The logprob of image captioning result of mplug is very small

Hello, I am using mplug to do image captioning task.

At inference, I use beam search to generate results, and I find that the topk_probs from beam_generator.translate_batch_scst() is very small(like 0.0014 after exp()), indicating that the confidence score of caption is very small, could you tell me if I am using the wrong output as confidence score of the caption, or why this confidence number is such small? Thx a lot!

I follow [this](https://github.com/alibaba/AliceMind/issues/78#issuecomment-1498459046) to write the code of inference.

Start with the inference code based on issue 78 and inspect beam_generator.translate_batch_scst(), especially how topk_probs is produced and interpreted. Reproduce the image-captioning run with beam search and determine whether the reported value is an intended score or a confidence value; done means the behavior and correct interpretation are clearly documented or the defect is isolated.

Written by the indexing model from the issue text.