How we're developing fastai_v1


(Stas Bekman) #42

and this :slight_smile: I’m being anal:

-    module = ''
+    module = '''#################################################
+### THIS FILE WAS AUTOGENERATED! DO NOT EDIT! ###
+#################################################
+'''

(Michael Skinner) #43

@jeremy timing is a funny thing isn’t it?

I just recently finished a very early draft of my own re-exploration of some key concepts behind the current fastai library: https://github.com/mcskinner/kerosene.

I refactored and rearranged what I saw as the core fitting fitting loop functionality, specifically with the ability to nicely program hyperparameters within an optimization, and then added a few bits of my own. This follows the lines of thinking in today’s LayerOptimizer and Callback stacks.

How do you think this could be most useful with respect to the development of fastai V1?


(Jeremy Howard) #44

Thanks for shaing @mcskinner - there’s some useful ideas there. As it happens, @sgugger is likely to be looking at doing a proper fit/optim/sched api in the next day or two, so this is indeed good timing. I’ll started a new thread for it, if you’re interested in getting involved.


(Jeremy Howard) #45

Here’s the new thread:


#46

I love how you proposed a whole new way to do all those things! Lots of food for thought.