danijar / layered

Clean implementation of feed forward neural networks
MIT License
240 stars 33 forks source link

Color training costs based on label value #27

Open danijar opened 8 years ago

danijar commented 8 years ago

Need to map label space from 0 to 1.

greninja commented 7 years ago

Can you describe more?

danijar commented 7 years ago

Sure. The idea is to plot the loss for each class as separate lines in the loss diagram. This requires to select colors for the classes. Mapping the possible labels between 0 and 1 would allow to color those lines from a Matplotlib color map. However, now I think hard coding colors from a palette (colorbrewer2.org) is better.

greninja commented 7 years ago

I did not understand what exactly is meant by 'class' here? Is it the different methods to train the NN given in gradient.py?

danijar commented 7 years ago

Haha I see. By classes I mean the possible labels in a classification dataset.