Open baiyuting opened 1 year ago
when I tried to use the same model to generate text through beamsearch and argmax, I found beamsearch get worse result. I use the generate_beam() and generate2() in clip_prefix_captioning_inference.ipynb to generate text. Is this a normal thing?
when I tried to use the same model to generate text through beamsearch and argmax, I found beamsearch get worse result. I use the generate_beam() and generate2() in clip_prefix_captioning_inference.ipynb to generate text. Is this a normal thing?