Reranking consumes too much memory space

[FatimaZulfiqar]

I want to apply reranking for the MSMT17 dataset but the algorithm consumes too much memory space. Is there a way to consume less memory by achieving similar results? I have limited resources available that why cannot afford to increase the specs. Any insight, solution, and suggestion to this problem will be appreciated.

[layumi]

Hi @FatimaZulfiqar You may try the half-precision data type float16。

[FatimaZulfiqar]

Okay thank you will try that but will the results be the same if I replace float32 with float16? Or will there be any effect on the result?

[layumi]

@FatimaZulfiqar
The result will be decrease by ~1% according to our experience.