how can i change backbone from resnet to efficientnet

[7eta]

How to convert resnet to efficientnet on Mask R-CNN backbone?

mask r cnn is too slow.. so i hope to change the backbone. but face the wall... help me..

[FrancescoMandru]

From your explanation it's hard to give you a complete answer. You can simply replace the backbone on your code with EfficientNet adjsuting the final shape of your tensor equal to the one provided by Resnet if you don't want to change also the rest of the code.

[VeeranjaneyuluToka]

I modified the Mask R-CNN ResNet101 backbone to EfficientNet. But accuracy is not as good as Mask R-CNN resnet101 with coco weights. Anybody is facing the same issue?

[atamazian]

This issue is not related to efficientnet package itself, so I think this issue should be closed.

[Rutvik552k]

I modified the Mask R-CNN ResNet101 backbone to EfficientNet. But accuracy is not as good as Mask R-CNN resnet101 with coco weights. Anybody is facing the same issue?

can you provide that code

[VeeranjaneyuluToka]

Hi @Rutvik552k ,

Sorry, i do not think i can provide implementation here and i have implemented that as part of my assignment in an organization.

[VeeranjaneyuluToka]

Actually i have done some R&D on this and found the reason as well. The accuracy was low because my efficientnet backbone coco weights are not generated in correct way. So i have tuned some hyper parameters, retrained the coco weights using efficientnet backbone, and used that frozen model in my transferring learning which worked fine later. I am sorry that i can not state here all the experiments that went through.