Adversarial bias mitigation #5269
Conversation
…s/adversarial-bias-mitigation
…s/adversarial-bias-mitigation
|
I have misspelled "adversarial" all my life 🤦🏻 . Seems like autocorrect bailed me out a lot. |
| !!! Note: | ||
| Intended to be used with `AdversarialBiasMitigator`. | ||
| trainer.model is expected to have `predictor` and `adversary` data members. |
There was a problem hiding this comment.
You could, if you wanted to, put in a check for this condition and throw an exception if it isn't met.
| } | ||
|
|
||
| trainer.model.predictor.zero_grad() | ||
| batch_outputs["loss"].backward() |
There was a problem hiding this comment.
If "loss" and "adversary_loss" don't use exactly the same computation graph, does that mean that parts of the computation graph of "adversary_loss" could stick around when we don't want them to?
There was a problem hiding this comment.
That's a really good point about part of the computation graph not getting erased! Upon further reading, it looks like the computation graph will stay around until adversary_loss goes out of scope. So I added this in the callback to remove all references to the adversary_loss in the graph and instead keep a view of the loss that's not in the graph:
# remove adversary_loss from computation graph
batch_outputs["adversary_loss"] = batch_outputs["adversary_loss"].detach()
allennlp/training/optimizers.py
Outdated
| @@ -278,6 +278,11 @@ def __init__( | |||
| " Alternatively, you can remove this optimizer from the provided `optimizers`" | |||
| " if it is not relevant to a particular parameter group." | |||
| ) | |||
| # default optimizer is required, but may not be used | |||
There was a problem hiding this comment.
"must not be used" or "using it is optional"?
There was a problem hiding this comment.
I know what you mean from context, but it would be easier to read if worded unambiguously.
| of some protected variable Z. Informally, "knowing Y would not help | ||
| you predict Z any better than chance" (Zaldivar et al., 2018). This |
There was a problem hiding this comment.
Should this be the other way round? "knowing Z would not help you predict Y"? Or is it stating that knowing the outcome shouldn't give you the information about the protected variable?
There was a problem hiding this comment.
The latter, it's stating that knowing the outcome shouldn't give you information about the protected variable.
|
|
||
| vocab : `Vocabulary` | ||
| Vocabulary of predictor. | ||
| predictor : `Model` |
There was a problem hiding this comment.
This is not strictly an issue. We use the term predictor differently in the library elsewhere; should we change the name here? If this is adhering to the paper's terminology, it's probably okay to keep it as is.
There was a problem hiding this comment.
Yes, I'm adhering to the paper's terminology.
Changes/additions proposed in this pull request:
AdversarialBiasMitigator, a Model wrapper to adversarially mitigate biases in predictions produced by a pretrained model for a downstream task. Tests are in: added AdversarialBiasMitigator tests and model allennlp-models#281.IndexOutOfBoundsExceptioninMultiOptimizerwhen checking if optimizer received any parameters.MultiOptimizerso that while a default optimizer is still required, an error is not thrown if the default optimizer receives no parameters.Depends on allenai/allennlp-models#281.