Please provide noisy label files for Office 31, Office Home

[devraj89]

Hi

Thanks for the wonderful code. Can you please provide the following (1) the noisy label files for Office 31 and Office Home dataset. In the code, only the noisy label file for the Amazon dataset has been provided.

(2) What is the noisy model used to corrupt the clean labels of the source data? Is it symmetric or pairflip? Can you please provide the code to generate the noise from our end?

(3) Can you please provide some information regarding how the feature corruption is being done? From the paper, it is written that "For feature corruption, each image is corrupted by Gaussian blur and salt-and-pepper noise with probability p_noise." -- are you corrupting the whole image or in patches ? is hale of the images corrupted with gaussian blur and the other half with salt-and-pepper? or are both the noise models applied sequentially on top of each other? What is the value of sigma in the case of gaussian blur or the value of the threshold for the salt-and-pepper noise?

I will be grateful if the code to generate the image corruption is provided.

Thanks Devraj

[jingzhengli]

I meet the same issue as you. Do you solve this issue?