Hi Sam,
I'm Srinivasa Pranav, a PhD student at CMU. We were interesting in recreating your experiments and comparing our method to HNNs.
In analyze-spring.ipynb, we had a hard time understanding why time is rescaled in the numerical integration using RK45 (from scipy's solve_ivp). This is in the integrate_models function under the "Quantitative Analysis" section:
# rescale time to compensate for noise effects, as described in appendix
t_span[1] *= 1 + .9*noise_std
t_eval *= 1 + .9*noise_std
This step has a significant impact on the results. We were unable to find information in the paper's appendix, so we hope you could guide us to a reference.
Also, since other experiments/tasks don't rescale time, we were interested to know why Task 1 requires it while other tasks do not.
Besides this, we really appreciate your efforts in neatly organizing your code and making it easy to understand.
Thank you!
Pranav
Hi Sam,
I'm Srinivasa Pranav, a PhD student at CMU. We were interesting in recreating your experiments and comparing our method to HNNs.
In
analyze-spring.ipynb, we had a hard time understanding why time is rescaled in the numerical integration using RK45 (from scipy'ssolve_ivp). This is in theintegrate_modelsfunction under the "Quantitative Analysis" section:This step has a significant impact on the results. We were unable to find information in the paper's appendix, so we hope you could guide us to a reference.
Also, since other experiments/tasks don't rescale time, we were interested to know why Task 1 requires it while other tasks do not.
Besides this, we really appreciate your efforts in neatly organizing your code and making it easy to understand.
Thank you!
Pranav