Naive Strategy #1491
-
|
Can you explain what is actually happening in Naive strategy? Thanks |
Beta Was this translation helpful? Give feedback.
Replies: 1 comment
-
|
Naive strategy is the most basic type of a continual learner. For every arriving experience, it iterates through the samples of the experience's dataset and updates the model with an SGD-based optimizer without any additional regularization or augmentation methods. It is also called sequential fine-tuning. If you are looking for more technical details, an SGD-based template (the super-class of the Naive strategy) in Avalanche provides a sequence of callbacks that are triggered one after another when you call "create dataloader"-> "get batch i" -> "forward" -> "calculate loss" -> "backward" -> etc. Here is a a plot from https://arxiv.org/pdf/2302.01766.pdf that shows some of the call backs in the Base and SGD templates: 
Link to the API doc and source code: |
Beta Was this translation helpful? Give feedback.
Naive strategy is the most basic type of a continual learner. For every arriving experience, it iterates through the samples of the experience's dataset and updates the model with an SGD-based optimizer without any additional regularization or augmentation methods. It is also called sequential fine-tuning.
If you are looking for more technical details, an SGD-based template (the super-class of the Naive strategy) in Avalanche provides a sequence of callbacks that are triggered one after another when you call
strategy.train([exp]). Here is an example at the high level:"create dataloader"-> "get batch i" -> "forward" -> "calculate loss" -> "backward" -> etc.
Here is a a plot from https://a…