About vgg19 result on high-resolution dataset

[soheejun]

Hi, thanks for your job and for sharing the code.

By the way, I made an experiment using vgg19 code in 'attack/cifar10/' on the different datasets (high-resolution dataset like caltech256, cubs200 (3, 224, 224)).
However, the results show very poor transferability.
The transferability score is 10% against the resnet50 model as the target model.

I used the original vgg19 bn model provided by torchvision.
I know the vgg19 bn model in the code is different from the original vgg19 bn model (has less fully connected layers in the classifier section of the whole model)
So, I add the linbp_relu function to the last 2 relu layer in the classifier section.
But the result had no improvement.

The hyper-parameters that I used are followed.
--epsilon 0.03 --niters 300 --ila_niters 100 --method linbp_ifgsm --save_dir data/cubs200/linbp_ifgsm --batch_size 32

I expect that the vgg19 will show good transferability with the high-resolution dataset just like resnet50.
Can you tell me is there anything I missed or misunderstood??

Again, I really appreciate your job!