Closed vaguenebula closed 1 year ago
Just realized
im still getting blue images, so it seems that applying the permutation in this manner gives a different result
@vaguenebula I also apply the permutation after the weight matching. The code after final_permutation in my code is basically manually applying the apply permutation function. I used the function before nbut changed it manually to see of there was anything up with the function itself. What exactly are you saving in torch.save ? updated_params ?
@vaguenebula I also apply the permutation after the weight matching. The code after final_permutation in my code is basically manually applying the apply permutation function. I used the function before nbut changed it manually to see of there was anything up with the function itself. What exactly are you saving in torch.save ? updated_params ?
For some reason, its giving me a good result. i will submit pull request so you can experiment with the code
@vaguenebula I also apply the permutation after the weight matching. The code after final_permutation in my code is basically manually applying the apply permutation function. I used the function before nbut changed it manually to see of there was anything up with the function itself. What exactly are you saving in torch.save ? updated_params ?
i fixed the blue image now. .its a mix of two models
@vaguenebula I also apply the permutation after the weight matching. The code after final_permutation in my code is basically manually applying the apply permutation function. I used the function before nbut changed it manually to see of there was anything up with the function itself. What exactly are you saving in torch.save ? updated_params ?
it might just be the wrong implementation of the method somehow. im not sure
wait whhhatt im getting model a back now
I thought i would make a separate issue for this to keep things organized. So i figured some stuff out. In the original code, the authors apply the permutation after making the final permutation. Here is what I changed in my fork:![image](https://user-images.githubusercontent.com/65990273/203147382-3ee75453-2f6b-4e06-886c-d1ee0b0c9a2f.png)
This does give a different model thats not biased. However, the resultant image is blue. I have no idea why.It turns out that I was normalizing the w_a and w_b, which was causing it to be blue