sharpness values are always bigger than the value reported in the paper

[Shiweiliuiiiiiii]

Hi,

I am trying to reproduce the sharpness with F1 model on MNIST. However, the sharpness values of BS=6000 are more than 200 or even 300 with different random seeds, which are larger than the value reported in the paper (57). I am wondering if this is due to the random seed I picked?

Many thanks.

[Billy1900]

Hi,

I am trying to reproduce the sharpness with F1 model on MNIST. However, the sharpness values of BS=6000 are more than 200 or even 300 with different random seeds, which are larger than the value reported in the paper (57). I am wondering if this is due to the random seed I picked?

Many thanks.

hi, it seems like you compute sharpness successfully, but I met some problems when I run the code.

ABNORMAL_TERMINATION_IN_LNSRCH

 Line search cannot locate an adequate point after 20 function
  and gradient evaluations.  Previous x, f and g restored.
 Possible causes: 1 error in function or gradient evaluation;
                  2 rounding error dominate computation.
std of sharpness:  0.0 mean of sharpness:  0.0

I read some blogs: https://stackoverflow.com/questions/34663539/scipy-optimize-fmin-l-bfgs-b-returns-abnormal-termination-in-lnsrch, it said gradient does not match with objective function. I am confused about that, could you please help me with that?