Open Seunghyeon-Cho opened 2 years ago
This finalized version is slightly different from the version used in paper, which perhaps is the reason that produces different results. I will look into the difference and see what causes the issue.
Thank you.
According to the paper, HBNODE should be better than NODE in loss value, but in the opposite case (NODE is better than HBNODE), so I left an issue.
Let me know if you find anything!
I have one more question.
In the Walker2D kinematic simulation experiment, is HBNODE taking longer training time than NODE a normal result?
If there is a checkpoint for the experiments, I wonder if you can share it. It seems that the model is heavily influenced by external influences such as random seeds.
for 5.3 i think the problem is the preshrink parameter is not the same for different tasks. I think in our experiments they were 0.1, but in this final version i forgot to include that. In walker 2d, we did 3 tries and all of them HBNODE is better in accuracy and taking longer time as well. Considering the limited resources i have right now, it'll take quite a long time to run that experiment again. I will try to find if i have the checkpoints, but it's almost a year and I am not sure whether I could find them.
Hello.
It's been a long time, but I'm asking again. I wonder if the results of one experiment in 5.3 solved the problem that cannot be restored.
Thank you.
Hi, Thanks a lot for the brilliant repo!
I ran the experiment in 5.3 experiments (Learning dynamical systems from irregularly-sampled time series Experiments) and the results are strange. I ran both NODE and HBNODE. According to the paper, the loss of HBNODE should be lower, but the train / val / test loss of NODE is lower.
Is it possible to check what's wrong with the code?