Reproduce Issue: Up-A CIFAR-100, Flowers 102

[dev-sungman]

Hi,

When I reproduce the linear-probe classification performance on CIFAR-100, Flowers 102,
I got weird results when Up-A R50 was used for the backbone.

In the paper, Up-A R50 outperforms ImageNet pre-trained R50, however, ImageNet pre-trained R50 outperforms Up-A R50 on CIFAR-100, Flowers 102 cases with large margins.

My configurations as below:
Model

Dataset

Others

I checked whether the pre-trained Up-A was successfully loaded or not.

Thanks in advance.

[dev-sungman]

If you could provide the linear-probe configuration that uses Up-A, it must be helpful to me.

[yinanhe]

In linear-probe setting, we select the learning rate and weight decay with a grid search of 3 logarithmically spaced, in which between 1e-4 and 0.1 for learning rate, between 1e-6 and 1e-4 for weight decay, we also searched two momentum 0.9 and 0.99. Besides that, 30 is also a work learning rate here.

[dev-sungman]

Thanks very much !