Re:Deep Learning with PyTorch – An Unofficial Startup Guide
by joshvm on Monday January 16, 2017 @06:49AM
(#13409473)
Sure:) Perhaps I should have made clear that I was referring to the fact that CS231N follows the recommendation in the blog post - i.e. lots of exercises where you implement things in primitive Numpy - rather than a broad "it's a better course". CS231N is frustrating in the real-world respect, although you're given some general recipes e.g. use ReLU, use Adam, etc.
That said, I'd be interested to see some actual results from people taking what they learned in fast.ai (with no prior knowledge)and applying it to a new test set. It's easy to be impressed with the examples where you're walked through it, but figuring out what works on your data is whole new kettle of fish!
However, my comment was about how much the theory helps in the real world (and I don't think my quote was misleading. I think the author still advocates learning theory before using the frameworks)
I think that the fast.ai approach of showing solutions, then explaining the theory is really good.
For example in lesson 2 Fast.ai shows how to finetune a VGG model to get (real!) state of the art performance on a real-wordl, two-class image recognition problem by adding an additional dense layer[1].
In CS231N (which I haven't done, although I have glanced through the notes) this isn't really said anywhere. It's kind of implied in a lot of places, and if you study it you'll be able to work it out for yourself.
That isn't bad, but I'm unconvinced that gives the huge "real world" advantage given the amount of time it takes to get to that point.
I strongly agree they are complimentary though.
[1] http://wiki.fast.ai/index.php/Lesson_2_Notes#Adding_a_Dense_...