Digit Recognizer is a Kaggle competition where using the dataset you have to create a classifier that can classify handwritten images into digits.

Here no pre-trained CNN or predefined architecture is used, this is a custom CNN architecture.

While doing this we'll go through

• Data augmentation using ImageDataGenerator
• Building custom CNN architecture
• Visualizing CNN (filters and feature maps)

• Getting started
• Kaggle Competition Results
• Models performance and visualizations
  • Visualizing Filters
  • Visualizing Feature Maps
• License

The notebook is available on Kaggle to work in the same environment where this notebook was created i.e. use the same version packages used, etc...

If you are interested in the model the you can find that in the Output section of the notebook.

The best model has an accuracy of 99.5%

Count plot for labels

The model is trained for 50epochs and below is the last epoch's results

Learning curves

Confidence matrix

Some predictions on the validation set

Image of the 96th filter of the 1st conv layer

Images for only first 20 filters in the 1nd conv layer

Images for only first 10 filters in the 2nd conv layer

We'll visualize feature maps for the digit 7

Feature maps by 1st conv layer

Feature maps by 2nd conv layer

APACHE LICENSE, VERSION 2.0