12 comments

[ 2.8 ms ] story [ 40.1 ms ] thread
https://github.com/blue-season/pywarm PyWarm is a lightweight, high-level neural network construction API for PyTorch. It enables defining all parts of NNs in the functional way.
I don't see much of a reason you want to clutter your code with that, if you need a wrapper around pytorch use fastai not some obscure library nobody has ever heard about.
Thank god you weren’t around when keras was released.
Use fastai if you want your code to be unreadable, unmaintable and prone to mysterious bugs.
Apart from the official tutorials what are the best resources out there for learning PyTorch?
What exactly are you looking for? I think Pytorch is so ergonomic that there is no need for other resources. The only gotchas I found were in the data loading utilities.
The #1 NLP repo on Github and the author of the most popular NLP course is moving from PyTorch to Tensorflow 2.0/Keras

https://twitter.com/GokuMohandas/status/1174497967232782336

>We’ll still use @PyTorch but more for research lessons, I’ll post more on this decision soon! I did deliberate for several weeks on this though but ultimately it sped up development and decreased overhead for the practical lessons by a lot.

> decreased overhead for the practical lessons by a lot.

What do they mean by that?

Presumably that it was easier to use to illustrate the actual content as opposed to implementation details. This is definitely the case, at least in TensorFlow 1.x there were a lot of details which made converting from [the way we think about the problem] to [practical implementation] more cumbersome. Some of these existed for good reason (eg. performance) while others were simply architectural cruft which the benefit of hindsight makes unnecessary.
This is really old, it came out in 2016.

Why did you reshare it ?

The framework which put my trust back into frameworks. Very good!