Kaggle Kernels Now Support GPU for Free


(Cedric Chee) #1

Good news! You can add an NVIDIA Tesla K80 GPU to your Kernel through the “Settings” tab. Here’s an example of training a model on a GPU.

a

I gotta try this with fastai soon.


#2

Looks good. Thanks for sharing! How can one use fastai with it? Thanks!


(Jeremy Howard) #3

You can use fastai with kaggle, since it is pip installable. Although the version on pypi is a little old now. I’ll try to update it soon.


(Martin) #4

Is there a way to use pip install git+https://github.com/fastai/fastai/... to install the latest version of FastAI?


(Alex) #5

should work pip install git+https://github.com/fastai/fastai.git


(Jeremy Howard) #6

I’ve just updated the version in pypi to the latest - don’t know how long it takes kaggle to update however. The new version is 0.7.0.


(Sanjaya Gebrial) #7

fastai is one of the packages that come pre-installed with kaggle kernel. Does it use pip or github to install them though?


(Cedric Chee) #9

FYI, I’ve moved this post to the new “Deep Learning” category.


(Jeremy Howard) #10

It uses pip.


(Davide Boschetto) #11

There’s a problem that I guess has to do with requirements on windows, so I’m tagging you @jeremy .
Doing “pip install fastai” on Windows 10 tries to look for “torch<0.4” and fails, saying that Pytorch does not provide packages for PyPI.
Any chance this has anything to do with fastai requirements? Does it really need torch<0.4.0 (No Windows)?


(Jeremy Howard) #12

This is a thread about Kaggle kernels - you’ll be more likely to get an answer if you post to or create a thread about windows (and you don’t need to tag me on questions that others can answer).

(As mentioned in the readme, pip isn’t recommended for installation - it’s only really there to support kaggle kernels, or those who can’t use conda for some reason.)


(Davide Boschetto) #13

I figured I could ask here given that you updated us about fastai 0.7.0 in this topic, but I understand. Thanks anyway, I’ll keep using git then!


(Kevin Mader) #14

For pytorch<0.4 on windows
peterj123 made a conda package that covers most functionality well, see https://github.com/pytorch/pytorch/issues/494


(Kouassi Konan Jean-Claude) #15

OK, me I used the github version putting fastai/fastai in the github field and it works as a charm.