lixilinx / psgd_tf Goto Github PK

Hi thank you for sharing this great work ; I had a question can the PSGD perform better then Adam algorithm when used for 3D object detection using CNN ?
Thank you

Hi,
thank you for this rigorous repo, the paper was delighting as well.
Today, first-order derivative optimizers are still prevalent in the landscape of automatic differentation frameworks (except for l-bfgs in pytorch, which is missing in tensorflow).
In my research, I happen to be bound on small scale recurrent neural networks (with less than 1000 parameters), and the usual Adam optimizer is not converging satisfactorily on my data exhibiting very long sequences.
This PSGD seems to be promising, yet I have my full framework running on TF2.3.
Is there going to be a TF2 version of the PSGD, and if not, is there any chance to get hints how to tackle this?
Thanks in advance and best regards

PSGD works great for the mentioned examples for it converges in lesser number of iterations as well as lesser time. But for models with large number of parameters, one iteration of PSGD would take a lot more time than that of first order methods i.e. SGD, RMSprop and ADAM. Is there a solution for models with large number of parameters?