Enable pip install by deferring conda-only imports

[keejkrej]

Summary

This PR makes the core micro-sam Python API importable and usable via pip without requiring conda-forge-only packages (nifty, vigra, python-elf).

Changes:

• Defer imports of nifty, vigra, and python-elf to the functions that use them
• Add pip-installable install_requires to setup.cfg (PyPI-available deps only)

Relates to #1113.

Motivation

Several dependencies used by micro-sam are only available on conda-forge. Importing micro_sam.util currently pulls them in at module load time, which breaks pip-only installs — even when the caller only needs the core SAM workflow (load model → compute embeddings → segment from prompts).

This matters for:

• Integration into applications that manage their own pip environments (e.g. Cell-ACDC)
• HPC / headless setups without a full conda stack

Why this doesn't affect the core neural network workflow

The SAM inference path (get_sam_model, precompute_image_embeddings, segment_from_points) depends on torch, segment-anything, and standard pip packages. It does not use nifty, vigra, or python-elf.

The deferred packages serve orthogonal infrastructure, not the model itself:

Package	Used for	When needed
nifty	Block tiling for large images	Tiled 2D/3D inference only
vigra	Eccentricity-based centroids (mode="v")	Optional prompt generation
python-elf	H5/N5/Zarr I/O, parallel connected components	Keyed file loading, mask post-processing

None of these participate in encoder/decoder forward passes. Users on pip who don't need tiling, vigra centroids, or elf I/O can install and run the core workflow without conda. Users who need those features can still conda install -c conda-forge nifty vigra python-elf as today.

Test plan

• from micro_sam.util import get_sam_model succeeds without nifty/vigra/python-elf installed
• from micro_sam.prompt_based_segmentation import segment_from_points succeeds without nifty installed
• Existing conda CI/tests pass unchanged
• Tiled inference still works when nifty is installed
• load_image_data(..., key=...) still works when python-elf is installed

Made with Cursor

[constantinpape]

Hi @keejkrej ,

thank you for your interest in improving the usability of micro_sam. We are currently working on another solution for this issue though and are creating a dedicated C++ library with all functionality that is needed and that can be installed via pip:
https://github.com/computational-cell-analytics/bioimage-cpp

This has the advantage that all features of micro_sam will be available to pip users. We are already in the process of migrating elf, see constantinpape/elf#141 and will migrate torch_em and then micro_sam next. I expect this to be done in 2-3 weeks.

In general, contributions to micro_sam (or other libraries developed by us) are very welcome, but I recommend to first open an issue for any larger contribution to see if working on this makes sense.

[keejkrej]

Thanks! Very interested in your new lib, for now I'll just install from my fork.