-
Notifications
You must be signed in to change notification settings - Fork 17
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Add tables weight-formats and pre-post-processing #43
Conversation
@oeway This would be my suggestion for listing the supported weight formats, pre-processing, and post-processing, as well as listing which consumer supports which option. Let me know what you think. Once we agree on a format I will try to fill the lists for pre/post-processing and ask others to double check. |
Thanks! Looks good to me! In addition to yes and no, there might be partial implementations, and we can add a footnote below the table. |
Ok, I will finish up everything as far as I can and then ask the others to fill in the gaps.
Good point; we can add this if it becomes relevant. |
I gathered all we had for pre/post-processing in #32, #37 and #39. |
Co-authored-by: Wei Ouyang <oeway007@gmail.com>
Update weight format compatibility
It looks very good! thank you I might have missed it, sorry, what is For the post-processing, should we allow basic thresholding? something like: - `intensity_threshold` binarize the intensity values of an image by thresholding them with a specific value.
- `kwargs`
- `threshold` For example [0.5, 75, 0.25] in the case of a 3 channel image where the channels are not thresholded jointly
- `reference_implementation` For the pre/post-processing support in @deepimagej we have kind of a solution but it's not clear for us whether it can be considered as "supported processing": both steps are given in IJ macros, so as long as they are not super complicated workflows, we are able to reproduce most of the defined pre/post processings. The problem is that we are not really reading the transformations from the yaml, but from the macros. (Comments are very welcome) |
My idea was to provide a link to the reference implementation for the given pre/post-processing, to make it easier to reimplement it. (I haven't provided any links yet, we can fill that in as we go.)
I think providing basic thresholding as post-processing is a good idea.
Are the macros provided separately by the users? |
Good!
Yes.
Yes but not really at this moment. What we have is a GitHub repo containing processing routines (macros), so the users can (manually) download and include them in the model package. Here, we can provide the ones specified as "bioimage.io macros" that are exactly what it is defined here. With the example input/output images, it is possible to know whether the model is working properly on different consumers, but still... |
Ok, then I would suggest that you mark the ones where you have a corresponding macro with yes (and you can also provide the link to macro). And we can then add a footnote that explains that the pre/post-processing currently needs to be selected manually in deepImageJ. |
Can we add a 'binarize' postprocessing step with |
@constantinpape the lower and upper percentile can be between |
Yes, good point. |
Would it be weird to disallow the upper percentile to be <1 as a sanity check? Of course it's a valid value, but someone might specify the percentiles in 0,1... and we wouldn't catch it |
Co-authored-by: FynnBe <thefynnbe@gmail.com>
Co-authored-by: FynnBe <thefynnbe@gmail.com>
Co-authored-by: FynnBe <thefynnbe@gmail.com>
Co-authored-by: FynnBe <thefynnbe@gmail.com>
for pre-/postprocessing
pointing to pytorch-bioimage-io instead, where comments have been added; avoiding to maintain the same model twice.
Tables update
Can I add one more request? 🙏 We use |
Sure! I will add it later and then would merge this PR. We can fill in what's missing in later PR, this one is becoming pretty long already... Any objections? |
Also, |
Ok, we should use the same name here, I will change it to min/max in the preprocessing.
Do you mean that the mode |
@constantinpape yes I see that the perceniles in |
No description provided.