The library uses float64 by default to preserve behavior of an earlier numpy version of the library. This is configured in factor.py. Usage of float32 may improve speed and reduce memory requirements of important operations, but could affect numerical stability in some cases. For this issue, we want to understand the impact of dtype on model estimation and provide guidance to users. Depending on the outcome of this analysis, we may want to disable float64 by default to match standard Jax behavior. We may also want to expand some of the APIs to make it easier to specify dtype rather than relying on global defaults.
The library uses float64 by default to preserve behavior of an earlier numpy version of the library. This is configured in factor.py. Usage of float32 may improve speed and reduce memory requirements of important operations, but could affect numerical stability in some cases. For this issue, we want to understand the impact of dtype on model estimation and provide guidance to users. Depending on the outcome of this analysis, we may want to disable float64 by default to match standard Jax behavior. We may also want to expand some of the APIs to make it easier to specify dtype rather than relying on global defaults.