Improvements to the code

[AlexEMG]

• make loops (for different shuffles)
• include evaluation code
• What if it crashes? Can we make a variant that restarts from intermediate snapshot?
• generally discuss coding style
• switch permission at the end of the script (so group can access it)
• assignment of models to gpus (this could be improved)

[AlexEMG]

Issue for all of you!

[sfmig]

I this could be a good starting point: script with contributions from Jonas, Sabri and me.

Things still to add:

• shuffles
• gpu assignment
• evaluation bit
• saving snapshots etc

[biol-jsh]

Regarding "What if it crashes? Can we make a variant that restarts from intermediate snapshot?"

we are wondering if the learning rate will take the start snapshot into account? I remember reading that it doesn't, but I might be wrong or it might have changed.

[sfmig]

Regarding "What if it crashes? Can we make a variant that restarts from intermediate snapshot?":
added here

For now no learning rate adjustments are implemented

[sfmig]

Regarding: 'include evaluation code'

This is in the code by @sabrinabenas here