A CNN trained from scratch on the EMNIST Balanced dataset to classify handwritten digits and letters. Includes a full training pipeline with live experiment tracking via Weights & Biases, and an interactive web app where you draw a character and get a prediction instantly.

• Handwritten digit + letter classification (47 classes)
• CNN built in PyTorch with BatchNorm and Dropout
• Training augmentation — random rotation, translation, and scaling
• Live experiment tracking with Weights & Biases (loss, accuracy, LR, gradients, sample predictions)
• Best model auto-saved to checkpoints during training
• Per-class accuracy breakdown with confusion matrix
• Local inference server (Flask) + canvas drawing web app
• All hyperparameters controlled from a single config file — no hardcoded values

• Python
• PyTorch + Torchvision
• Weights & Biases
• Flask + Flask-CORS
• scikit-learn
• PyYAML
• Pillow
• NumPy
• Jupyter Notebook

emnist-classifier/
│
├── data/
│   ├── raw/                  ← EMNIST downloads here automatically
│   └── processed/
│
├── src/
│   ├── model.py              ← CNN architecture
│   ├── train.py              ← Training loop + WandB logging
│   ├── evaluate.py           ← Test evaluation + per-class report
│   ├── dataset.py            ← Data loading, transforms, augmentation
│   └── utils.py              ← Config loader, checkpointing, AverageMeter
│
├── configs/
│   └── config.yaml           ← All hyperparameters live here
│
├── notebooks/
│   └── exploration.ipynb     ← Visualize samples and class distribution
│
├── app/
│   ├── server.py             ← Local Flask inference server
│   └── index.html            ← Drawing web app
│
├── checkpoints/              ← best_model.pth saved here after training
├── requirements.txt
└── .gitignore

This project uses the EMNIST Balanced dataset:

• 47 classes — digits (0–9) and letters (visually ambiguous pairs like C/c are merged)
• 112,800 training samples, 18,800 test samples
• 28x28 grayscale images, same format as MNIST

Dataset downloads automatically via torchvision.datasets.EMNIST on first run.

Classes: 0 1 2 3 4 5 6 7 8 9
         A B C D E F G H I J K L M N O P Q R S T U V W X Y Z
         a b d e f g h n q r t

git clone https://github.com/DagaVedant/EMNIST-Character-Classifier.git
cd emnist-classifier

pip install -r requirements.txt

Free account at wandb.ai, then:

wandb login

To disable WandB, set wandb.enabled: false in configs/config.yaml.

python src/train.py

• Downloads EMNIST automatically (~550MB, one time)
• Trains for 15 epochs on GPU (GTX 1650 ≈ 20–30 min)
• Saves best checkpoint to checkpoints/best_model.pth
• Streams all metrics to WandB in real time

# Console report only
python src/evaluate.py

# Also log confusion matrix to WandB
python src/evaluate.py --wandb

# Start the inference server (requires a trained checkpoint)
python app/server.py

Then open app/index.html in your browser. Draw a character on the canvas and click Classify to see the top 5 predictions with confidence bars.

Raw EMNIST images (28x28 grayscale)
        │
        ├─ Orientation fix (EMNIST is rotated by default)
        ├─ Augmentation: rotation ±10°, translate 5%, scale 0.9–1.1
        ├─ Normalize
        │
        ▼
   Conv Block 1: Conv→BN→ReLU→Conv→BN→ReLU→MaxPool→Dropout
        │
   Conv Block 2: Conv→BN→ReLU→Conv→BN→ReLU→MaxPool→Dropout
        │
   Classifier:   Flatten→Linear(512)→ReLU→Dropout→Linear(47)
        │
        ▼
   Top-5 predictions with confidence scores

Training — Adam optimizer with ReduceLROnPlateau scheduler. Validation split held out from training data. Best model saved automatically based on val accuracy.

Inference — canvas drawing is sent as a base64 PNG to the Flask server, preprocessed (inverted, cropped to bounding box, resized to 28x28), and passed through the model.

Everything lives in configs/config.yaml:

The model is evaluated using:

• Overall test accuracy
• Per-class precision, recall, and F1
• Confusion matrix (via WandB)

Expected test accuracy: ~85–88% on EMNIST Balanced.

• Add symbols and custom shape classes with synthetic data generation
• Export model to ONNX for faster inference
• Deploy web app publicly with a cloud-hosted model
• Add real-time classification as you draw (no click needed)
• Experiment with ResNet or EfficientNet backbones

This project is open-source and available under the MIT License.

Developed as a personal machine learning project exploring CNNs, experiment tracking, and end-to-end model deployment.