A Python Jupyter notebook with a simple computer vision implementation for creating a very acceptable submission for the Digit Recognizer challenge in Kaggle: https://www.kaggle.com/competitions/digit-recognizer
The image above shows:
- The digit number 4777
- A slightly rotated version of itself
- A slightly zoomed version of itself
The algorithm performs these augmentations for every digit in the dataset, effectively increasing its size to three times the original. The labels are triplicated to compensate for this.
The implemented algorithm returned a testing accuracy of
In the image above, both the training/test accuracy and error can be seen for three epochs, generated by the notebook as is.