Which platform should I use for deep learning?

Part 1 (2019)
1

Hi everyone,

I just getting started the part 1, there are so many options available for setting jupyter notebook, I really can’t decide which platform I should be using, I DON’T want to make any mistakes and setup again on another platform later, I would like to know which platforms are you guys using for the course? And which one is the best?

Thanks
Gary

2

It depends how much money you are willing to spend, and also how much you want to set up from scratch.

If you don’t want to spend any money, I would recommend Kaggle Kernels or Google Colab.

If you go to the course website under server setup there are instructions for many different services.

2 Likes
3

I have used paperspace and floydhub.
I would not recommend paperspace for multiple reasons, but an easy one is they don’t have an auto-shutoff for the server…so if you forget to shutdown the server you’ll wake up to a nice fat bill.
In addition, they will conveniently, automatically and without any notice add a ‘public ip’ option charge deep in the accessory menu…and you’ll find out about it with a mysterious charge for ‘add ons’ from them every month ever after, even when you haven’t used the service at all b/c you left for a better platform (they did refund me though and explained the convoluted manner to remove it…in a menu I never used nor had any reason to go into).

Floydhub does have autoshutdown and I like their interface overall. It’s more expensive per hour than PaperSpace,but as noted, one forgetful lack of shutting down on Paperspace and Floydhub will look very cheap.
Floydhub has been having issues this past week where you can sit 15-20 minutes with the server ‘queuing’ b/c they have no free GPU systems, and that’s getting old at this point when you have work to do. Also, the load time of the server once allocated is often quite slow (I’ve had it ‘loading’ for up to 8 minutes…which you are paying for while it’s unusable).

Anyway, that’s my feedback - hopefully someone may have better/more robust recommendations vs Floydhub and if so I’m certainly interested in trying it out.

1 Like
4