better results by retraining with "step" learning rate policy in solver_ctc.prototxt, than with "inv"

[rohitsaluja22]

Hey, I am getting better results by retraining with "step" learning rate policy in solver_ctc.prototxt, than with "inv".. Just wondering why.. What learning rate policy you used for solver_ctc.prototxt in your experiments?

[MichalBusta]

Hi,
we have been using step policy - for big model

• no proper experiments with optimization has been done (we are low with GPU resources)

some experiments with policy you can find here: (but the task is different) https://github.com/ducha-aiki/caffenet-benchmark

[rohitsaluja22]

Oh. I asked this question because you have shared "inv" in https://github.com/MichalBusta/DeepTextSpotter/blob/master/models/solver_ctc.prototxt

[MichalBusta]

But shared "small" model has been trained with "inv" - and we did not run comparison ...