Pulling models from trackastra leads to error due to writting restriction in containerized microsam

[rjlopez2]

We are building a container with napari-microsam in our computer cluster runing on a virtual desktop that runs on an openondemand app.
After loading a sample image and recreate the tutorial for object tracking with trackastra we encounter the following error:

INFO: Computed embeddings for the 'Light Microscopy' model of size 'base'.
---------------------------------------------------------------------------
OSError                                   Traceback (most recent call last)
File /opt/conda/envs/napari-microSAM_env/lib/python3.13/site-packages/micro_sam/sam_annotator/_widgets.py:2013, in AutoSegmentWidget.__call__(self=<micro_sam.sam_annotator._widgets.AutoTrackWidget object>)
   2007     kwargs = {
   2008         "pred_iou_thresh": self.pred_iou_thresh,
   2009         "stability_score_thresh": self.stability_score_thresh,
   2010         "box_nms_thresh": self.box_nms_thresh,
   2011     }
   2012 if self.volumetric and self.apply_to_volume:
-> 2013     worker = self._run_segmentation_3d(kwargs)
        self = <micro_sam.sam_annotator._widgets.AutoTrackWidget object at 0x7fd5c8d09910>
        kwargs = {'center_distance_threshold': 0.4, 'boundary_distance_threshold': 0.5, 'min_size': 100}   2014 elif self.volumetric and not self.apply_to_volume:
   2015     i = int(self._viewer.dims.point[0])

File /opt/conda/envs/napari-microSAM_env/lib/python3.13/site-packages/micro_sam/sam_annotator/_widgets.py:2103, in AutoTrackWidget._run_segmentation_3d(self=<micro_sam.sam_annotator._widgets.AutoTrackWidget object>, kwargs={'boundary_distance_threshold': 0.5, 'center_distance_threshold': 0.4, 'min_size': 100})
   2100     self._viewer.layers["auto_segmentation"].data = segmentation
   2101     self._viewer.layers["auto_segmentation"].refresh()
-> 2103 result = seg_impl()
   2104 update_segmentation(result)

File /opt/conda/envs/napari-microSAM_env/lib/python3.13/site-packages/micro_sam/sam_annotator/_widgets.py:2083, in AutoTrackWidget._run_segmentation_3d.<locals>.seg_impl()
   2080     pbar_signals.pbar_update.emit(1)
   2082 pbar_signals.pbar_reset.emit()
-> 2083 segmentation, lineages = track_across_frames(
        segmentation = array([[[   0,    0, ...,    0,    0],
        [   0,    0, ...,    0,    0],
        ...,
        [   0,    0, ...,    0,    0],
        [   0,    0, ...,    0,    0]],

       [[   0,    0, ...,    0,    0],
        [   0,    0, ...,    0,    0],
        ...,
        [   0,    0, ...,    0,    0],
        [   0,    0, ...,    0,    0]],

       ...,

       [[   0,    0, ..., 1084, 1084],
        [   0,    0, ..., 1084, 1084],
        ...,
        [   0,    0, ...,    0,    0],
        [   0,    0, ...,    0,    0]],

       [[   0,    0, ...,    0,    0],
        [   0,    0, ...,    0,    0],
        ...,
        [   0,    0, ...,    0,    0],
        [   0,    0, ...,    0,    0]]],
      shape=(84, 512, 512), dtype=uint32)
        timeseries = array([[[132, 128, ..., 131, 126],
        [125, 136, ..., 119, 112],
        ...,
        [132, 130, ..., 127, 131],
        [128, 131, ..., 132, 131]],

       [[116, 114, ..., 114, 114],
        [111, 115, ..., 116, 117],
        ...,
        [121, 125, ..., 120, 118],
        [120, 124, ..., 116, 116]],

       ...,

       [[116, 113, ..., 116, 113],
        [115, 119, ..., 121, 113],
        ...,
        [123, 127, ..., 122, 126],
        [122, 123, ..., 115, 127]],

       [[118, 118, ..., 120, 110],
        [118, 119, ..., 127, 116],
        ...,
        [122, 119, ..., 124, 118],
        [117, 121, ..., 126, 121]]], shape=(84, 512, 512), dtype=uint8)
        pbar_signals = <micro_sam.sam_annotator._widgets.PBarSignals object at 0x7fd5c8d0bc80>   2084     timeseries, segmentation,
   2085     verbose=True, pbar_init=pbar_init,
   2086     pbar_update=lambda update: pbar_signals.pbar_update.emit(1),
   2087 )
   2088 pbar_signals.pbar_stop.emit()
   2089 return (segmentation, lineages)

File /opt/conda/envs/napari-microSAM_env/lib/python3.13/site-packages/micro_sam/multi_dimensional_segmentation.py:638, in track_across_frames(timeseries=array([[[132, 128, ..., 131, 126],
        [125,..., 126, 121]]], shape=(84, 512, 512), dtype=uint8), segmentation=array([[[   0,    0, ...,    0,    0],
        [...  0]]],
      shape=(84, 512, 512), dtype=uint32), gap_closing=None, min_time_extent=None, verbose=True, pbar_init=<function AutoTrackWidget._run_segmentation_3d.<locals>.seg_impl.<locals>.pbar_init>, pbar_update=<function AutoTrackWidget._run_segmentation_3d.<locals>.seg_impl.<locals>.<lambda>>, output_folder=None)
    635 if gap_closing is not None and gap_closing > 0:
    636     segmentation = _preprocess_closing(segmentation, gap_closing, pbar_update)
--> 638 segmentation, lineage = _tracking_impl(
        segmentation = array([[[   0,    0, ...,    0,    0],
        [   0,    0, ...,    0,    0],
        ...,
        [   0,    0, ...,    0,    0],
        [   0,    0, ...,    0,    0]],

       [[   0,    0, ...,    0,    0],
        [   0,    0, ...,    0,    0],
        ...,
        [   0,    0, ...,    0,    0],
        [   0,    0, ...,    0,    0]],

       ...,

       [[   0,    0, ..., 1084, 1084],
        [   0,    0, ..., 1084, 1084],
        ...,
        [   0,    0, ...,    0,    0],
        [   0,    0, ...,    0,    0]],

       [[   0,    0, ...,    0,    0],
        [   0,    0, ...,    0,    0],
        ...,
        [   0,    0, ...,    0,    0],
        [   0,    0, ...,    0,    0]]],
      shape=(84, 512, 512), dtype=uint32)
        np = <module 'numpy' from '/opt/conda/envs/napari-microSAM_env/lib/python3.13/site-packages/numpy/__init__.py'>
        timeseries = array([[[132, 128, ..., 131, 126],
        [125, 136, ..., 119, 112],
        ...,
        [132, 130, ..., 127, 131],
        [128, 131, ..., 132, 131]],

       [[116, 114, ..., 114, 114],
        [111, 115, ..., 116, 117],
        ...,
        [121, 125, ..., 120, 118],
        [120, 124, ..., 116, 116]],

       ...,

       [[116, 113, ..., 116, 113],
        [115, 119, ..., 121, 113],
        ...,
        [123, 127, ..., 122, 126],
        [122, 123, ..., 115, 127]],

       [[118, 118, ..., 120, 110],
        [118, 119, ..., 127, 116],
        ...,
        [122, 119, ..., 124, 118],
        [117, 121, ..., 126, 121]]], shape=(84, 512, 512), dtype=uint8)
        min_time_extent = None
        output_folder = None    639     timeseries=np.asarray(timeseries),
    640     segmentation=segmentation,
    641     mode="greedy",
    642     min_time_extent=min_time_extent,
    643     output_folder=output_folder,
    644 )
    645 return segmentation, lineage

File /opt/conda/envs/napari-microSAM_env/lib/python3.13/site-packages/micro_sam/multi_dimensional_segmentation.py:578, in _tracking_impl(timeseries=array([[[132, 128, ..., 131, 126],
        [125,..., 126, 121]]], shape=(84, 512, 512), dtype=uint8), segmentation=array([[[   0,    0, ...,    0,    0],
        [...  0]]],
      shape=(84, 512, 512), dtype=uint32), mode='greedy', min_time_extent=None, output_folder=None)
    576 def _tracking_impl(timeseries, segmentation, mode, min_time_extent, output_folder=None):
    577     device = "cuda" if torch.cuda.is_available() else "cpu"
--> 578     model = Trackastra.from_pretrained("general_2d", device=device)
        device = 'cuda'    579     lineage_graph = model.track(timeseries, segmentation, mode=mode)
    580     track_data, parent_graph, _ = graph_to_napari_tracks(lineage_graph)

File /opt/conda/envs/napari-microSAM_env/lib/python3.13/site-packages/trackastra/model/model_api.py:152, in Trackastra.from_pretrained(cls=<class 'trackastra.model.model_api.Trackastra'>, name='general_2d', device='cuda', download_dir=None)
    136 @classmethod
    137 def from_pretrained(
    138     cls, name: str, device: str | None = None, download_dir: Path | None = None
    139 ):
    140     """Load a pretrained Trackastra model.
    141 
    142     Available pretrained models are described in detail in pretrained.json.
   (...)    150         Trackastra model instance.
    151     """
--> 152     folder = download_pretrained(name, download_dir)
        name = 'general_2d'
        download_dir = None    153     # download zip from github to location/name, then unzip
    154     return cls.from_folder(folder, device=device)

File /opt/conda/envs/napari-microSAM_env/lib/python3.13/site-packages/trackastra/model/pretrained.py:66, in download_pretrained(name='general_2d', download_dir=PosixPath('/opt/conda/envs/napari-microSAM_env/lib/python3.13/site-packages/trackastra/.models'))
     63 else:
     64     download_dir = Path(download_dir)
---> 66 download_dir.mkdir(exist_ok=True, parents=True)
        download_dir = PosixPath('/opt/conda/envs/napari-microSAM_env/lib/python3.13/site-packages/trackastra/.models')     67 try:
     68     url = _MODELS[name]

File /opt/conda/envs/napari-microSAM_env/lib/python3.13/pathlib/_local.py:722, in Path.mkdir(self=PosixPath('/opt/conda/envs/napari-microSAM_env/lib/python3.13/site-packages/trackastra/.models'), mode=511, parents=True, exist_ok=True)
    718 """
    719 Create a new directory at this given path.
    720 """
    721 try:
--> 722     os.mkdir(self, mode)
        self = PosixPath('/opt/conda/envs/napari-microSAM_env/lib/python3.13/site-packages/trackastra/.models')
        mode = 511    723 except FileNotFoundError:
    724     if not parents or self.parent == self:

OSError: [Errno 30] Read-only file system: '/opt/conda/envs/napari-microSAM_env/lib/python3.13/site-packages/trackastra/.models'

our setup ("napari --info" command bellow):

(base) [rl18y689@gnode28 development]$ apptainer exec --nv napari-microSAM-1.6.2.sif /opt/launchers/start_napari.sh --info
WARNING: could not determine DPI
WARNING:vispy:could not determine DPI
WARNING: could not determine DPI
WARNING:vispy:could not determine DPI
WARNING: could not determine DPI
WARNING:vispy:could not determine DPI
WARNING: could not determine DPI
WARNING:vispy:could not determine DPI
napari: 0.6.6 (from conda)
Platform: Linux-5.14.0-570.58.1.el9_6.x86_64-x86_64-with-glibc2.34
System: Rocky Linux 9.7 (Blue Onyx)
Python: 3.13.11 | packaged by conda-forge | (main, Dec  6 2025, 11:24:03) [GCC 14.3.0]
Qt: 5.15.15
PyQt5: 5.15.11
NumPy: 2.3.5
SciPy: 1.16.3
Dask: 2025.11.0
VisPy: 0.15.2
magicgui: 0.10.1
superqt: 0.7.6
in-n-out: 0.2.1
app-model: 0.4.0
psygnal: 0.15.0
npe2: 0.7.9
pydantic: 2.11.10

OpenGL:
  - PyOpenGL: 3.1.10
  - GL version:  4.6.0 NVIDIA 550.163.01
  - MAX_TEXTURE_SIZE: 32768
  - GL_MAX_3D_TEXTURE_SIZE: 16384

Screens:
  - screen 1: resolution 1824x1089, scale 1.0

Optional:
  - numba: 0.63.0
  - triangle not installed
  - napari-plugin-manager: 0.1.8
  - bermuda: 0.1.6
  - PartSegCore: 0.15.14

Experimental Settings:
  - Async: False
  - Autoswap buffers: False
  - Triangulation backend: Fastest available

Settings path:
  - /storage/homefs/rl18y689/.vibe/.config/napari/napari-microSAM_env_6a700390b4d4d8d8b66f95ee9e18d42c59217c64/settings.yaml

Launch command:
  - /opt/conda/envs/napari-microSAM_env/bin/napari --info

Plugins:
  - micro-sam: 1.6.2 (26 contributions)
  - napari: 0.6.6 (88 contributions)
  - napari-console: 0.1.4 (0 contributions)
  - napari-svg: 0.2.1 (2 contributions)
  - skan: 0.13.0 (4 contributions)

and here is the yaml file to recreate our python environment:

name: napari-microSAM_env
channels:
  - conda-forge
dependencies:
  - pip
  - pyqt
  - napari
  - micro_sam
  - pip:
    - git+https://github.com/ChaoningZhang/MobileSAM.git
    - trackastra
  

I wander if there is way to change or allow to choose a safer location where the models would be downloaded to as is done here to pull the sample data?:

micro-sam/micro_sam/sample_data.py

Line 4 in 32f07b0

by setting the environment variable: MICROSAM_CACHEDIR

Or to use the same cache folder where the example images are downloaded?

[rjlopez2]

Further Looking at the source code from trackastra seem that the origin of the problem could be there:
https://github.com/weigertlab/trackastra/blob/8f3e96e84926160ab3d871eedabc3f448cf68834/trackastra/model/pretrained.py#L60C5-L60C24

I will also poke them to see how to address this issue.

[rjlopez2]

This issue has been addressed by the trackastra developers here weigertlab/trackastra#49.
closing

[anwai98]

Hi @rjlopez2,

Thanks for opening the issue and apologies for the late response (was trying to meet some deadlines). That sounds great that it's fixed now! (all thanks to @bentaculum ;)).

Let us know if you have anymore questions. And thanks for your interest in micro-sam!