Installing micro-sam headless

[ElpadoCan]

Hello,

First of all, congratulations on the publication, well-deserved!

Secondly, I was wondering if it's possible to install micro-sam headless, without GUI libraries (such as qt-main).

I'm asking because I'm planning to integrate micro-sam into Cell-ACDC GUI as plugin. However, in Cell-ACDC the GUI libraries are installed via pip and (as we all know) conda does not communicate well with pip.

I would only need the core functionalities of micro-sam without the GUI.

Is that possible? I could not find anything in the documentation. I apologise if I missed it.

If not possible yet, how hard would it be to implement for you? This would also enable installation of micro-sam on HPC clusters where often there is no desktop environment.

Thanks!

Francesco

[constantinpape]

Hi Francesco,

yes installation without the GUI libraries is no problem. We have designed the package in such a way that the GUI part is encapsulated into sam_annotator. The rest of the micro_sam library can be used without GUI-related dependencies.

So you should be able to set up an environment without these lines from the env file:

• https://github.com/computational-cell-analytics/micro-sam/blob/master/environment.yaml#L7-L8
• https://github.com/computational-cell-analytics/micro-sam/blob/master/environment.yaml#L12

to get all required dependencies except GUI related ones.

Let us know if you run into any issues with this or if you have any further questions.

[ElpadoCan]

Hi Constantin,

Thanks for the quick reply!

Oh yeah, you are right, I'm stupid, I didn't see you also had the pip install -e . installation option, I knew I was missing something :D Thanks!

Speaking of pip, are you planning to release on PyPi? A micro-sam package on PyPI without dependencies (for software like Cell-ACDC that automatically manages environments for users) would be extremely helpful, and it seems like you have everything in place for that.

Cheers,
Francesco

[constantinpape]

Speaking of pip, are you planning to release on PyPi? A micro-sam package on PyPI without dependencies (for software like Cell-ACDC that automatically manages environments for users) would be extremely helpful, and it seems like you have everything in place for that.

Unfortunately that is not possible right now, because we have (non-gui) dependencies that are conda-only. Removing those would be a major effort. What would be possible in a relative easy way is to have a micro-sam headless version on conda-forge.

[ElpadoCan]

Got it, makes sense. Headless on conda-forge would also help, true! Thanks

[constantinpape]

Headless on conda-forge would also help, true!

I will try to set this up next week; will let you know.

[ElpadoCan]

Hi Constantin,

I was trying to set up the headless environment; however, by installing the packages one by one, I found out that when I install the package torch_em, other GUI pacakges get installed, such as magicgui and superqt. Should I remove torch_em too and install it with pip install -e .?

Thanks!

[constantinpape]

Should I remove torch_em too and install it with pip install -e .?

Yes, that's the way to go for now. (torch_em only depends "shallowly" on the GUI packages, so it should work).

I will then also provide a headless package for torch_em to facillitate installation.

(Feel free to leave this open, I will follow up here once the headless packages are available)

[ElpadoCan]

Actually, also elf is installing qtpy and superqt even if I install it with pip, so I guess I would have to go with --no-deps there?

[constantinpape]

Yep exactly, I guess this ended up in the dependencies in the pip file. As before, this is not deeply integrated and should work just fine without the GUI packages. If there is an issue let me know; it should be easy to resolve by updating 1-2 imports.

(And I will add a headless package of this to the list)

[ElpadoCan]

Yes, sure, I'm happy to test and compile a minimal set of packages needed for headless installation.

[ElpadoCan]

Ok, I can report that magicgui, qtpy, etc. are coming from the package skan installed by elf. I guess that complicates things?

In any case, I tried installing everything without skan and now when I run from micro_sam.util import get_sam_model I get the follwing error

>>> from micro_sam.util import get_sam_model
OMP: Error #15: Initializing libomp.dll, but found libiomp5md.dll already initialized.
OMP: Hint This means that multiple copies of the OpenMP runtime have been linked into the program. That is dangerous, since it can degrade performance or cause incorrect results. The best thing to do is to ensure that only a single OpenMP runtime is linked into the process, e.g. by avoiding static linking of the OpenMP runtime in any library. As an unsafe, unsupported, undocumented workaround you can set the environment variable KMP_DUPLICATE_LIB_OK=TRUE to allow the program to continue to execute, but that may cause crashes or silently produce incorrect results. For more information, please see http://openmp.llvm.org/

At the moment, I cannot tel where it is coming from. I installed everything with pip except for vigra that I installed from conda-forge, but since import vigra works fine, it's probably not the source of the issue?

[constantinpape]

Ok, I can report that magicgui, qtpy, etc. are coming from the package skan installed by elf. I guess that complicates things?

Ok, that's good to know! skan is not essential for elf. I can convert it to an optional dependency.

At the moment, I cannot tel where it is coming from. I installed everything with pip except for vigra that I installed from conda-forge, but since import vigra works fine, it's probably not the source of the issue?

Two questions / points:

• Is this an error or just a warning? You could try loading and using the model.
• My guess that this is some issue with either numpy or pytorch. Can you try importing the two and see if this triggers the same output?

[ElpadoCan]

• Is this an error or just a warning? You could try loading and using the model.
• My guess that this is some issue with either numpy or pytorch. Can you try importing the two and see if this triggers the same output?

Actually, you are right, it's only a warning. I will try to use the model and report back. In the meantime I can trigger the warning with these minimal imports

import torch 
from skimage.segmentation import relabel_sequential

I have a feeling that it is caused by some lower-level library but at the moment I cannot tell more because import numpy after import torch works fine

EDIT: Individually, both import torch and from skimage.segmentation import relabel_sequential do not trigger the warning