About adversarial learning

[alceubissoto]

First of all, congratulations on your paper! I thoroughly enjoyed reading it, and I found the ideas presented to be very interesting.

I’ve been working on reproducing your results, and I have a question regarding the training process. Based on my understanding, following the instructions to run ILI-Pred4-0.py and ILI-Pred4-1.py sequentially will:

1. Perform pre-training on the feature extractor (5 epochs),
2. Separate the environments (5 epochs), and
3. Run the invariant learning model (10 epochs).

From reading the paper, I was under the impression that steps 2 and 3 would be performed in an adversarial manner, with the environments being re-learned as the feature extractor improves, continuing in this manner until convergence. Was there a loop around these two steps in the paper's experiments?

I might have misunderstanded the paper, or overlooked something in the code. Would you please help me clarify this?

Thank you!