Aleksei Shabanov

One of the possible solutions is to allow elements with the same distance to the query to share the same `k-th` place. This approach is used in Hyp-ViT: https://github.com/htdt/hyp_metric/blob/be2b829b21c279ab874f113c648c0296be89134d/helpers.py#L70

Closed because of: https://github.com/OML-Team/open-metric-learning/pull/382#issuecomment-2155060075

@dapladoc changing tolerance is okay

@DaloroAT What do you think? Does bringing the additional logic worth the benefits of metrics' stability? PS. Initially, I thought that the required extra logic would be smaller and simpler...

@dapladoc let's implement what we've decided offline: 1. measure overhead 2. introduce a flag for this functionality available for developers only (so, we don't expose is to config) 3. depends...

@dapladoc hey! just a reminder :)

Let me close it for now due to inactivity. Metrics code has also been significantly modified. The order of predictions is determined in `RetrievalResults`, so, I'm not sure this problem...

from @dapladoc : > I thought about some algorithm to find quantiles online, in a batch mode. I found that there is P^2 algorithm [(link)](https://www.cse.wustl.edu/~jain/papers/ftp/psqr.pdf) and here is a discussion...

from me: > @dapladoc thank you for your thoughts and links! please, don't start working on cython implementation (except if you really want this), let's discuss it for now. I...

The graphs above seem okay to me, it's I consider fnmr@fmr only as an additional (auxiliary) metric. So, it's not super important 30% or 32% of positive distances are smaller...