Training fails to converge

Question

Training fails to converge

Closed this issue 3 years ago · 2 comments

I use the dafault data but only change the cores used. However, the Training floats at 60%. And when i only change the algorithm to Adam, it even get worse. the training stays at 12% and cannot improve.

Answer 1 · 2022-03-25T20:40:48.000Z

For Adam, you might need to reduce the learning rate to like 0.1, 0.05. The default data (a-Ag RRAM) has pretty bad non-linearity, it might not converge very well.

Answer 2 · 2022-03-26T04:46:07.000Z

thanks. I see. At 2022-03-26 04:40:58, "NeuroSim" ***@***.***> wrote: For Adam, you might need to reduce the learning rate to like 0.1, 0.05. The default data (a-Ag RRAM) has pretty bad non-linearity, it might not converge very well. — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: ***@***.***>