Neural Network Character Recognition

This project is a submission-focused, from-scratch NumPy ANN for handwritten character recognition. It is simplified to the core requested workflow: train and run a GUI using the handwritten dataset.

Submission Scope

• Main dataset: character_fonts (with handwritten data).npz
• Main model type: Multi-layer Perceptron (MLP)
• Main interface: gui_app.py

Implemented Components

• Multi-layer perceptron (MLP) with backpropagation for multi-class character classification.
• Tkinter GUI to train/test the MLP with adjustable hyperparameters and live visual feedback.
• NPZ-based dataset loading and train/test split.

Project Structure

• src/ann/data.py: dataset loading and train/test split.
• src/ann/mlp.py: from-scratch MLP classifier.
• src/ann/utils.py: encoding and evaluation helpers.
• train_mlp.py: train and evaluate multiclass MLP.
• gui_app.py: interactive training and prediction GUI with live plots.

Install

c:/python313/python.exe -m pip install -r requirements.txt

Train MLP (A-Z)

c:/python313/python.exe train_mlp.py --dataset "character_fonts (with handwritten data).npz" --epochs 15 --max-samples 120000 --val-size 0.1 --early-stopping-patience 5

The trainer now detects number of classes dynamically from the dataset labels. If the dataset includes class_names in NPZ metadata, predictions and confusion matrix labels use those names.

Outputs:

• Model: models/mlp_az.npz
• Curves: outputs/mlp_training_curve.png
• Confusion matrix: outputs/mlp_confusion_matrix.png

Run GUI

c:/python313/python.exe gui_app.py

In the GUI, you can:

• Pick a dataset (.npz).
• Set hidden layers, learning rate, epochs, batch size, max samples, validation split, and early stopping.
• Train the MLP with live loss/accuracy curves.
• Draw a character on canvas and predict it.
• Upload an image and view top class probabilities.

Notes

• The full dataset is large; start with --max-samples to iterate faster.
• The MLP supports relu or sigmoid hidden activation.
• GUI training uses a background worker thread with safe main-thread UI updates.