the AB-TTA strategy #16
Description
Hi authors,
Thanks for your great work! I found the DGTR paper very inspiring and I am currently trying to reproduce the results using your code.
I have a question regarding the Adversarial-Balanced Test-Time Adaptation (AB-TTA) strategy mentioned in Section 3.4 of the paper. My understanding is that AB-TTA is designed to refine grasp poses during the inference phase, which likely involves back-propagation to optimize the hand parameters.
However, while going through test.py, I noticed that the inference seems to perform a single forward pass under @torch.no_grad(), and I wasn't able to locate the iterative optimization loop for TTA.
Could you kindly guide me on where the AB-TTA logic is implemented in the current repository? If this part of the code hasn't been released yet, would it be possible to update it or provide some guidance on how to implement it within the current framework?
Thanks for your time and help!