one question about kernel 5*5

[alexliyang]

when train cifar10, initial size is 9696 , and when downscale size is less than kernel 55 ,how to deal these channels?

[martinkersner]

Hi @alexliyang!

Thanks for a good question! The input size 96x96 is just an example. You should use larger input for MobileNetV3 architecture if you want to get correct/valid results. I will update repository to reflect this constraint.

Cheers, Martin

[alexliyang]

and I meet another question, I want to use this model to classify 7 classes, when training acc is about
0.9, I save model and test , the reuslt is very bad, always tell me the class id is 2 , no other class ids

[martinkersner]

Do you evaluate your model at the end of every epoch? What is the validation accuracy?