Hey there,
this is just to let you know I just released v2.0.0 of the nnInteractive backend.
The main features are:
- speed improvements
- replaced interactions tensor to be a compressed blosc2 ndarray. This substantially reduced RAM footprint of interactions at negligible compute overhead
- reduced peak VRAM consumption by factor 2-3x when autozoom is triggered
- compatibility with soon-to-be-released new nnInteractive model
Overall the most noticeable improvements should be for very large images.
It has some minor API changes that I have summarized here. Take a look at the TLDR section.
If possible, please use the new optimized code paths for lasso/scribble where only a crop is passed to the backend, that has substantial performance implications!
Let me know in case you have questions!
Best,
Fabian
Hey there,
this is just to let you know I just released v2.0.0 of the nnInteractive backend.
The main features are:
Overall the most noticeable improvements should be for very large images.
It has some minor API changes that I have summarized here. Take a look at the TLDR section.
If possible, please use the new optimized code paths for lasso/scribble where only a crop is passed to the backend, that has substantial performance implications!
Let me know in case you have questions!
Best,
Fabian