Difference in the way of copying the samples from task to buffer memory of GEM and A-GEM in update_memory function

[amalapuram]

Hi avalanche team,

While going through the implementation of GEM and A-GEM, I found the following differences in storing the samples from task to buffer memory.

In GEM implementation (lines 127 to 158,https://github.com/ContinualAI/avalanche/blob/master/avalanche/training/plugins/gem.py), authors are copying from the first sample of the current task till the sample with index value equal to self.patterns_per_experience into the buffer memory. In other words, self.patterns_per_experience sized several samples from the first indexed sample of the current task are copied into the buffer memory.

However, in A-GEM (line 128,https://github.com/ContinualAI/avalanche/blob/73210af748d5de74e1ae73c00f525454b4129cd2/avalanche/training/plugins/agem.py) current task samples are stored in the buffer memory using the random_split function.

My question here is why there is a different strategy for GEM and A-GEM. Technically, A-GEM is an efficient variant of GEM in terms of the number of constraints in the objective function. Please let me know, If I miss anything

[AntonioCarta]

If I remember correctly, GEM and A-GEM do not specify the buffer policy, so any reasonable policy should be ok. I don't think there is any specific reason why the two policies are different, we just implemented the algorithms in different moments. We could improve Avalanche implementation by making the buffer policy a parameter, so that users can change it.

Of course if you need a specific policy it's easy to change the plugins to achieve what you want.

[amalapuram]

Thanks, @AntonioCarta, for your response. With the current avalanche version, is there scope to specify the buffer policies for these methods?

[AntonioCarta]

For the moment, the easy solution would be to make your own GEM by copying the original GEM plugin and modify it to use a different Avalanche buffer.