" and "Embarked";
• returns a clean, prepped titanic dataset;

titanic-feature-pipeline.py:

• can be run either locally or with MODAL;
• applies titanic_prep function to yield the prepped titanic data set as titanic_df;
• creates Hopsworks feature group and stores features and labels of titanic_df into Hopsworks;

titanic-training-pipeline.py:

• can be run either locally or with MODAL;
• loades features and labels from Hopsworks feature store either as feature view or creates feature view from existing feature group;
• Train-Test-Split with 80/20;
• runs K-nearest-neighbours algorithm on train split;
• runs evaluation on test split and visualized as confusion matrix;
• Uploads model to the Hopsworks Model Registry;

huggingface-spaces-titanic folder:

• folder containing main app for user interface and requirements;
• app let's the user input feature values and returns a prediction

titanic-feature-pipeline-daily.py:

• same as titanic-feature-pipeline-daily when BACKFILL=TRUE;
• when BACKFILL=FALSE: generates new random passenger (either survivor or victim) and adds to Hopsworks feature group;
• ranges for survivor/victim features were determined according to min/max values or existing categories of a feature from the original titanic data set filtered for the labels respectively;

titanic-batch-inference-pipeline.py:

• can be run either locally or with MODAL on a daily schedule;
• gets model from Hopsworks registry and batch data from feature view;
• predicts if the synthetic passengers survived or not;
• downloads appropriate images for prediction and creates up to date confusion matrix when there have been 2 predictions made to date;
• stores predictions to Hopsworks;

huggingface-spaces-titanic-monitor folder:

• folder containing main app for user interface and requirements;
• app shows the most recent synthetic passenger prediction and real label, and a confusion matrix with historical prediction performance;

General notes

• everyone aged less than 1 year old is considered as 1 year old person;
• randomly generated passengers ranges for victims and survivors are very similar which may result in very ambiguous prediction results