Error OOM memory exhausted using lev_marq
fikri2992 opened this issue · comments
when using levemberg marquardt optimizer error OOM memory exhausted with shape[13537203,13537203]
but not with momentum with same shape
input is 45120 size with 38 dataset
and sigmoid 300
softmax 3
and then OOM and need fully restart on python runtime
it's expected, since Levenberg-Marquardt constructs NxN matrix, where N is number of parameters
(45120 * 300 + 300) + (300 * 3 + 3) == 13537203
It's better to use LM algorithm for smaller samples and networks
wow thanks, I did not notice that I have this NVIDIA k80 to run this and still get an error
any advice to reduce the networks if I'm still using this peripheral and input size?
you might try to reduce dimensionality of the input, maybe using PCA or SVD, if you manage to reduce it from 45120 to maybe 200 then everything should work fine.
awesome thanks, this case solved will try that PCA or SVD