Hi @kpei, did you make any progress on this? It recently come to our attention that using pm.dist_math.switch might break the gradient. You can see this post for a case study, and here for some recommendation/workaround.
junpenglao
17
Related topics
| Topic | Replies | Views | Activity | |
|---|---|---|---|---|
| Switch point Metropolis tuning | 25 | 5095 | May 19, 2018 | |
| Custom sampling step | 36 | 2765 | November 8, 2022 | |
| Sampling time is very long (minutes per sample) | 13 | 4674 | November 2, 2017 | |
| Geometric variable initialization | 1 | 468 | April 14, 2018 | |
| Multivariate Hypergeometric model has very low number of effective samples | 4 | 1262 | April 22, 2019 |