DIGITS docker - googlenet training accuracy

[saracloud]

Hi,

Trying to train set of images for classification using googlenet. The DBs and training is done at GPU cloud where DIGITS docker runs without any issue.

Ensured that the image size meets googlenet network default image size requirement. Minimum 30 epochs are done, still i see that the accuracy is less.

Kindly advise, what is way to improve this. Do i need to modify predefined network to improve accuracy.

I tried my level best to find some info at your developer portal, didn't really find the appropriate one, kindly let me know if this issue is already addressed.

Thanks,