Training Accuracy problem

[leochli]

Hi @farmingyard ,

Have you trained with your prototxt and do you reach the paper acc?

I followed the resnext training policy yet have 1%~2% difference below the paper acc.

I'd be appreciated if there're any tricks to share.

thx!

[farmingyard]

@LeoLee96
My prototxt is not the original paper structure but its 1by2 of each feature maps, so we don't know its perfect acc. WHAT is your current acc?

[leochli]

Hi @farmingyard ，
I haven't tried your prototxt yet. I trained with the original paper structure and get top1 70.7% now. I'll try ur 0.5x version asap

[farmingyard]

@LeoLee96
You can reference to shicai's result, which is more or less with yours.
https://github.com/shicai/MobileNet-Caffe

[borisgribkov]

Does 1by2 mean 0.5 MobileNet in the original paper? It seems that yes according to the number of kernels. I trained a network using your prototxt, final accuracy is close to 63%. (63.7% was published for 0.5 network)

[Ai-is-light]

Why the BN layer of your deploy.prototxt does not set "use_global_stats: true"
Thanks @farmingyard

[borisgribkov]

Finally I have increased number of kernels twice, this should be original MobileNet, final accuracy 69.5%, getting of over 70% accuracy is still open question.