How to prepare pytorch model weight?

[RorschachChen]

I tried to train ResNet18 from torchvision on CIFAR10 from scratch and with a 90+% accuracy, but when I evaluate the fourier heatmap of my model weight, I found that the result is not like the example. Do you have any idea about the reason or how can I get the pretrained model weight properly? Thanks for replying.