Att mask

[glkuzi]

Small bugfixes:

• added attention mask to the BERT input
• changed the number of frozen layers in BERT based models from 10 to 12
• added mapping to ints for the gender label in the BIOS dataset
• added missed dependencies
• changed best model selection from val loss criterion to val accuracy

[HanXudong]

Hi Gleb,

Many thanks for your comments again!

Following our previous discussions, I've made the following changes:

• update the Python version from 3.7 -> 3.8, and the NumPy version from 1.21 to 1.22 for security reasons. See https://github.com/HanXudong/fairlib/security/dependabot/2 for details.
• include the attention mask as input to BERT-based models
• replace tuples with dictionaries for each minibatch iteration
• update the classifiers and discriminators (for adversarial learning) to support the additional inputs
• include the label mapping for BIOS dataset

I will then close the pull request.

Best, Xudong