freq_mask_path and space_mask_path

[llllly26]

Hi,
I want to know that, If I want to use freq_mask and space_mask on my own datasets, how can I generate it?
Thanks!

[AwakerMhy]

You can calculate the average value over the dataset for each pixel in each image to generate the space_mask, and calculate the average amptitude over the dataset of each frequency component in each image after Fourier transform for the freq_mask, as mentioned in Sec 4.3 of our paper. Emprically, we find that about 200 images randomly sampled from the dataset is enough for calculating both masks.

[llllly26]

You can calculate the average value over the dataset for each pixel in each image to generate the space_mask, and calculate the average amptitude over the dataset of each frequency component in each image after Fourier transform for the freq_mask, as mentioned in Sec 4.3 of our paper. Emprically, we find that about 200 images randomly sampled from the dataset is enough for calculating both masks.

thanks!
By the way, could you release code about this? These parameters are necessary to generate high quality image, so I hope to keep align with your work.

[AwakerMhy]

Sure, I will add a demo about how to calculate the masks soon.

[AwakerMhy]

We have added the file mask_gen_demo.py.py, which is a demo showing how to generate frequency and space mask for an image dataset. We take CIFAR10 as an example. We hope you find it helpful.

[llllly26]

We have added the file mask_gen_demo.py.py, which is a demo showing how to generate frequency and space mask for an image dataset. We take CIFAR10 as an example. We hope you find it helpful.

thanks!

[llllly26]

We have added the file mask_gen_demo.py.py, which is a demo showing how to generate frequency and space mask for an image dataset. We take CIFAR10 as an example. We hope you find it helpful.

Sorry @AwakerMhy , another confusion that, if different categories of images are different scale, need to [padding] and [mask] to same shape, but without using torchvision.transforms to Resize, in this scenario, does it make sense to calculate average?

[AwakerMhy]

For other datasets in which the raw images has different sizes, you need to resize them to have the same size and before calculating the masks.