A convolutional neural network achieving 99.08% accuracy on MNIST, built with PyTorch. Features a production-style pipeline with reproducibility, CLI configuration, structured logging, and comprehensive error analysis.
Input [1, 28, 28]
→ Conv2d(1→32, 3×3) → BatchNorm → ReLU → MaxPool(2×2) [32, 14, 14]
→ Conv2d(32→64, 3×3) → BatchNorm → ReLU → MaxPool(2×2) [64, 7, 7]
→ Conv2d(64→64, 3×3) → BatchNorm → ReLU [64, 7, 7]
→ Flatten → Linear(3136→128) → ReLU → Dropout(0.5)
→ Linear(128→10)
| Metric | Value |
|---|---|
| Test Accuracy | 99.08% |
| Parameters | 458,890 |
| Training Time | 35s (5 epochs, Apple MPS) |
| Total Errors | 92 / 10,000 |
pip install torch torchvision matplotlib numpy
# Train with defaults
python mnist_cnn_classifier.py
# Custom hyperparameters
python mnist_cnn_classifier.py --epochs 10 --batch_size 256 --lr 0.0005- Reproducibility — fixed seeds across all random generators for deterministic results
- Data augmentation — random rotation (±10°) and translation (±10%)
- Batch normalization — after each conv layer for stable training
- Dropout (0.5) — in the classifier head to reduce overfitting
- LR scheduling — ReduceLROnPlateau for adaptive learning rate decay
- Auto device selection — CUDA → MPS → CPU detection
├── mnist_cnn_classifier.py Training & evaluation pipeline
├── outputs/
│ ├── best_model.pth Best model weights
│ ├── checkpoint.pth Full training checkpoint
│ ├── 01_samples.png Dataset samples
│ ├── 02_class_distribution.png Class balance
│ ├── 03_training_curves.png Loss & accuracy curves
│ ├── 04_confusion_matrix.png Prediction analysis
│ ├── 05_wrong_predictions.png Most confident errors
│ ├── 06_correct_predictions.png Correct prediction samples
│ ├── 07_conv_filters.png Learned conv filters
│ └── 08_feature_maps.png Layer 1 activations
└── README.md
- Python 3.10+
- PyTorch 2.x
- torchvision
- matplotlib, numpy
MIT