whatr does the reward mode score mean?

[wh0x]

The score of the reward model is obtained directly from the final linear layer without undergoing normalization, which feels somewhat peculiar. What is the value range of the reward model's score, and what does it represent?

[Kenkenzaii]

The reward model is a regression network that outputs a continuous score for inpainting quality. We normalize the score using a fixed mean and standard deviation to ensure consistency and stable training. See here https://github.com/Kenkenzaii/PrefPaint/blob/6001d6cbda003a0d3fff23f7723c204ece3521e6/InpaintReward.py#L100.

[Kenkenzaii]

The score of the reward model is obtained directly from the final linear layer without undergoing normalization, which feels somewhat peculiar. What is the value range of the reward model's score, and what does it represent?

The config.group setting is now correctly set to True.

[wh0x]

Thank you for your reply. I would like to know what the loss reduction curve looks like during training, for example, from what value to what value it decreases. Recently, I've been trying to apply this method to an object removal task, but the loss curve is quite unusual, sometimes even showing negative values. I wonder if this is an issue with my code or some other reason. Also, do you think a reward model would be helpful for the removal task?