Contextualized bias mitigation #5176

ArjunSubramonian · 2021-05-05T08:18:02Z

Changes proposed in this pull request:

A Model wrapper to mitigate biases in contextual embeddings upon training/finetuning on a downstream task.

…g-debiasing

…s/post-processing-debiasing

…s/contextualized-bias-mitigation

epwalsh

Just a couple of comments so far.

epwalsh · 2021-05-25T18:52:19Z

allennlp/fairness/bias_mitigator_applicator.py

+    # Currently doing this manually because difficult to
+    # dynamically forward __getattribute__ due to
+    # behind-the-scenes usage of dunder attributes by torch.nn.Module
+    # and both BiasMitigatorWrapper and base_model inheriting from Model


I've seen this done before, so should be possible. See https://github.com/facebookresearch/fairscale/blob/b54eed1bd039e9ac73b82947e85113f78805c8eb/fairscale/nn/data_parallel/fully_sharded_data_parallel.py#L567-L572 for an example.

We had a discussion about this. This was my initial implementation, but the problem is that I want all function calls (not just the missing ones) to be delegated to base_model, but all non-function attributes to be the ones that belong to bias_mitigator_applicator. There is __getattribute__ (__getattr__ is only for missing attribute calls), and I tried messing around with this, but because torch.nn.Module does some behind-the-scenes things with __getattribute__, this created so many stack overflow errors (__getattribute__ kept calling itself), and I didn't feel comfortable including that in the library. I think this is why you had recommended just manually delegating calls. I explored this for way longer than I should have lol (probably 10 hours), and there's just not a good way to identify if something is a function in Python; you can only tell if something seems like a callable.

epwalsh · 2021-05-25T19:00:15Z

allennlp/fairness/evaluate_bias_mitigation.py

+@Subcommand.register("evaluate-bias-mitigation")
+class EvaluateBiasMitigation(Subcommand):


We usually define subcommands in allennlp.commands. You could just move this file over there.

I'm considering just not including this in the core library because it's too specific of a command, @AkshitaB had the great idea to keep it in the guide, so I will just remove it from this PR.

@ArjunSubramonian I think you missed this. :)

epwalsh · 2021-05-25T19:00:56Z

allennlp/data/dataset_readers/snli.py

@@ -0,0 +1,122 @@
+from typing import Dict, Optional


Was this just copied over from allennlp-models?

yes, it's used in the test, but the environment in which the test is run doesn't have allennlp-models.

It's fine to have some tests in allennlp-models for this code. For instance, we have some checklist tests there, because that's where the models are.

I added the test under allennlp-models.

We can remove this file now, right?

allennlp/fairness/bias_mitigator_applicator.py

AkshitaB · 2021-05-18T22:55:56Z

allennlp/commands/__init__.py

@@ -20,6 +20,7 @@
 from allennlp.common.plugins import import_plugins
 from allennlp.common.util import import_module_and_submodules
 from allennlp.commands.checklist import CheckList
+from allennlp.fairness.evaluate_bias_mitigation import EvaluateBiasMitigation


Are we including this?

Sorry, we are not!

AkshitaB · 2021-05-18T22:57:56Z

allennlp/data/dataset_readers/__init__.py

@@ -19,3 +19,4 @@
 from allennlp.data.dataset_readers.sequence_tagging import SequenceTaggingDatasetReader
 from allennlp.data.dataset_readers.sharded_dataset_reader import ShardedDatasetReader
 from allennlp.data.dataset_readers.text_classification_json import TextClassificationJsonReader
+from allennlp.data.dataset_readers.snli import SnliReader


This is present in allennlp-models right? Do we need to have it here?

yes, it's used in the test, but the environment in which the test is run doesn't have allennlp-models.

AkshitaB · 2021-05-18T22:59:54Z

allennlp/fairness/bias_direction_wrappers.py

+    """
+
+    def __init__(self):
+        self.direction = None


We should specify the expected types for direction and noise.

AkshitaB · 2021-05-25T20:38:00Z

allennlp/fairness/bias_mitigator_wrappers.py

+        raise NotImplementedError
+
+
+# TODO: remove equalize words from evaluation words


Is this for later, or for this PR?

This is for later, the current assumption is that the user won't have any overlap between the embeddings for which to mitigate bias and the equalize set.

AkshitaB · 2021-05-25T20:43:11Z

allennlp/fairness/bias_direction_wrappers.py

+        self,
+        seed_words_file: Union[PathLike, str],
+        tokenizer: Tokenizer,
+        direction_vocab: Vocabulary = None,


direction_vocab: Optional[Vocabulary] = None

Have done this for all instances of Vocabulary = None

AkshitaB · 2021-05-25T20:51:11Z

test_fixtures/fairness/experiment.jsonnet

+  "validation_data_path": "test_fixtures/fairness/snli_dev.jsonl",
+  "test_data_path": "test_fixtures/fairness/snli_test.jsonl",
+  "model": {
+    "type": "allennlp.fairness.bias_mitigator_applicator.BiasMitigatorApplicator", 


Does it need to be a full path?

In this case, yes, because BiasMitigatorApplicator is in allennlp.fairness and FromParams only looks at the modules imported in model's init.py. I tried placing BiasMitigatorApplicator in model's init.py, but this caused a circular dependency issue for some reason - BiasMitigatorApplicator was using some fairness module classes that hadn't been imported yet, etc.

Co-authored-by: Pete <[email protected]>

…b.com/allenai/allennlp into arjuns/contextualized-bias-mitigation

AkshitaB

Minor comments.

AkshitaB · 2021-06-02T21:20:29Z

allennlp/data/dataset_readers/snli.py

@@ -0,0 +1,122 @@
+from typing import Dict, Optional


We can remove this file now, right?

AkshitaB · 2021-06-02T21:23:13Z

allennlp/fairness/evaluate_bias_mitigation.py

+@Subcommand.register("evaluate-bias-mitigation")
+class EvaluateBiasMitigation(Subcommand):


@ArjunSubramonian I think you missed this. :)

* added linear and hard debiasers * worked on documentation * committing changes before branch switch * committing changes before switching branch * finished bias direction, linear and hard debiasers, need to write tests * finished bias direction test * Commiting changes before switching branch * finished hard and linear debiasers * finished OSCaR * bias mitigators tests and bias metrics remaining * added bias mitigator tests * added bias mitigator tests * finished tests for bias mitigation methods * fixed gpu issues * fixed gpu issues * fixed gpu issues * resolve issue with count_nonzero not being differentiable * added more references * fairness during finetuning * finished bias mitigator wrapper * added reference * updated CHANGELOG and fixed minor docs issues * move id tensors to embedding device * fixed to use predetermined bias direction * fixed minor doc errors * snli reader registration issue * fixed _pretrained from params issue * fixed device issues * evaluate bias mitigation initial commit * finished evaluate bias mitigation * handles multiline prediction files * fixed minor bugs * fixed minor bugs * improved prediction diff JSON format * forgot to resolve a conflict * Refactored evaluate bias mitigation to use NLI metric * Added SNLIPredictionsDiff class * ensured dataloader is same for bias mitigated and baseline models * finished evaluate bias mitigation * Update CHANGELOG.md * Replaced local data files with github raw content links * Update allennlp/fairness/bias_mitigator_applicator.py Co-authored-by: Pete <[email protected]> * deleted evaluate_bias_mitigation from git tracking * removed evaluate-bias-mitigation instances from rest of repo * addressed Akshita's comments * moved bias mitigator applicator test to allennlp-models * removed unnecessary files Co-authored-by: Arjun Subramonian <[email protected]> Co-authored-by: Arjun Subramonian <[email protected]> Co-authored-by: Arjun Subramonian <[email protected]> Co-authored-by: Arjun Subramonian <[email protected]> Co-authored-by: Akshita Bhagia <[email protected]> Co-authored-by: Pete <[email protected]>

Arjun Subramonian and others added 25 commits April 13, 2021 01:57

added linear and hard debiasers

79c6c33

worked on documentation

e23057c

committing changes before branch switch

fcc3d34

committing changes before switching branch

7d00910

finished bias direction, linear and hard debiasers, need to write tests

668a513

finished bias direction test

91029ef

Commiting changes before switching branch

396b245

finished hard and linear debiasers

a8c22a1

finished OSCaR

ef6a062

bias mitigators tests and bias metrics remaining

2c873cb

added bias mitigator tests

d97a526

added bias mitigator tests

8460281

finished tests for bias mitigation methods

5a76922

Merge remote-tracking branch 'origin/main' into arjuns/post-processin…

85cb107

…g-debiasing

fixed gpu issues

8e55f28

fixed gpu issues

b42b73a

fixed gpu issues

37d8e33

resolve issue with count_nonzero not being differentiable

31b1d2c

merged main into post-processing-debiasing

a1f4f2a

added more references

36cebe3

Merge branch 'main' of https://github.com/allenai/allennlp into arjun…

88c083b

…s/post-processing-debiasing

fairness during finetuning

86081ee

finished bias mitigator wrapper

ae592d8

added reference

2501b8c

updated CHANGELOG and fixed minor docs issues

f664dfb

ArjunSubramonian requested review from AkshitaB and epwalsh May 5, 2021 08:18

ArjunSubramonian self-assigned this May 5, 2021

Arjun Subramonian and others added 2 commits May 5, 2021 15:43

move id tensors to embedding device

595449d

Merge branch 'main' into arjuns/contextualized-bias-mitigation

dc4793f

Arjun Subramonian and others added 10 commits May 11, 2021 11:25

forgot to resolve a conflict

254676f

Merge branch 'main' of https://github.com/allenai/allennlp into arjun…

26d8dff

…s/contextualized-bias-mitigation

Refactored evaluate bias mitigation to use NLI metric

1ae5e99

Added SNLIPredictionsDiff class

e2cc38e

ensured dataloader is same for bias mitigated and baseline models

c34cf31

finished evaluate bias mitigation

fdb9ea7

Merge branch 'main' into arjuns/contextualized-bias-mitigation

3efffd2

Update CHANGELOG.md

c47de58

Merge branch 'main' of https://github.com/allenai/allennlp into arjun…

2b8cf09

…s/contextualized-bias-mitigation

Replaced local data files with github raw content links

33d6267

ArjunSubramonian mentioned this pull request May 23, 2021

Added binary gender bias-mitigated RoBERTa model for SNLI allenai/allennlp-models#268

Merged

epwalsh reviewed May 25, 2021

View reviewed changes

AkshitaB suggested changes May 25, 2021

View reviewed changes

AkshitaB reviewed May 25, 2021

View reviewed changes

ArjunSubramonian and others added 5 commits May 25, 2021 16:55

Update allennlp/fairness/bias_mitigator_applicator.py

ec53a05

Co-authored-by: Pete <[email protected]>

deleted evaluate_bias_mitigation from git tracking

4afb7f2

removed evaluate-bias-mitigation instances from rest of repo

21bed9d

Merge branch 'arjuns/contextualized-bias-mitigation' of https://githu…

fefcbad

…b.com/allenai/allennlp into arjuns/contextualized-bias-mitigation

addressed Akshita's comments

972ea60

ArjunSubramonian requested review from AkshitaB and epwalsh May 26, 2021 00:36

moved bias mitigator applicator test to allennlp-models

b4011cb

AkshitaB suggested changes Jun 2, 2021

View reviewed changes

AkshitaB and others added 2 commits June 2, 2021 14:24

Merge branch 'main' into arjuns/contextualized-bias-mitigation

4d7fffb

removed unnecessary files

22a5964

AkshitaB approved these changes Jun 2, 2021

View reviewed changes

Merge branch 'main' into arjuns/contextualized-bias-mitigation

bd727dd

ArjunSubramonian merged commit b92fd9a into main Jun 2, 2021

ArjunSubramonian deleted the arjuns/contextualized-bias-mitigation branch June 2, 2021 23:14

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Contextualized bias mitigation #5176

Contextualized bias mitigation #5176

ArjunSubramonian commented May 5, 2021

epwalsh left a comment

epwalsh May 25, 2021

ArjunSubramonian May 25, 2021 •

edited

Loading

epwalsh May 25, 2021

ArjunSubramonian May 25, 2021

AkshitaB Jun 2, 2021

epwalsh May 25, 2021

ArjunSubramonian May 26, 2021

AkshitaB May 26, 2021

ArjunSubramonian Jun 2, 2021

AkshitaB Jun 2, 2021

AkshitaB May 18, 2021

ArjunSubramonian May 25, 2021

AkshitaB May 18, 2021

ArjunSubramonian May 26, 2021

AkshitaB May 18, 2021

ArjunSubramonian May 26, 2021

AkshitaB May 25, 2021

ArjunSubramonian May 26, 2021

AkshitaB May 25, 2021

ArjunSubramonian May 26, 2021

AkshitaB May 25, 2021

ArjunSubramonian May 25, 2021

AkshitaB left a comment

AkshitaB Jun 2, 2021

AkshitaB Jun 2, 2021

		@Subcommand.register("evaluate-bias-mitigation")
		class EvaluateBiasMitigation(Subcommand):

		raise NotImplementedError


		# TODO: remove equalize words from evaluation words

Contextualized bias mitigation #5176

Contextualized bias mitigation #5176

Conversation

ArjunSubramonian commented May 5, 2021

epwalsh left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

ArjunSubramonian May 25, 2021 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

AkshitaB left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

ArjunSubramonian May 25, 2021 •

edited

Loading