Comparing Machine Learning Models for Cancer Prediction using Weights & Biases

This repository contains code for training predictive models on the Cancer_Dataset using Weights & Biases (wandb) for tracking experiments.

Table of Contents

1. About Saturdays.AI
2. Dataset
3. Exploratory Data Analysis (EDA)
4. Training
5. Conclusion
6. Acknowledgements

About Saturdays.AI

This project is a part of my internship at Saturdays.AI, an organization dedicated to democratizing Artificial Intelligence education and fostering a global community of AI practitioners. By participating in the internship, I had the opportunity to apply the concepts and techniques learned during the program to real-world problems, such as the cancer prediction task in this repository.

Dataset

The dataset can be found in the data/Cancer_Dataset.csv file. It contains features related to cell nuclei extracted from breast cancer samples. The target variable is the diagnosis (benign or malignant).

Exploratory Data Analysis (EDA)

A superficial exploratory data analysis has been performed in the EDA.ipynb notebook. The analysis focused on understanding the types of features (all continuous numerical variables) and the target variable (categorical, with 'B' for benign and 'M' for malignant).

Training

Model training is carried out in the train.py script. To run the training, open a terminal and execute the following command:

python train.py

You can customize the training by providing arguments to the parse_args() function in the script. The available arguments are:

--seed: The random seed used for reproducibility (default: 42) --log_preds: Whether or not to log model predictions (default: False) --data_path: Path to the CSV dataset file (default: "data/Cancer_Data.csv") --test_size: Proportion of the dataset to be included in the test split (default: 0.33) --model_index: Index of the model to use for training (default: 0). The available models are: 0: LogisticRegression 1: RidgeClassifier 2: SGDClassifier 3: PassiveAggressiveClassifier 4: Perceptron For example, to train a RidgeClassifier model, you can run:

python train.py --model_index 1

This will override the default model (LogisticRegression) and train a RidgeClassifier model instead.

You can also use the provided YAML configuration file to run a hyperparameter sweep using Weights & Biases. This will help you find the best combination of hyperparameters for your model.

First, create the sweep using the YAML configuration file:

wandb sweep sweep.yaml

This command will generate a sweep ID. Copy the sweep ID, and use it to launch the sweep agent:

wandb agent SWEEP_ID

Replace "SWEEP_ID" with the actual ID generated in the previous step. This command will run multiple training runs with different hyperparameter combinations as specified in the YAML configuration file. You can track the progress of your sweep and visualize the performance of different hyperparameter configurations in the Weights & Biases dashboard.

Conclusion

By leveraging Weights & Biases for tracking experiments, we were able to train and compare multiple models on the Cancer Dataset. This approach enables efficient model selection and aids in improving the overall performance of the predictive models. As you continue working on this project, consider exploring additional models, feature engineering techniques, or hyperparameter tuning to further improve the results.

Acknowledgements

I would like to express my gratitude to Saturdays.AI, where I had the opportunity to participate in a practical learning experience. Their support and guidance have been instrumental in my growth and the development of this project.