Question: (active) training loop about batchbald_redux HOT 3 CLOSED

Dear Pietro,

thank you for your question!

The train_loader is using the dataset active_learning_data.training_dataset. It is not immutable but changes as new samples are acquired. The overlying dataloader will thus iterate over the updated dataset.

I wanted for batchbald_redux to be independent of baal. It is based on the original BatchBALD code, which might or might not have served as an inspiration to baal.

In more practical terms, batchbald_redux uses toma which allows for automatic adaptation of GPU memory usage which potentially might allow the code to run where baal's implementation might not. (I haven't looked at that implementation in a while.)

I hope this answers your question somewhat.

Best wishes,
Andreas

from batchbald_redux.

Hi Andreas,

I really appreciated your reply. I have been experimenting with it lately and found that it fact it does change without reloading it. Thanks to your comment now I understand why :)

Re: batchbald_redux - baal relationship, I get it and I think that was a good choice as in this way the library remains self-contained. I am currently using batchbald_redux on a project and finding it amazing (+ toma is the real cherry on the cake). I was wondering about the best way to reach out to you in case I have comments/questions. Would Github issues work for this?

Once again, thank you very much for your consideration.

Best,
Pietro

from batchbald_redux.

Dear Pietro,

thanks and sorry for the late reply. I'm happy you like the library.

GitHub is one way to reach me. You can also message me on Twitter.

Best wishes,
Andreas

from batchbald_redux.