Paper implementation

[ehfo0888]

Hello,

Thank you very much for sharing your code.

when I studied the paper I came across 2 implementation detail. I see that the setting you choose are the same as the hyper parameter for the inpainting?

is there an option in the code for regular training the one with 1024 batch size? what is the difference between option 1 and 2 in the screen above?

[Janspiry]

Hi, although the implementation uses the same learning rate, the batch_size is far from 1000, so maybe it would be better to implement it conditionally (you can adjust the batch_size option in the configuration). The second question I did not understand too well ?

[ehfo0888]

In the training details it mentioned the number of training step is 1M but in the Diffusion Hyper Parameter section it says only 2000 training. I am confused, what are the differences?

[Janspiry]

2000 is time-step of diffusion \alpha, not the training step

[ehfo0888]

Thank you for your time.

1 - how can I increase the number of training steps in the code? 2- what is the mask you used in the pre-trained model? is it hybrid? I run the colab and changed the type of mask during the testing to free-form but the model couldn't fill the edges? how do you think I should change the setting to cover this issue?

[Janspiry]

1. Change the n_iter: https://github.com/Janspiry/Palette-Image-to-Image-Diffusion-Models/blob/136b29f58d0af6e5db9f3655d2891f5a855fcdaa/config/inpainting_celebahq.json#L137
2. hybrid, the unmasked region have changed, Is there any change somewhere else?

[ehfo0888]

The number of iterations when running the training script is way less than that with the current setting which is 9332.

The only change I did is to change the mask during the test in the path file from center to free form and changed the "n_step" In the test from 1000 to 3000. It seems even though I change the mask type it still uses a centering mask!!

What is the correct way of setting a mask for such a picture? and what type of mask do you recommend?

[Janspiry]

Did you change these in notebook?

[Janspiry]

Yes, you can find it in https://github.com/Janspiry/Palette-Image-to-Image-Diffusion-Models/blob/136b29f58d0af6e5db9f3655d2891f5a855fcdaa/data/util/mask.py#L232, and you need to give the mask manually. Get more images by https://github.com/Janspiry/Palette-Image-to-Image-Diffusion-Models/blob/136b29f58d0af6e5db9f3655d2891f5a855fcdaa/config/inpainting_celebahq.json#L71