shuiyigt

Results 2 issues of shuiyigt

批量送入不同size的inputs,无论是embedding还是reranker,最后平均到每条数据的耗时都几乎一样,这是为啥? 例如inputs分别输入20、40、60,总耗时正比例递增,最后每条数据的耗时相似。 按理说不是batch越大,矩阵运算会节约一点时间,大batch的单条平均耗时会更小吗?有点不明白,望解惑。

Paper shows several steps about enhance perception, containing element descption, dense captioning, state transition captioning, SoM, QA, grounding. I wonder if the FIRST THREE steps are only trained for adjusting...