Closed qin-yu closed 1 month ago
Looks very nice! :) It's much cleaner.
Since we are discussing this widget, I have two more ideas/suggestions:
patch_halo
for a standard user? To me, the default always worked the best. Maybe we should also remove it? Single Patch
checkbox may be cryptic. What could be a better alternative? Maybe "Single Patch (Lower Memory Usage)
"
- Is there ever a need to manually adjust the
patch_halo
for a standard user? To me, the default always worked the best. Maybe we should also remove it?
I originally planned to implement automatic halo sizing when I noticed that the halo size was set to a constant value. However, I shifted focus to correct the halo implementation in PlantSeg and pytorch-3dunet, and the initial plan was set aside. I’ll proceed based on this paper: https://nvlpubs.nist.gov/nistpubs/jres/126/jres.126.009.pdf. My belief is that group norm and wrong halo implementation (and perhaps constant halo size) worked in previous settings because the images were always similar. Now I trained new official models with batch norm that reduce hallucination, fixed halo padding that reduce tiling artefact (it's not too wrong to claim we "removed tiling artefact"). The next step is to automatically set a minimal halo size based on the architecture so that the prediction is as exact as theoretically possible and completely removes tiling artefact.
In the end, yes we remove the user-input halo, but do not use a constant. What do you think?
- The
Single Patch
checkbox may be cryptic. What could be a better alternative? Maybe "Single Patch (Lower Memory Usage)
"
Thanks for the advice, I'll fix it tomorrow!
Cool paper, a great work retraining the moodels! :) I like the idea of havig the perfect halo size calculated from the architecture. My point is just that exposing the halo
parameter to the users is probably more confusing than helpful.
- The
Single Patch
checkbox may be cryptic. What could be a better alternative? Maybe "Single Patch (Lower Memory Usage)
"
Hey Lorenzo, I don't actually understand the significance of having "Single Patch". If the model runs on host then PlantSeg finds batch size 1, otherwise finds the best batch size. I guess we can remove it and always find the best batch size for users?
But I made the changes you requested.
There is one use case though: when someone has only one graphics card and plan to run PlantSeg for many images while using the same card. Yes, let's keep it.
- The
Single Patch
checkbox may be cryptic. What could be a better alternative? Maybe "Single Patch (Lower Memory Usage)
"Hey Lorenzo, I don't actually understand the significance of having "Single Patch". If the model runs on host then PlantSeg finds batch size 1, otherwise finds the best batch size. I guess we can remove it and always find the best batch size for users?
But I made the changes you requested.
Update
There is one use case though: when someone has only one graphics card and plan to run PlantSeg for many images while using the same card. Yes, let's keep it.
Yeah, that's exactly the use case. In general, if you have a single graphic card, it's nicer to keep a bit of Slack for other apps.
Changes
This PR improves #247
Improve UI: Change "Single Patch" to "Batch Size" and make auto-batch-size-discover the default
Choose to show PlantSeg or all models in dropdown list from BioImage.IO Model Zoo
Remove PlantSeg model zoo filters for BioImage.IO models
Before this PR