gyhandy / channel-wise-lightweight-reprogramming Goto Github PK

Thanks for your excellent work. I noticed that an adaptive task oracle is proposed in section 3.2. But i could not find it in the code. Is it a one-hot vector or something else? Could you please give me some hint about how to implement it? Thanks in advance!

Hi, first of all - thank you for writing such a well-thought out paper , with a great codebase.

I was going through the training pipeline, and was curious about how exactly the inference would take place in a Continual learning setting? According to the paper, there will be a pool of task-specific CLRs. But in the code I can see that, a separate model is trained for each task.
In that case, can you explain how the task performance graphs were generated? Since the models are completely isolated from each other, isn't this more or less similar to normal fine-tuning on top of a large pretrained backbone?

Apologies for the long question. I am just curious about how to utilize it for continual learning - i.e:, how it's actually tackling the Catastrophic Forgetting issue.