Running gradslam at higher processing speed about gradslam HOT 2 CLOSED

Hi @apoorvkh,

Indeed, processing speeds and code optimization are some of our focus areas in subsequent releases. (and yes, decreasing numiters is expected to cause degradation in reconstruction quality)

One design choice we've made early on was to have gradslam (at least initial versions) research-ready. This included preferring pure pytorch code over (CUDA and/or JIT) among other calls made.

We're happy to have suggestions from the community on these aspects.

Regarding the specific codeblock you pointed to, the operation there is a CPU -> GPU memcpy and it is indeed a bottleneck. That said, there might indeed be ways to speed things up. We will review dataloading in the near-term. If you find anything, do not hesitate to get back to us

from gradslam.

Thanks @krrish94! That would be good and I will let you know if I find any solutions in the near term.

from gradslam.