ChenyangSi

@jsjxyjl 谢谢关注，开源code需要一系列的流程，我们会尽快开源。

@lllller thanks for your interest. We are cleaning the code. And the code needs to be reviewed by our engineering team. After that, the code can be released.

@cxq1 thanks! relative log amplitudes are plotted with https://colab.research.google.com/github/xxxnell/how-do-vits-work/blob/transformer/fourier_analysis.ipynb. Fourier spectrum of feature maps is visualized with the function np.fft.fft2()

@cxq1 did you mean the visualization of Grad-CAM activation maps？

@lijain We have released the code.

Code has been released.

Code has been released.

Code has been released.

@hzlbbfrog yes, first do FFT with fft2(), then do a channel-wise average pooling

@cxq1 Grad-CAM: https://github.com/sail-sg/poolformer/blob/main/misc/cam_image.py yes, "Avepool" and "upsample" can be regarded as low-frequency, but we use these two operations mainly to reduce the computational cost.