Handwritten digit OCR

Handwritten digit OCR is a set of Jupyter notebooks notebooks that train convolutional neural networks to recognise handwritten digits. It is written in Python with TensorFlow and Keras, and each notebook opens in Google Colab.

Experiments

The notebooks compare training set-ups on the same task:

• a CNN trained for 50 and for 200 epochs;
• the same with early stopping, which ends training when validation stops improving;
• transfer learning from VGG16, a network pre-trained on ImageNet, including a variant with a higher learning rate;
• a notebook comparing the results.

The trained models are saved in Keras format alongside them.

See also

• Machine learning and data
• Machine learning