search

DingXiaoH / RepLKNet-pytorch

Scaling Up Your Kernels to 31x31: Revisiting Large Kernel Design in CNNs (CVPR 2022)
MIT License
867 stars 86 forks source link

Accuracy of EMA is even lower than the trained model #24

Closed Shiweiliuiiiiiii closed 2 years ago

Shiweiliuiiiiiii commented 2 years ago

Hi Authors,

Thanks again for your excellent work. I am using your structure and running some experiments, but found that the accuracy of the EMA model is 1% lower than the original trained model. Have you noticed this problem? Or I have missed some important implementation about this.

Thanks, Shiwei

TheGreatTreatsby commented 2 years ago

I wonder the input size when you train,and how much pictures in your dataset?

DingXiaoH commented 2 years ago

When you train a model from scratch , the accuracy of EMA may be even lower than the trained model, which is expected. EMA usually works while finetuning a model.