How to conduct batch inference

[stdKonjac]

Hi, thanks for the great work!
I wonder if this repo supports batch inference so that I can process multiple images at once?

[Luciennnnnnn]

Sure, we’ve exported batch_evaluate, which is designed to process multiple images at once. I recommend referring to our reward server implementation, which leverages the batch inference capability of EditScore.

[stdKonjac]

Sure, we’ve exported batch_evaluate, which is designed to process multiple images at once. I recommend referring to our reward server implementation, which leverages the batch inference capability of EditScore.

Thanks for the quick response!
One more small issue: I see batch_evaluate in the EditScore class, but the PQ_prompt setup differs from the evaluate function:

# single evaluate
PQ_prompt_final = self.model.prepare_input(image_prompts[-1], self.PQ_prompt)
# batch evaluate
PQ_prompt = [self.model.prepare_input(image_prompt, self.PQ_prompt) for image_prompt in image_prompts]

Just want to double-check if all implementations are correct? (Based on my understanding, the second one should also use image_prompt[-1] to fetch the edited image)

[Luciennnnnnn]

You’re right — it’s a bug. The expected behavior should be image_prompt[-1]. The current results were produced using the incorrect code. I’ll rerun the experiment to see how this bug affects RL performance and share the results with you once they’re ready. Thanks!

[stdKonjac]

Sure. Feel free to close this issue once this bug is resolved 👍