jluethi
commented
2 years ago Great summary of the discussion at yesterday's meeting @tcompa :)
This suggests that:
- It is beneficial to add a little bit of same-GPU parallelism.
- This wont' scale to anything more than small batches of 2-3 sites. And for larger datasets (i.e. more Z planes), the maximum allowed batch_size could even be one.
To me, this would suggest that parallelization is interesting, but needs to either be a user parameter (easy) or we need to be able to estimate the load (hard). I'd go for a user parameter to start with. Then we can collect some experience what works well and maybe come up with good heuristics for a given model or such. Also, really good if this is something that can be varied in this task, because a user may e.g. decide to run cellpose at pyramid layer 1 (once this becomes possible using e.g. the ROI approach described above) and then could potentially run ~10 sites at once (given 10 Z layers) etc.
=> The flexibility is conceptually interesting, when and how we'll use it remains a bit of an open question though that we'll figure out more later I'd assume
tcompa
Hi @gusqgm and @jluethi, let's discuss here the details of the image labeling task.