S2P2 does not compute acc and rmse metrics during training/evaluation

[NansyJourney]

Hello,

I am using EasyTPP to train and evaluate the S2P2 model. To verify this is not a dataset-specific issue, I tested on your provided Taxi dataset (from the EasyTPP benchmark) using the example configuration from examples/configs/experiment_config.yaml.

When I add metrics: ['acc', 'rmse'] to the S2P2 training config, the model trains successfully but no accuracy or RMSE values are reported in the logs — only loglike is shown. This behavior is consistent across both the Taxi dataset and my custom dataset.

For comparison, RMTPP, NHP, and THP all correctly report acc and rmse during training and evaluation.

My questions:

1. Is it possible to compute acc and rmse for S2P2 using the thinning algorithm, similar to other models?
2. If yes, what is the correct configuration or code change needed to enable this?
3. If not, is this a known limitation of the current S2P2 implementation in EasyTPP, and are there plans to support it?

Thank you!

[iLampard]

Sorry for later reply. We will let the author to reply you.

[yuxinc17]

Thanks for your interest in S2P2. While our model is integrated into the core EasyTPP library for general use and as a reference implementation, the specific configurations are explained in Appendix C.1, C.3, and D.1. The code modifications used to reproduce our paper's results are hosted in our forked repository (as mentioned in Footnote 3 on page 8 of our paper).

As noted in our README and Appendix D.1, we chose to depart from the default EasyTPP thinning-based evaluation to produce a faster RMSE estimator for the next immediate event with much lower variance.

We recommend checking our paper for the full context. Hope this helps!