Better scores are acieved than the one in your paper

[masashi-y]

Hello, thank you for sharing the great codebase.
I cloned the codebase and ran all.py to reproduce the scores in your paper.
Yes, I successfully observe your hyperbolic cone model acieves:

 ======> best OVERALL  f1 CONES test = 92.80; CONES valid = 92.60 - after 260 epochs.

but after that, it shows significantly improved score:

 ======> best OVERALL  f1 CONES test = 96.60; CONES valid = 96.60 - after 360 epochs.

It is a bit confusing. Why didn't you report this score in the paper?
and which one should I compare my method against?

[octavian-ganea]

Thanks for sharing your experimental findings. We stated in the "Training details" of our papers the number of epochs used to train all models. It is indeed possible that some of these models benefit from longer training. My suggestion is to train each method until convergence and no improvement is observed.

[masashi-y]

Thank you for the very prompt reply.
Now I understand that was because the score was achieved outside 100 (initialization) + 200 (cone training) epochs.
On top of that, I realized the score was in an experiment with 90% of transitive closure instances.
I think I need to be careful on these points..
Thank you for the suggestion as well!

[masashi-y]

Sorry for the repeated questions. I tried to reproduce the results in the case where the embedding size = 10, which resulted in failure.

The best obtained score is using 90% of transitive closure instances with:

 ======> best OVERALL  f1 CONES test = 89.50; CONES valid = 84.80 - after 360 epochs.

I changed the following dims to 10 (I believe the former is not relevant anyway). Is there any other place I need to change in order to reproduce the scores?
https://github.com/dalab/hyperbolic_cones/blob/master/params.py#L15
https://github.com/dalab/hyperbolic_cones/blob/master/params.py#L51