Fastai v2 on Kaggle Gotchas

Note as of today kaggle notebooks docker image has fastai v2 and pytorch 1.6.0 installed, so most of this hoop jumping is no longer necessary.

Kaggle notebooks do not currently have fastai v2 installed as standard so it has to be installed. After setting up a notepad on Kaggle I ran into a few problems so thought I'd list them here to hopefully help someone else. 

Wrong version of torch

Fastai v2 requries pytorch version 1.6.0 however Kaggle notebooks come with pytorch 1.5.0. So we must install an upgraded version of pytorch first:

!pip install torch==1.6.0+cu101 torchvision==0.7.0+cu101 -f https://download.pytorch.org/whl/torch_stable.html

Wrong version of korina or allenlp

!pip install --upgrade kornia

!pip install allennlp==1.1.0.rc4

Old version of fast ai

from fastai2.basics import *

AttributeError: 'str' object has no attribute '__stored_args__'

Before fast ai was officially released the code lived in the fastai2 package. You will still see a lot of references in older posts and notebooks to this. However you should be referencing just the fast ai classes e.g. 

    from fastai.vision import * 

Note: This error message is also seen when you have an incompatible version of fastai and fastcore installed. It is recommended that you upgrade fastai and fastcore.