Training a CNN with the EMNIST Balanced dataset with different kernel matrix sizes and evaluating accuracy
In this python notebook, the model layers are printed after training to get a better look on the inner workings of a Convolutional Neural Network:
Kernel sizes tested:
- 8
- 16
- 24
- 64 + 128
Also a simple dense layer instead of the convolutional layer was tested, which resulted in overfitting:
Same test was done for a few Chinese characters, just to assess how it would work with something different:
