running environment

[SylvanLiu]

Thanks for contributing such a reproduction for the gorgeous essay.
However while running those *.ipynb files, I actually suffered dozens of problems which I found are arisen by the python and libraries versions not match.

So I think it would be better if you can add some constraint or requirements to the readme page.
As your codes can cause confilct with the latest version of pytorch and fastai.

================================================================
Firstly, it can be ran with with 3.6.9
And for the lib version conflicts, I mainly solved by the following two steps.

step I:
pip3 install torch==1.4.0
pip3 install torchvision==0.5.0
The higher versions will lead the program into {Exception: It's not possible to apply those transforms to your dataset: expected scalar type Float but found Double}

step II:
pip3 install "sentencepiece<0.1.90" wandb tensorboard albumentations pydicom opencv-python scikit-image pyarrow kornia
catalyst captum neptune-cli

Also noteworthy is that 'neptune-cli' may have potential issues while installing, so you should download the file and install it offline first instead.
https://files.pythonhosted.org/packages/ba/ae/0cdcb9d20bf3b2324d21cfd11dc05246972e9759e7609739c6ce0ef01639/neptune-cli-2.8.23-1.tar.gz

step III:
For the rest of libraries needed, like fastai; SlideRunner and SlideRunner.dataAccess, they are okay to be installed up to date.

================================================================
Potential Issue:
If you have no confidence with your network of where you are.
It may suffer some potential problems while downloading the dataset files from 'figshare' by using any python network-request libraries like the demo in the 'Setup.ipynb'. no matter urllib.request or python standard lib 'requests'.
I tried a lot of methods, like make up my request headers and cookies, but all not working,
The final method helps me download them all in final, is also the most 'noob' way: inserting every url link into your chrome...
And the files are all successfully downloaded by the chrome default downloader instead of python scripts of any kine.