Digit Recognizer

This project is a small neural network built from scratch with NumPy to classify handwritten digits from the Kaggle Digit Recognizer dataset.

Setup

Install the dependencies:

pip install -r requirements.txt

Place the Kaggle train.csv in one of the locations the script already checks, for example:

models/data/train.csv

You can also pass a custom path when running the script:

python models/neural_network.py --data path/to/train.csv

Run

Train the model and generate example prediction images:

python models/neural_network.py

The script will save sample predictions to outputs/.

Results

I ran the model on models/data/train.csv with the default settings in the script and got the following training accuracy progression:

Iteration	Accuracy
0	0.1148
500	0.8503
1000	0.8830
1500	0.8974
2000	0.9054
2500	0.9109
3000	0.9151
3500	0.9184
4000	0.9210
4500	0.9236
4950	0.9253

Example predictions from the final run:

• Index 30: predicted 7, label 7
• Index 3: predicted 8, label 8
• Index 50: predicted 8, label 8
• Index 8: predicted 0, label 0

Generated images:

• outputs/prediction_30.png
• outputs/prediction_3.png
• outputs/prediction_50.png
• outputs/prediction_8.png

Notes

• The script uses a simple 2-layer neural network with ReLU and softmax.
• The training accuracy above is on the training split used by the script, not a separate Kaggle leaderboard score.
• If you want to run fewer iterations for a quick test, use --iterations.

Dataset

Download the data from Kaggle:

https://www.kaggle.com/competitions/digit-recognizer/data