Cross Validation

[osivaz61]

Hi,

I have two questions:

First : Why didn't you use K Fold Cross Validation? Second : What is the reason use different learning rate for classifier? Is it for faster convergence?

I am trying to adapt CSRA to EfficientNetB3 on my multi-label dataset. Although I try various head and lambda numbers, I am getting worse results according to baseline model. What is your opinion? Is there also something different to try?

Also there is class imbalance in my dataset. Is there need to make data augmentation to prevent class imbalance? Is CSRA a method affected by data augmentation?

Thanks

[Kevinz-code]

Hi, Thanks for your question.

For your first question, can refer to ASL, Q2L for a comprehensive pipeline of multi-label classification code (which contains standard validation code similar to ours).

The ten times larger learning rate of the classifier is used for faster convergency.

As far as we're concerned, there is a dropout layer between the avgpool the the final FC in EfficientnetB3, which might cause different struture settings when CSRA is applied. Try using resnet for a stable improvement.

For imblance, refer to recent multi-label long tail classification paper, where specific loss has been added to cure performance degradation.

Best,

[osivaz61]

Hi,

I have an idea about CSRA performance. I want to know your idea also.

In voco2007 dataset there are 5011 images and 7306 objects. The multi-label objects ratio is (7306-5011)/5011 = % 45.79 In coco dataset there are 82081 images and 241035 objects. The multi-label objects ratio is (241035-82081)/82081 = % 193.64

The best performance gain was obtained for the coco dataset (for VIT-L16 (80.4 -> 86.5) %6.1).

For my dataset multi-label ratio is %9. I think the reason is why I can not get better results is lower multi-label ratio. I am right?

Thanks Orhan SIVAZ