Getting deeper and deeper into the weeds of deep learning and loving it :D @mcgenergy   github.com/morganmcg1 

First Open Source Pull Request Accepted :)

THE PR

Building on all the amazing work done by @lessw2020 on the Fastai Forums combining RAdam and AdamQH with Lookahead (resulting in Ranger and RangerQH respectively), I ported AdamQH over to fastai v2 to enable the use of RangerQH in the updated library…. making it my first ever open source PR!

https://github.com/fastai/fastai_dev/pull/295

I’ve pushed code to production at work in Facebook, but it feels different (great!) to push to a public resource that has benefitted myself and many other so much.

The Twist

More homework!

Screenshot 2019-11-16 10.38.27.png

Jeremy Howard had the great idea of making the Quasi Hyperbolic element a parameter that other steppers could call, making it more flexible and applicable to other optimizers. I still have to look into whether my slow brain can figure it out, but on the surface it sounds like a promising idea.

More work <3

[1/4][Kaggle] RSNA Intracranical Hemorrhage Competition Series

Mish-EfficientNet, a more usable EfficientNet?