[BUG] Dimension mismatch after distributed segmentation of 2D image

[drhochbaum]

I successfully ran distributed segmentation of a 2D zarr array, but the code fails after processing all the chunks:

---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)
Cell In[21], line 24
     13 cluster_kwargs = {
     14     'n_workers':1,    # if you only have 1 gpu, then 1 worker is the right choice
     15     'ncpus':8,
     16     'memory_limit':'64GB',
     17     'threads_per_worker':1,
     18 }
     20 # run segmentation
     21 # outputs:
     22 #     segments: zarr array containing labels
     23 #     boxes: list of bounding boxes around all labels (very useful for navigating big data)
---> 24 segments, boxes = distributed_eval(
     25     input_zarr=data_zarr,
     26     blocksize=(2048, 2048),
     27     write_path='/Users/XXX/YYY/data_chunk.zarr',
     28     model_kwargs=model_kwargs,
     29     eval_kwargs=eval_kwargs,
     30     cluster_kwargs=cluster_kwargs,
     31 )

File ~/miniforge3/envs/FISH/lib/python3.12/site-packages/cellpose/contrib/distributed_segmentation.py:357, in cluster.<locals>.create_or_pass_cluster(*args, **kwargs)
    355     with cluster_constructor(**kwargs['cluster_kwargs']) as cluster:
    356         kwargs['cluster'] = cluster
...
--> 182     raise RuntimeError('structure and input must have equal rank')
    183 for ii in structure.shape:
    184     if ii != 3:

RuntimeError: structure and input must have equal rank
------------------------------------------------------------------------------------------

I chunked a 2D image of DAPI stained nuclei into a zarr array with chunks = (2048,2048)

Then I ran the distributed segmentation:

# Run CELLPOSE3 distributed segmentation
from cellpose.contrib.distributed_segmentation import distributed_eval

# parameterize cellpose however you like
model_kwargs = {'gpu':True, 'model_type':'nuclei'}  # can also use 'pretrained_model'
eval_kwargs = {'diameter':100,
            'z_axis':0,
            'channels':[1,0],
            'do_3D':False,
}

# define compute resources for local workstation
cluster_kwargs = {
    'n_workers':1,    # if you only have 1 gpu, then 1 worker is the right choice
    'ncpus':8,
    'memory_limit':'64GB',
    'threads_per_worker':1,
}

# run segmentation
# outputs:
#     segments: zarr array containing labels
#     boxes: list of bounding boxes around all labels (very useful for navigating big data)
segments, boxes = distributed_eval(
    input_zarr=data_zarr,
    blocksize=(2048, 2048),
    write_path='/Users/XXX/YYY/chunked_data.zarr',
    model_kwargs=model_kwargs,
    eval_kwargs=eval_kwargs,
    cluster_kwargs=cluster_kwargs,
)

After running through all the chunks the error above is thrown. The log file shows that nuclei were successfully detected in a majority of the chunks.

dask_worker_0.log

[RichieHakim]

This issue can be recreated with a random noise image as well:

image = np.random.randint(low=0, high=255, size=(41142, 86593, 1), dtype=np.uint8)

# RUN CELLPOSE3
from cellpose import denoise, io

io.logger_setup() # run this to get printing of progress

# DEFINE CELLPOSE MODEL
dn = denoise.DenoiseModel(gpu=use_gpu, model_type="denoise_nuclei")
imgs_dn = dn.eval(image.transpose(2,0,1),
                  diameter=100., 
                  channels=[1, 0],
)
 
# Run CELLPOSE3 distributed segmentation
from cellpose.contrib.distributed_segmentation import numpy_array_to_zarr
data_zarr = numpy_array_to_zarr('/path/to/file.zarr', image[:,:,0], chunks=(2048, 2048))

# Run CELLPOSE3 distributed segmentation
from cellpose.contrib.distributed_segmentation import distributed_eval

# parameterize cellpose however you like
model_kwargs = {'gpu':True, 'model_type':'cyto3'}  # can also use 'pretrained_model'
eval_kwargs = {'diameter':100,
            'z_axis':0,
            'channels':[1,0],
            'do_3D':False,
}

# define compute resources for local workstation
cluster_kwargs = {
    'n_workers':1,    # if you only have 1 gpu, then 1 worker is the right choice
    'ncpus':8,
    'memory_limit':'64GB',
    'threads_per_worker':1,
}

# run segmentation
# outputs:
#     segments: zarr array containing labels
#     boxes: list of bounding boxes around all labels (very useful for navigating big data)
segments, boxes = distributed_eval(
    input_zarr=data_zarr,
    blocksize=(2048, 2048),
    write_path='/path/to/file.zarr',
    model_kwargs=model_kwargs,
    eval_kwargs=eval_kwargs,
    cluster_kwargs=cluster_kwargs,
)

It fails on block (20, 42) every time, just like the OP's issue.

[RichieHakim]

this PR I think solves this problem: https://github.com/MouseLand/cellpose/pull/1123