Working with the lib, I've come to the conclusion that it might be useful to have a meta trainer. This trainer would be used to do a series of experiments (k-fold, hyper parameters seach, etc).
This would be particularly useful coupled with the tasks at initialization and at the end (producing reports, etc.)
Instead of epochs, we could add another level : experiments.
Of course, per epochs and updates tasks would be directly transferred to its children.
Working with the lib, I've come to the conclusion that it might be useful to have a meta trainer. This trainer would be used to do a series of experiments (k-fold, hyper parameters seach, etc).
This would be particularly useful coupled with the tasks at initialization and at the end (producing reports, etc.) Instead of epochs, we could add another level : experiments. Of course, per epochs and updates tasks would be directly transferred to its children.
Any other ideas that'd be useful?