I’m really neutral about this, but I think that talking in polish on public forum is kinda rude, so maybe some folks would prefer doing it on slack.
So who’s crazy enough to watch the lectures in the middle of the night beside me?
I’m really glad that GCP offers free credits and way more than we need to complete this course. I have a PC that could handle DL, but I’ve tried to setup all the stuff needed for fast.ai on it and wasted 2 weekends on it.
The ‘normal’ setup has Tesla100 (http://course-v3.fast.ai/start_gcp.html) and I don’t care much about the price, because it will be hard to use whole 300$ even for both parts of the course. I really like that GCP has more readable interface compared to AWS.
Is there a big performance jump between T80 and T100?
I’ve run the notebook and player around a bit and got 0.999 accuracy on mnist dataset, just using unfreezing and lr_finder. Had 0 idea what I was doing, but it was fun.
yeah that’s a lot, just need to remember to stop the instance
but then it’s for 1 year, so part 1 part 2 kaggle other
$300/300 days = 1 hour a day at $1/hour
It was linux, ubuntu, my main problem was …well everything lol, one time there were some issues with CUDA, another time with something else. I don’t really remember (it was sometime in the spring), but I do remember the fact that I wanted to throw the PC out of the window.
I’m going to try again, but this time with docker, so at least it won’t take so much time (to know it won’t work . Today I’ve got GCP up and going within like 10 minutes and it made me so happy.
I dont remember what cause more problem but full set up of GPU environment on my desktop took me 2-3 days including OPENCV compilation from source Pytorch best to install by conda installer
hi @jeremy
just wanted to get your opinion as there seem to be some confusion at our thread ‘Study Group Polska’ about if we can use Polish language there or should stick to English?
from your post i get the impression that native language is welcome here?
I set up fastai (previous version) on ubuntu 16, ubuntu 16 gnome, ubuntu 18 without any problems.
I think the trick is to NOT install CUDA on your own, let env do everything. And avoid pip when possible.
I just have lots on env right now
one for keras, other for vanilla pytorch, other for fastai, another for some DL style transfer demo. Because each requires different CUDAs and diff ver of pytorch etc
When I tried installing CUDA following instructions from nvidia, I ended up re-installing everything
If you don’t menage to make it work, follow Jeremy advice, and just use Google Cloud. It takes too much time and attention, that would be better spend on focusing on lessons. And tackle setting up own rig after the course.