adversarial debiasing (Updated)

[giandos200]

Following the paper "Mitigating Unwanted Biases with adversarial learning", i've added the pretraining of both the Network (Classifier and Adversarial one). I've also tried to update the adversary_loss_weight as sqrt(epoch) but the result same to be not effective, so i left constant the value of 0.1. However, with the Pretraining methodologies with all the dataset (German,Compas, Adult) the Fairness Metric have generally a relevant improvement (DI from 0.58 to 0.96 in adult, german['sex'] from 1.34 to 1, compass['sex'] from 0.75 to 0.76). Of course for each dataset should be done a Hyperparameter tuning for achieving best result. Furthermore, the code has been migrated to TF2. resolve #276

[lgtm-com[bot]]

This pull request introduces 1 alert when merging 00eacb01091136d6dfd17803615a5645485c27e8 into a5dac73786265415138ed6e2f8c926d9ee968ee5 - view on LGTM.com

new alerts:

• 1 for Unused import

[lgtm-com[bot]]

This pull request introduces 4 alerts when merging 1999fa344eaeedb337945acbbd52bb4102798752 into 963df2e4eea807bd5765fee9f1c594500bdcbb5b - view on LGTM.com

new alerts:

• 2 for Unused local variable
• 2 for Unused import