Jiazheng Xu comments

Results 58 comments of


                                            Jiazheng Xu

How to use HuggingFace Data?

Yes. Like this: `load_dataset("THUDM/ImageRewardDB", "8k_group")`

test custom model acc : 0.0%

Discussion is very welcome and I hope I can help. I noticed that your test_test1.json has exactly the same rewards, and it's reasonable to assume that it's probably the same...

dependency_links are ignored; manual installation of CLIP is requried

Thanks a lot! We have updated it in `setup.py`.

关于纯图像的得分评估

Thanks a lot to @hkunzhe for sharing and explaining!

有没有可能把ImageReward用在生图过程中指导降噪？

当然可以尝试。事实上，ImageReward可以像classifier guidance那样指导生成。 > Sure try. In fact, ImageReward can guide generation like classifier guidance.

About the score range

When you test with ImageReward, the higher ImageReward, the better. The reward is normalized to have a mean of 0 and a standard deviation of 1. When testing stable diffusion...

About the reward score

You can refer to Section 2.2 RM Training for details.

Training ImageReward model on different budgets

Thanks for your discussion! Firstly, the hyperparameters are for 8k budget (but the shuffle may differ, so it's worth trying a little bit different ones). Secondly, see https://github.com/THUDM/ImageReward/blob/main/train/src/ImageReward.py#L87-L99.

Training ImageReward model on different budgets

The 8k hyper-parameters should only need smaller adjustments to accommodate 1k/2k/4k.

训练问题

可以减少Transformer参数的可训练部分。 > The trainable parameters of Transformer Blocks in ImageReward can be reduced.