I am looking for a bit of advice about starting the course. I have been considering starting the course for a year, and was about to do it, but I see everyone excited about the new fastai v1, and that the course using it will come out in 2019.
Would you recommend:
Start the course now, and use fastai v0.7?
Start the course now, and use fastai v1?
Wait until the new version of the course comes out, and use v1 then?
You should start now and use fastai v0.7 as the notebooks are written!
Seriously, you can easily finish the course now until the new version comes out. If you wait until january or february you easily miss out on 2-3 months of time you could spend learning the stuff that is already in the fastai courses.
And the API changes from v0.7 to v1.0 are not that big. Furthermore thatās not the point! If you have a solid understanding of how to train neural networks the api is just a means to an end.
Also I would start on paperspace. There are plenty of 15$ coupon codes to find here in the forum and it is the easiest setup. It would be very frustrating to spend like 10h setting up the environment alone. And I had no luck with Google Colab so far. I did the first 4 lessons on paperspace and I have maybe spend 5$ or so because of the coupon code.
It depends on how comfortable you are with code. If you are somewhat comfy, fastai 1 is well documented and you should be able to re-write the 0.7 notebooks with a modest amount of research. Expect a few curveballs, but a lot of those involve fastai 1 being simpler than 0.7, and āwhy do I only write one line of code when the notebook has six lines?ā The other five lines are probably defaults in the one line. You can still tweak those parameters, but your initial models can bypass those decisions and just roll with the defaults.
Definitely start now. Practice, practice, practice. This is the key to build and gain intuition around a concept. And this takes a long time.
The API changes is huge. I think you intended to say that the changes in DL concepts (and knowledge) from the old course to the new course are not that huge.
Thanks all ā I see the consensus ā I just signed up for paperspace yesterday, and hope to start on the content later today (itās my first time working on a cloud server, so there was a bit to learn there )
One last advice question for now: When going through the course, do you recreate the notebooks from scratch? Or do you generally copy the notebooks from git, and run/edit/rewrite as you go?
Jeremy recommends watching the videos, later trying to replicate the model from scratch, and only checking the existing notebooks when you get stuck. Not everybody is able to pull this off, so do as much you can, as long as it is more than just clicking straight through the notebooks. Spend as much time in the code as possible, and a few videos in you will be rolling your own.
I asked myself same question and decided to go ahead with 2k18: Jeremy recommends going through the material a first time fast to learn the concepts and then re-watch the videos and dig deeper. So planning doing first pass on 2k18 and will dig deeper on 2k19.
As for Collaboratory it is a easy setup and a great platform, I combined info from different blog/forum posts and get it set up in a couple of hours, I can share the notebook on Github if you want so you can get it ready in couple of minutes. So far very happy with it, I always have 11GB GPU for myself, so same than paper space but free
Thanks a lot! Iāve just started this course two days ago and I canāt access Paperspace for some reason, Crestle is giving issues once in a while. It could be a network issue from my side I believe. Anyway I donāt mind the extra hassle of setting my own development environment through Google Colab and your notebooks will be of great help for me!
Also beware of Google. I set up google ML engine for this course and within 4 days some glitch billed me Ā£7000 !! They have been sorting this out for 3 weeks and given no feedback and no promises that the bill will be waived! So iāve switched to paperspace it is simple and has no weird hidden charges to it.