A from-scratch multiclass classifier implementing logistic regression with a one-vs-all scheme. Includes a gradient-descent training loop with convergence checks and performance evaluation to verify correctness and generalization.
The project offers features such:
- One-vs-all training with class-wise parameters and unified argmax prediction.
- Gradient descent with learning rate/tolerance controls and loss curve logging.
- Feature scaling, bias term, and train/validation splits for fair evaluation.
- Metrics: accuracy, confusion matrix, and per-class precision/recall/F1.
- Simple save/load of weights and a CLI/notebook for classifying new samples.
Implemented logistic regression from scratch with a one-vs-all multiclass setup, feature scaling, bias term, and a gradient-descent trainer. Convergence checks and loss-curve logging with train/validation splits; evaluates accuracy, precision/recall/F1, and a confusion matrix. Simple CLI/notebook to save/load weights and classify new samples for reproducible experiments and baseline benchmarking.