Digit Recognizer (MNIST) — CNN (Kaggle)

This project trains a Convolutional Neural Network (CNN) on the Kaggle Digit Recognizer dataset (MNIST-like, 28×28 grayscale images) and generates a submission.csv file for Kaggle competition submission.

What this project does

• Loads train.csv and test.csv from Kaggle input dataset
• Preprocesses pixels (normalize to 0–1 and reshape to 28×28×1)
• Splits data into train/validation
• Trains a CNN using TensorFlow/Keras
• Predicts digits (0–9) for the test set
• Creates submission.csv in Kaggle working directory

Model Architecture (Baseline CNN)

• Conv2D(32) → MaxPool
• Conv2D(64) → MaxPool
• Flatten → Dense(128) → Dropout(0.5)
• Dense(10, softmax)

Advanced CNN (Model 2)

An advanced CNN is also included:

• Multiple Conv blocks with BatchNormalization
• Dropout regularization
• EarlyStopping + ModelCheckpoint (best_advanced_cnn.h5)

Dataset

Kaggle: Digit Recognizer

• train.csv: label + 784 pixel columns
• test.csv: 784 pixel columns

How to Run (Kaggle Notebook)

1. Open the notebook: digitdetector.ipynb
2. Attach the Kaggle dataset: digit-recognizer
3. Run all cells

After running, you will get:

• submission.csv ✅

Output

submission.csv format:

• ImageId (1..N)
• Label (predicted digit)

Requirements

See requirements.txt.

Notes

• The notebook uses the Kaggle path:
  • /kaggle/input/digit-recognizer/train.csv
  • /kaggle/input/digit-recognizer/test.csv
• For local run, you must download the dataset and update the paths.

License

MIT (recommended)