CNNMnist is a mini-project to apply convolutional neural network to MNIST dataset. The whole flow contains two parts:
- Pipeline for model training: task_01_cnn_mnist_pipeline.py
- Script to classify an input image by a trained model: task_02_cnn_mnist_predict.py
- Model training pipeline can be started by
task_01_cnn_mnist_pipeline.py -i <input_dataset_dir> - Recognize the digit on an input image can done by
task_02_cnn_mnist_predict.py -i <input_image> -m <trained_model>
Demonstration of the steps above can be found in demo_task_01_cnn_mnist_pipeline.ipynb and demo_task_02_cnn_mnist_predict.ipynb. Details about the model setup, modules of pipeline steps can be found in class CNNMnist.
Based on training history, we can conclude the model from epoch 5 is good for the purpose of this mini-project.
Out best model performs quite consistently across training, validation and test sets without overfitting.
| Dataset | Loss | Accuracy |
|---|---|---|
| Training set (48k) | 0.0644 | 0.9863 |
| Validation set (12k) | 0.0536 | 0.9859 |
| Test set (10k) | 0.0436 | 0.9874 |
Given a input image and our trained model, digit on the image is correctly predicted.
- Numpy
- Keras and Tensorflow
- Dataset CNNMnist You can run ./data/download_data.sh to download it.