Metrics for GKE - Githubissues

arreason / CategoryLearning

Where Category Theory meets Machine Learning

1 stars 0 forks source link

Metrics for GKE #35

Open Kuraao opened 3 years ago

Kuraao commented 3 years ago

add functions to compute usual metrics for GKE. The overall idea is:

We want to find the best candidates to complete a triplet source-label>target. We might be missing any of the three items (source, label or target).
After the matching operation, rank the available candidates according to their scores.
Compute the metric you want.

metrics which are often used are:

MR (mean rank)
MRR (mean reciprocal rank, i.e the mean of 1/rank)
top@k (proportion of ranks not larger than k)

This will also be added to logging operations (using wandb)