Closed matanhalevy closed 2 years ago
Hi is this repo actively maintained/will this issue be assigned?
Hey Matan,
Sorry for the delay on this. I've looked into this a bit and it seems like it might be a limitation on Huggingface. If I'm not mistaken, Huggingface doesn't allow abstract classes for whatever reason. The specific ValueError that you're hitting was added via #8602 and there is a comment on that line.
I'm happy to help dig into this more if you're able to help me reproduce the error.
Closing this comment out given the original ValueError in this issue was removed from huggingface.
Hi I've been trying to debug this for a few days, I'm using the MinDiff remediation similarly to how it's used in the tutorial. Using the debugger I see the 'mask' value is None from the caller. The base model I am using to remediate unfairness is
TFBertForSequenceClassification
from HuggingFace's transformers library.Relevant Versions: python 3.6.9 tensorflow 2.3.1 tensorflow-model-remediation 0.1.3 transformers 4.3.2
Stacktrace:
Code to reproduce (I can't share my data, but I've tried loading the BERT weights from a TF trained Transformers with the same issue. I've also commented out wrapping the TFBertForSequenceClassification in tf.keras model but it does not make a difference to the stacktrace.
Thank you!