Base strategy adaptation for revisiting the exact same data with self.batches

[Mattdl]

Hi! I'm trying to finetune an independent model within a strategy on the exact same data as seen for the current batch.
However, the losses for this model I want to add to the loss of the model being optimized currently in the base strategy.

In practice, this proves difficult as before/after_forward happen before the calculating the loss fo the current network. But after loss+=criterion(...) for the current task I would like to process the loss for the independently finetuned model, so I can just add to the loss as well within the strategy.

What I tried now is to implement it in the before_backward, and to iterate over the exact same data:
def before_backward(self, strategy, **kwargs): for mb_task_id, (mb_x, mb_y) in strategy.current_dataloader[strategy.mb_it].items(): .... # Process
However, I get an error because I can't index the dataloader. I think this would be solved if there was a self.batches in the base strategy which allows to easily iterate over the batches again.
This would allow to:
def before_backward(self, strategy, **kwargs): for mb_task_id, (mb_x, mb_y) in self.batches.items(): .... # Process

[AntonioCarta]

If I understand correctly your problem, you may be able to solve it by modifying also the criterion?
Either way, I think you are right that we need to provide access to batches. That part of the BaseStrategy was written before AvalancheDataset supported multi-task data.

I will work on this (#452).

[Mattdl]

Thanks for the quick reply! So to be clear, the batches (and self.mb_it) refer to minibatches for multiple tasks for a single round of processing? For me this was not directly clear from the code, so might be interesting to put in the Base_Strategy documentation (:
Anyway, in that case, I only have 1 batch in batches, so I can just use the self.mb_x to solve my problem in the before_backward.

Thanks for the help!

[AntonioCarta]

yes, you are correct. It is only needed to average the loss among multiple tasks before doing the backpropagation. I will add more documentation about the internal state of the BaseStrategy soon.