Raw implementation of MNIST cnn trained for 2389 classes of apple emojis
Based on MNIST CNN, using emoji database as basis of my augmented one (need to pull that).
All preparation functions are placed in augment.py, '''train.py''' describes model and how to interact with data.
Also you can just use final.h5 and predict.py to predict emojis.
The model was trained on .png file with 4 channels but this .pngs were read as 'rgb' so in predict.py you can see strange behaviour of adding green background to recognizable .png.
Train again with greyscale