search

THUDM / ImageReward

[NeurIPS 2023] ImageReward: Learning and Evaluating Human Preferences for Text-to-image Generation
Apache License 2.0
1.12k stars 62 forks source link

Can I use ReFL method to train another dataset for a new reward model? #96

Open LOGIC-10 opened 2 weeks ago

LOGIC-10 commented 2 weeks ago

Hi, thanks for your fantastic work! I am trying to understand how to use the ReFL method to train a new reward model with a custom dataset. Specifically, I am unsure how to structure my dataset and integrate it into the current codebase. The current code reads image files in a way that is not immediately clear to me.

The current code reads image files in a way that seems to rely on specific identifiers rather than absolute file paths. This makes it unclear how to apply my own dataset. Thanks if you can give a guidance on how to ensure my dataset is correctly loaded and processed for training.

Thanks a lot!

xujz18 commented 1 day ago
image

You can refer to this section and notice that there is a make dataset function to process data.