adrianjav / ml-code-base Goto Github PK

Example of usage:

@timed
def train(model, x, y):
    ....

Create a class to read, write and acces arguments in a global way.

Design choices:

• Modifing the type of an argument in run time is not allowed (although it is allow to change non-leaf nodes)
• As changing leaf-nodes as args.leaf = smth may lead to ask whether this is a global or local change, it is not allow. Instead one should write args.load({'leaf': smth}, scope='global') and args.load({'leaf': smth}, scope='local') instead.

Features:

• Change arguments locally and globally.
• Global access.
• Load from strings, dictionaries and yaml files.
• Write the (global) settings in a file after the run has finished (transparently + configurable).
• Choose whether or not remove the file after completion (folder?)

Closed Questions:

• How do we mix the guard class with the arguments? Sometimes saving might not be desired.
  • A. Globally configurable.

Open Questions:

Write a module for automaticly create the directory system for the experiments. Those directories will be accesed via a global variable called dirs.

Configurable options:

• Structure of the folders. Default: root -> logs, plots, checkpoints

Create a module that allow to save checkpoints of certain objects if some kind of exception happens.

Functionalities:

• Declare classes that have to be saved using decorators (@failsafe)
• Declare classes that have to be saved by inherating from a metaclass.
  • There you will say which loader/savers use (or take the default from the class methods)
• Allow instances to modify their behaviour on the fly. on instance creation.
• Declare which functions are prone to suffer this, e.g., the train method when using a Pytorch's model (we don't care if it crashes after training).

• There we will put the try/catch
• Specify whether or not save at the end? Change the destructors?
• How can we allow it to save using a different criteria? (best model)

• If we tell the library to load a checkpoint, no other code should change (transparency).

• Modify the constructor functions.
• There should be no difference between the model when it was saved and after being loaded.
• How can we distinguish different instances of the same object on the code? This methods should be used for special classes only. We use a counter of objects.

Open questions:

• Should this module take care of changes in the code after crashing?