Modifying training scripts in auto3dseg

[kretes]

I wanted to give a try to auto3dseg.
Due to setup of my environment and pytorch/pytorch#973 I am using a workaround where I call torch.multiprocessing.set_sharing_strategy("file_system") at the beginning of any of my training scripts.
With Auto3DSeg - the default training fails, so I go to work_dir/{algo}_{fold}/train.py and add this line - then it works.

The problem is that this is not a scalable solution, since it requires manual modification.
Question : Is it possible to specify that as 'custom code to be included in training script'?
Or maybe there is a way to modify the templates once and use them since now on?
The last resort is patching script files with sed or something like this, but I would really like to avoid that.

[kretes]

Ok, it seems there is a way to do it once.
One need to copy the generated templates, modify them and later provide AutoRunner with templates_path_or_url param pointing to the dir with custom-modified scripts. I'm still to see if that works ok