a naive question

[chichengfengxue]

We greatly appreciate your outstanding work. I would like to ask you a naive question. I used a checkpoint with a metric of 55.7 to evaluate on a single RTX 3090 GPU, and the result was 55.5. Could this be related to my testing environment? In fact, during my reproduction process, the evaluation results on multiple GPUs were also different from those on a single GPU, even when using the same checkpoint. I would really appreciate your insights!

[Linwei-Chen]

It is very likely related to your testing environment. I'm not entirely sure, but you could try testing with 4 or 8 GPUs to see if there are any differences.

[chichengfengxue]

It is very likely related to your testing environment. I'm not entirely sure, but you could try testing with 4 or 8 GPUs to see if there are any differences.

Thank you for your prompt response. Can you briefly talk about how to draw "Fig. 11. Quantitative frequency analysis"? Do you analyze a single image or all of the images in a dataset?

[Linwei-Chen]

Thanks for your interest. You can draw it by calculating the power of each set of frequencies, or you can simply choose the diagonal frequency (refer to ICLR 2022, "How Do Vision Transformers Work?") on the entire dataset.