Pre-trained googlenet performance dropped when the input image was normalized with imagenet statistics

vision
1

Hi,
I have used the GoogleNet pre-trained model and normalized the input image by the Imagenet statistics, but significantly the performance dropped. But when I used the input tensor in the range of [0,1], I saw that the model worked better. Did anyone have any experience with this issue?

1 Like
2

What kind of data are you using at the moment and what accuracies do you get using both approaches?

3

Thank you very much for your quick reply. Let me explain my task for you. I worked on Imagenet validation dataset to train an adversarial perturbation generator in which a freezed googlenet model was used.

I saw this line of code and I inferred that when we set pretrained=True in Googlenet constructor, the forward function normalized the input image differently (seems it works with different statistics).

About the accuracy, I did not directly calculate it but in terms of fooling rate in adversarial attack task which I achieved, I saw that the input image with the range [0,1] had far better performance.