An end-to-end Machine Learning project featuring a Deep Neural Network built from scratch using NumPy. This repository includes pre-trained weights, allowing you to run the interactive drawing application immediately.

• Inference-Ready: Includes mnist_weights.pkl so you can use the model without training.
• Neural Network from Scratch: Logic for forward/backpropagation and gradient descent implemented without ML frameworks (no TensorFlow/PyTorch).
• Interactive GUI: A Tkinter-based whiteboard to draw digits and get real-time predictions.
• Optimized Preprocessing: Automatic image scaling and inversion to match the MNIST dataset format.

The model is a 2-layer dense neural network:

• Input Layer: 784 neurons (28x28 pixel images).
• Hidden Layer: 64 neurons with ReLU activation.
• Output Layer: 10 neurons (digits 0-9) with Softmax activation.

• Accuracy: ~95% on MNIST test data.
• Training Method: Stochastic Gradient Descent (SGD).

Ensure you have Python installed along with these dependencies:

pip install numpy pandas matplotlib pillow

1. Clone the Repo: Ensure the mnist_weights.pkl file is in the same directory as your main script.
2. Launch: Run the Python script containing the Tkinter code.
3. Interact: * Draw a digit (0-9) on the white canvas using your mouse.
   • Click Predict to see the model's result.
   • Click Clear to wipe the canvas for a new attempt.

To ensure the model recognizes your drawings accurately, the GUI performs these automated steps whenever you click Predict:

• Capture: Saves the current canvas drawing as a PostScript file.
• Convert: Changes the image to Grayscale (L mode).
• Invert: Flips the colors to white-on-black to match the MNIST training standard.
• Resize: Squashes the drawing down to exactly 28x28 pixels.
• Normalize: Divides pixel values by 255 to feed the network values between 0 and 1.

If you wish to retrain the model from scratch or experiment with hyperparameters:

1. Data Setup: Ensure you have the mnist_train.csv file located in a folder named data/.
2. Execute Training: Run the gradient_descent function within the training script.
3. Save: The script will automatically overwrite the existing mnist_weights.pkl with your newly optimized parameters.