Skip to main content

A serverless client for reading and writing Neuroglancer Precomputed volumes both locally and on cloud services.

Project description

Build Status PyPI version SfN 2018 Poster codecov DOI

CloudVolume: IO for Neuroglancer Datasets

from cloudvolume import CloudVolume

vol = CloudVolume('gs://mylab/mouse/image', parallel=True, progress=True)
image = vol[:,:,:] # Download a whole image stack into a numpy array from the cloud
vol[:,:,:] = image # Upload an entire image stack from a numpy array to the cloud

label = 1
mesh = vol.mesh.get(label)
skel = vol.skeleton.get(label)

CloudVolume is a serverless Python client for random access reading and writing of Neuroglancer volumes in "Precomputed" format, a set of representations for arbitrarily large volumetric images, meshes, and skeletons. CloudVolume is typically paired with Igneous, a Kubernetes compatible system for generating image hierarchies, meshes, skeletons, and other dependency free jobs that can be applied to petavoxel scale images.

Precomputed volumes are typically stored on AWS S3, Google Storage, or locally. CloudVolume can read and write to these object storage providers given a service account token with appropriate permissions. However, these volumes can be stored on any service, including an ordinary webserver or local filesystem, that supports key-value access.

The combination of Neuroglancer, Igneous, and CloudVolume comprises a system for visualizing, processing, and sharing (via browser viewable URLs) petascale datasets within and between laboratories. A typical example usage would be to visualize raw electron microscope scans of mouse, fish, or fly brains up to a cubic millimeter in physical dimension. Neuroglancer and Igneous would enable you to visualize each step of the process of montaging the image, fine tuning alignment vector fields, creating segmentation layers, ROI masks, or performing other types of analysis. CloudVolume enables you to read from and write to each of these layers. Recently, we have introduced the ability to interact with the graph server ("PyChunkGraph") that backs proofreading automated segmentations via the graphene:// format.

You can find a collection of CloudVolume accessible and Neuroglancer viewable datasets at, an open data project by some of our collaborators.


  • Random access to petavoxel Neuroglancer images, meshes, and skeletons.
  • Nearly all output is immediately visualizable using Neuroglancer.*
  • Reads graph server backed proofreading volumes (via graphene://).
  • Serverless (except graphene://) and multi-cloud.

Detailed Highlights

  • Multi-threaded, supports multi-process and green threads.
  • Memory optimized, supports shared memory.
  • Lossless connectomics relevant codecs (compressed_segmentation, compresso, fpzip, zfpc, png, and brotli)
  • Understands image hierarchies & anisotropic pixel resolutions.
  • Accomodates downloading missing tiles (fill_missing=True).
  • Accomodates uploading compressed black tiles to erasure coded file systems (delete_black_uploads=True).
  • Growing support for the Neuroglancer sharded format which dramatically condenses the number of files required to represent petascale datasets, similar to Cloud Optimized GeoTIFF, which can result in dramatic cost savings.
  • Reads Precomputed meshes and skeletons.
  • Includes viewers for small images, meshes, and skeletons.
  • Only 3 dimensions + RBG channels currently supported for images.
  • No data versioning.


Cloud-volume is regularly tested on Ubuntu with 3.7, 3.8, 3.9 and 3.10. We officially support Linux and Mac OS. Windows is community supported. After installation, you'll also need to set up your cloud credentials if you're planning on writing files or reading from a private dataset. Once you're finished setting up, you can try reading from a public dataset.

pip Binary Installation

pip install cloud-volume # standard installation

CloudVolume depends on several PyPI packages which are Cython bindings for C++. We have provided compiled binaries for many platforms and python versions, however if you are on an unsupported system, pip will attempt to install from source. In that case, follow the instructions below.

Windows Note: If you get errors related to a missing C++ compiler, this blog post might help you:

Optional Dependencies

Tag Description Dependencies
boss boss:// format support intern
test Supports testing pytest
mesh_viewer mesh.viewer() GUI vtk
skeleton_viewer skeleton.viewer() GUI matplotlib
all_viewers All viewers now and in the future. vtk, matplotlib
dask Supports converting to/from dask arrays dask[array]


pip install cloud-volume[boss,test,all_viewers]

pip Source Installation

C++ compiler required.

sudo apt-get install g++ python3-dev # python-dev if you're on python2
pip install numpy
pip install cloud-volume

Due to packaging problems endemic to Python, Cython packages that depend on numpy require numpy header files be installed before attempting to install the package you want. The numpy headers are not recognized unless numpy is installed in a seperate process that runs first. There are hacks for this issue, but I haven't gotten them to work. If you think binaries should be available for your platform, please let us know by opening an issue.

Manual Installation

This can be desirable if you want to hack on CloudVolume itself.

git clone
cd cloud-volume

# With virtualenvwrapper
mkvirtualenv cv
workon cv
# With only virtualenv
virtualenv venv
source venv/bin/activate

sudo apt-get install g++ python3-dev # python-dev if you're on python2
pip install numpy # additional step needed for accelerated compressed_segmentation and fpzip
pip install -e . # without optional dependencies
pip install -e .[all_viewers] # with e.g. the all_viewers optional dependency


You'll need credentials only for the services you'll use. If you plan to use the local filesystem, you won't need any. For Google Storage (setup instructions here), default account credentials will be used if available and no service account is provided.

If neither of those two conditions apply, you need a service account credential. If you have your credentials handy, you can provide them like so as a dict, JSON string, or a bare token if the service will accept that.

cv = CloudVolume(..., secrets=...)

However, it may be simpler to save your credential to disk so you don't have to always provide it. google-secret.json is a service account credential for Google Storage, aws-secret.json is a service account for S3, etc. You can support multiple projects at once by prefixing the bucket you are planning to access to the credential filename. google-secret.json will be your defaut service account, but if you also want to also access bucket ABC, you can provide ABC-google-secret.json and you'll have simultaneous access to your ordinary buckets and ABC. The secondary credentials are accessed on the basis of the bucket name, not the project name.

mkdir -p ~/.cloudvolume/secrets/
mv aws-secret.json ~/.cloudvolume/secrets/ # needed for Amazon
mv google-secret.json ~/.cloudvolume/secrets/ # needed for Google
mv boss-secret.json ~/.cloudvolume/secrets/ # needed for the BOSS
mv matrix-secret.json ~/.cloudvolume/secrets/ # needed for Matrix
mv tigerdata-secret.json ~/.cloudvolume/secrets/ # needed for Tigerdata

aws-secret.json and matrix-secret.json

Create an IAM user service account that can read, write, and delete objects from at least one bucket.



You can create the google-secret.json file here. You don't need to manually fill in JSON by hand, the below example is provided to show you what the end result should look like. You should be able to read, write, and delete objects from at least one bucket.

  "type": "service_account",
  "project_id": "$YOUR_GOOGLE_PROJECT_ID",
  "private_key_id": "...",
  "private_key": "...",
  "client_email": "...",
  "client_id": "...",
  "auth_uri": "",
  "token_uri": "",
  "auth_provider_x509_cert_url": "",
  "client_x509_cert_url": ""


Note: used to be called chunkedgraph-secret.json. This is still supported but deprecated.

If you have a token from Graphene/Chunkedgraph server, create the cave-secret.json file as shown in the example below. You may also pass the token to CloudVolume(..., secrets=token).

  "token": "<your_token>"

Note that to take advantage of multiple credential files, prepend the fully qualified domain name (FQDN) of the server instead of the bucket for GCS and S3. For example,


CloudVolume supports reading and writing to Neuroglancer data layers on Amazon S3, Google Storage, The BOSS, and the local file system.

Supported URLs are of the forms:


The format or protocol fields may be omitted where required. In the case of the precomputed format, the format specifier is optional.

Format Protocols Default Example
precomputed gs, s3, http, https, file, matrix, tigerdata Yes gs://mybucket/dataset/layer
graphene gs, s3, http, https, file, matrix, tigerdata graphene://gs://mybucket/dataset/layer
boss N/A boss://collection/experiment/channel
n5 gs, s3, http, https, file, matrix, tigerdata n5://gs://mybucket/dataset/layer

Supported Formats

Supported Protocols

  • gs: Google Storage
  • s3: Amazon S3
  • http(s): (read-only) Ordinary Web Servers
  • file: Local File System (absolute path)
  • matrix: Princeton Internal System (run in large part by Seung Lab)
  • tigerdata: Princeton Internal System (run by Princeton OIT)

CloudVolume also supports alternative s3 aliases via CloudFiles.

info Files - New Dataset

Neuroglancer relies on an info file located at the root of a dataset layer to tell it how to compute file locations and interpret the data in each file. CloudVolume piggy-backs on this functionality.

In the below example, assume you are creating a new segmentation volume from a 3d numpy array "rawdata". Note Precomputed stores data in Fortran (column major, aka CZYX) order. You should do a small test to see if the image is written transposed. You can fix this by uploading rawdata.T. A more detailed example for uploading a local volume is located here.

from cloudvolume import CloudVolume

info = CloudVolume.create_new_info(
    num_channels    = 1,
    layer_type      = 'segmentation',
    data_type       = 'uint64', # Channel images might be 'uint8'
    # raw, png, jpeg, compressed_segmentation, fpzip, kempressed, zfpc, compresso
    encoding        = 'raw', 
    resolution      = [4, 4, 40], # Voxel scaling, units are in nanometers
    voxel_offset    = [0, 0, 0], # x,y,z offset in voxels from the origin
    mesh            = 'mesh',
    # Pick a convenient size for your underlying chunk representation
    # Powers of two are recommended, doesn't need to cover image exactly
    chunk_size      = [ 512, 512, 16 ], # units are voxels
    volume_size     = [ 250000, 250000, 25000 ], # e.g. a cubic millimeter dataset
vol = CloudVolume(cfg.path, info=info)
vol[cfg.x: cfg.x + cfg.length, cfg.y:cfg.y + cfg.length, cfg.z: cfg.z + cfg.length] = rawdata[:,:,:]
Encoding Image Type Lossless Neuroglancer Viewable Description
raw Any Y Y Serialized numpy arrays.
png Image Y Y Multiple slices stiched into a single PNG.
jpeg Image N Y Multiple slices stiched into a single JPEG.
compressed_segmentation Segmentation Y Y Renumbered numpy arrays to reduce data width. Also used by Neuroglancer internally.
compresso Segmentation Y Y Lossless high compression algorithm for connectomics segmentation.
fpzip Floating Point Y Y* Takes advantage of IEEE 754 structure + L1 Lorenzo predictor to get higher compression.
kempressed Anisotropic Z Floating Point N** Y* Adds manipulations on top of fpzip to achieve higher compression.
zfpc Alignment Vector Fields N*** Y* zfp stream container.

* Not integrated into official Neuroglancer yet, but available on a branch (fpzip, zfpc). ** Lossless if your data can handle adding and then subtracting 2. *** Lossless by default, but you probably want to use the lossy mode.

Note on compressed_segmentation: To use, make sure compressed_segmentation_block_size is specified (usually [8,8,8]. This field will appear in the info file in the relevant scale.

Note on zfpc: To configure, use the fields zfpc_rate, zfpc_precision, zfpc_tolerance, zfpc_correlated_dims in the relevant scale of the info file.


# Basic Examples
vol = CloudVolume('gs://mybucket/retina/image')
vol = CloudVolume('gs://mybucket/retina/image', secrets=token, dict or json)
vol = CloudVolume('gs://bucket/dataset/channel', mip=0, bounded=True, fill_missing=False)
vol = CloudVolume('gs://bucket/dataset/channel', mip=[ 8, 8, 40 ], bounded=True, fill_missing=False) # set mip at this resolution
vol = CloudVolume('gs://bucket/datasset/channel', info=info) # New info file from scratch
image = vol[:,:,:] # Download the entire image stack into a numpy array
image =, mip=2, renumber=True) # download w/ smaller dtype
uniq = vol.unique(bbox, mip=0) # efficient extraction of unique labels
listing = vol.exists( np.s_[0:64, 0:128, 0:64] ) # get a report on which chunks actually exist
exists = vol.image.has_data(mip=0) # boolean check to see if any data is there
listing = vol.delete( np.s_[0:64, 0:128, 0:64] ) # delete this region (bbox must be chunk aligned)
vol[64:128, 64:128, 64:128] = image # Write a 64^3 image to the volume
img = vol.download_point( (x,y,z), size=256, mip=3 ) # download region around (mip 0) x,y,z at mip 3
pts = vol.scattered_points([ (x1,y1,z1), (x2,y2,z2) ]) # download voxel labels located at indicated points
# download image files without decompressing or rendering them. Good for caching!
files = vol.download_files(bbox, mip, decompress=False) 

# Server
vol.viewer() # launches neuroglancer compatible web server on http://localhost:1337

# Microviewer
img = vol[64:1028, 64:1028, 64:128]
img.viewer() # launches web viewer on http://localhost:8080

# Meshes # save 12345 as ./12345.ply on disk[12345, 12346, 12347]) # merge three segments into one file, file_format='obj') # 'ply' and 'obj' are both supported
vol.mesh.get(12345) # return the mesh as vertices and faces instead of writing to disk
vol.mesh.get([ 12345, 12346 ]) # return these two segids fused into a single mesh
vol.mesh.get([ 12345, 12346 ], fuse=False) # return { 12345: mesh, 12346: mesh }
vol.mesh.put(meshes) # works for unsharded legacy only
vol.mesh.delete(segids) # works for unsharded meshes only

mesh.viewer() # Opens GUI. Requires vtk.

# Skeletons
skel = vol.skeleton.get(12345)
vol.skeleton.upload_raw(segid, skel.vertices, skel.edges, skel.radii, skel.vertex_types)

# specified in nm, only available for datasets with a generated index
skels = vol.skeleton.get_by_bbox( Bbox( (0,0,0), (500, 500, 500) ) )
vol.skeleton.spatial_index # None if not available

skel.empty() # boolean

bytes = skel.encode() # encode to Precomputed format (bytes)
skel = Skeleton.decode(bytes) # decode from PrecomputedFormat

skel = skel.crop(slices or bbox) # eliminate vertices and edges outside bbox
skel = skel.consolidate() # eliminate duplicate vertices and edges
skel3 = skel.merge(skel2) # merge two skeletons into one
skel = skel.clone() # create copy
skel = Skeleton.from_swc(swcstr) # decode an SWC file
skel_str = skel.to_swc() # convert to SWC file in string representation
skel.viewer() # Opens GUI. Requires matplotlib

skel.cable_length() # sum of all edge lengths
skel = skel.downsample(2) # reduce size of skeleton by factor of 2

skel1 == skel2 # check if contents of internal arrays match
Skeleton.equivalent(skel1, skel2) # ...even if there are differences like differently numbered edges

# Parallel Operation
vol = CloudVolume('gs://mybucket/retina/image', parallel=True) # Use all cores
vol.parallel = 4 # e.g. any number > 1, use this many cores
data = vol[:] # uses shared memory to coordinate processes under the hood

# Shared Memory Output (can be used by other processes)
vol = CloudVolume(...)
# data backed by a shared memory buffer
# location is optional (defaults to vol.shared_memory_id)
data = vol.download_to_shared_memory(np.s_[:], location='some-example')
vol.unlink_shared_memory() # delete the shared memory associated with this cloudvolume
vol.shared_memory_id # get/set the default shared memory location for this instance

# Shared Memory Upload
vol = CloudVolume(...)
vol.upload_from_shared_memory('my-shared-memory-id', # do not prefix with /dev/shm
    bbox=Bbox( (0,0,0), (10000, 7500, 64) ))

# Download or Upload directly with Files
# The files must be in Precomputed raw format.
vol.download_to_file('/path/to/file', bbox=Bbox(...), mip=0) # bbox is the download region
vol.upload_from_file('/path/to/file', bbox=Bbox(...), mip=0) # bbox is the region it represents

# Transfer w/o Excess Memory Allocation
vol = CloudVolume(...)
# single core, send all of vol to destination, no painting memory
vol.transfer_to('gs://bucket/dataset/layer', vol.bounds)

# Caching, default located at $HOME/.cloudvolume/cache/$PROTOCOL/$BUCKET/$DATASET/$LAYER/$RESOLUTION
# You can also set the cache location using
# cache=str or with environment variable CLOUD_VOLUME_CACHE_DIR
vol = CloudVolume('gs://mybucket/retina/image', cache=True) # Basic Example
image = vol[0:10,0:10,0:10] # Download partial image and cache
vol[0:10,0:10,0:10] = image # Upload partial image and cache

# Resizing and clearing the LRU in-memory cache
vol = CloudVolume(..., lru_bytes=num_bytes) # >= 0, 0 means disabled
vol.image.lru.resize(num_bytes) # same
len(vol.image.lru) # number of items in lru
vol.image.lru.nbytes # size in bytes (not counting LRU structures, nor recursive)
vol.image.lru.items() # etc, also functions as a dict

# Evaluating the on-disk Cache
vol.cache.list() # list files in cache at this mip level
vol.cache.list(mip=1) # list files in cache at mip 1
vol.cache.num_files() # number of files at this mip level
vol.cache.num_bytes(all_mips=True) # Return num files for each mip level in a list
vol.cache.num_bytes() # number of bytes taken up by files, size on disk can be bigger
vol.cache.num_bytes(all_mips=True) # Return num bytes for each mip level in a list

vol.cache.enabled = True/False # Turn the cache on/off
vol.cache.path = Str # set the cache location
vol.cache.compress = None/True/False # None: Link to cloud setting, Boolean: Force cache to compressed (True) or uncompressed (False)

# Deleting Cache
vol.cache.flush() # Delete local cache for this layer at this mip level
vol.cache.flush(preserve=Bbox(...)) # Same, but preserve cache in a region of space
vol.cache.flush_region(region=Bbox(...), mips=[...]) # Delete the cached files in this region at these mip levels (default all mips)

# Using Green Threads
import gevent.monkey

cv = CloudVolume(..., green_threads=True)
img = cv[...] # now green threads will be used

# Dask Interface (requires dask installation)
arr = cv.to_dask()
arr = cloudvolume.dask.from_cloudvolume(cloudpath) # same as to_dask
res = cloudvolume.dask.to_cloudvolume(arr, cloudpath, compute=bool, return_store=bool)

CloudVolume Constructor

    cloudpath:str, mip:int=0, bounded:bool=True, 
    autocrop:bool=False, fill_missing:bool=False, cache:CacheType=False, 
    compress_cache:CompressType=None, cdn_cache:bool=True, 
    progress:bool=INTERACTIVE, info:dict=None, provenance:dict=None,
    compress:CompressType=None, compress_level:Optional[int]=None, 
    non_aligned_writes:bool=False, parallel:ParallelType=1, delete_black_uploads:bool=False, 
    background_color:int=0, green_threads:bool=False, use_https:bool=False,
    max_redirects:int=10, mesh_dir:Optional[str]=None, skel_dir:Optional[str]=None, 
    agglomerate:bool=False, secrets:SecretsType=None, 
    spatial_index_db:Optional[str]=None, lru_bytes:int = 0
  •  agglomerate: (bool, graphene only) sets the default mode for downloading
      images to agglomerated (True) vs watershed (False).
  •  autocrop: (bool) If the specified retrieval bounding box exceeds the
        volume bounds, process only the area contained inside the volume. 
        This can be useful way to ensure that you are staying inside the 
        bounds when `bounded=False`.
  •  background_color: (number) Specifies what the "background value" of the
      volume is (traditionally 0). This is mainly for changing the behavior
      of delete_black_uploads.
  •  bounded: (bool) If a region outside of volume bounds is accessed:
        True: Throw an error
        False: Allow accessing the region. If no files are present, an error 
            will still be thrown. Consider combining this option with 
            `fill_missing=True`. However, this can be dangrous as it allows
            missing files and potentially network errors to be intepreted as 
  •  cache: (bool or str) Store downs and uploads in a cache on disk
          and preferentially read from it before redownloading.
        - falsey value: no caching will occur.
        - True: cache will be located in a standard location.
        - non-empty string: cache is located at this file path
        After initialization, you can adjust this setting via:
        `cv.cache.enabled = ...` which accepts the same values.
        Note: This cache is totally separate from the LRU controlled by 
  •  cdn_cache: (int, bool, or str) Sets Cache-Control HTTP header on uploaded 
      image files. Most cloud providers perform some kind of caching. As of 
      this writing, Google defaults to 3600 seconds. Most of the time you'll 
      want to go with the default. 
      - int: number of seconds for cache to be considered fresh (max-age)
      - bool: True: max-age=3600, False: no-cache
      - str: set the header manually
  •  compress: (bool, str, None) pick which compression method to use.
  •      None: (default) gzip for raw arrays and no additional compression
          for compressed_segmentation and fpzip.
          False=no compression, Overrides defaults
          'gzip': Extension so that we can add additional methods in the future 
                  like lz4 or zstd. 
          'br': Brotli compression, better compression rate than gzip
          '': no compression (same as False).
  •  compress_level: (int, None) level for compression. Higher number results
        in better compression but takes longer.
      Defaults to 9 for gzip (ranges from 0 to 9).
      Defaults to 5 for brotli (ranges from 0 to 11).
  •  compress_cache: (None or bool) If not None, override default compression 
        behavior for the cache.
  •  delete_black_uploads: (bool) If True, on uploading an entirely black chunk,
        issue a DELETE request instead of a PUT. This can be useful for avoiding storing
        tiny files in the region around an ROI. Some storage systems using erasure coding 
        don't do well with tiny file sizes.
  •  fill_missing: (bool) If a chunk file is unable to be fetched:
        True: Use a block of zeros
        False: Throw an error
  •  green_threads: (bool) Use green threads instead of preemptive threads. This
      can result in higher download performance for some compression types. Preemptive
      threads seem to reduce performance on multi-core machines that aren't densely
      loaded as the CPython threads are assigned to multiple cores and the thrashing
      + GIL reduces performance. You'll need to add the following code to the top
      of your program to use green threads:
          import gevent.monkey
  •  lru_bytes: (int) number of bytes used to cache recently used image 
      tiles in memory. This is an in-memory cache and is completely separate from
      the `cache` parameter that handles disk IO. Tiles are stripped over only their
      second stage compression.
  •  info: (dict) In lieu of fetching a neuroglancer info file, use this one.
        This is useful when creating new datasets and for repeatedly initializing
        a new cloudvolume instance.
  •  max_redirects: (int) if > 0, allow up to this many redirects via info file 'redirect'
        data fields. If <= 0, allow no redirections and access the current info file directly
        without raising an error.
  •  mesh_dir: (str) if not None, override the info['mesh'] key before pulling the
      mesh info file.
  •  mip: (int or iterable) Which level of downsampling to read and write from.
        0 is the highest resolution. You can also specify the voxel resolution
        like mip=[6,6,30] which will search for the appropriate mip level.
  •  non_aligned_writes: (bool) Enable non-aligned writes. Not multiprocessing 
        safe without careful design. When not enabled, a 
        cloudvolume.exceptions.AlignmentError is thrown for non-aligned writes. 
    parallel (int: 1, bool): Number of extra processes to launch, 1 means only 
        use the main process. If parallel is True use the number of CPUs 
        returned by multiprocessing.cpu_count(). When parallel > 1, shared
        memory (Linux) or emulated shared memory via files (other platforms) 
        is used by the underlying download.
  •  progress: (bool) Show progress bars. 
        Defaults to True in interactive python, False in script execution mode.
  •  provenance: (string, dict) In lieu of fetching a provenance 
        file, use this one. 
  •  secrets: (dict) provide per-instance authorization tokens. If not provided,
      defaults to looking in .cloudvolume/secrets for necessary tokens.
  •  skel_dir: (str) if not None, override the info['skeletons'] key before 
      pulling the skeleton info file.
  •  spatial_index_db: (str) A path to an sqlite3 or mysql database that follows 
      the following uri schema. sqlite is assumed if no scheme is present in 
      the uri.
      Igneous generated datasets include a JSON based spatial
      database that tiles the dataset. This can be fast enough up to about 100 TVx
      datasets. Above that, a proper database is required for efficient queries.
      We provide multiple SQL database types that the index can be hosted on.
  •  use_https: (bool) maps gs:// and s3:// to their respective https paths. The 
      https paths hit a cached, read-only version of the data and may be faster.

CloudVolume Methods

Better documentation coming later, but for now, here's a summary of the most useful method calls. Use help(cloudvolume.CloudVolume.$method) for more info.

  • create_new_info (class method) - Helper function for creating info files for creating new data layers.
  • refresh_info - Repull the info file.
  • refresh_provenance - Repull the provenance file.
  • bbox_to_mip - Covert a bounding box or slice from one mip level to another.
  • slices_from_global_coords - deprecated, why not use bbox_to_mip? Find the CloudVolume slice from MIP 0 coordinates if you're on a different MIP. Often used in combination with neuroglancer.
  • reset_scales - Delete mips other than 0 in the info file. Does not autocommit.
  • add_scale - Generate a new mip level in the info property. Does not autocommit.
  • commit_info - Push the current info property into the cloud as a JSON file.
  • commit_provenance - Push the current provenance property into the cloud as a JSON file.
  • image - Access image operations directly.
    • download - Download bounding boxes from a given mip level.
    • upload - Upload images to bounding boxes at a given mip level.
    • transfer_to - Transfer data without painting a container array to avoid out of memory errors.
    • exists - Check which chunk files exist in a given bounding box.
    • delete - Delete chunks in a given bounding box at a given mip level.
  • mesh - Access mesh operations
    • get - Download an object. Can merge multiple segmentids
    • save - Download an object and save it in .obj format. You can combine equivialences into a single object too.
  • skeleton - Access Skeletons
    • get - Download an object.
    • upload - Save a skeleton object to the cloud.
  • cache - Access cache operations
    • enabled - Boolean switch to enable/disable cache. If true, on reading, check local disk cache before downloading, and save downloaded chunks to cache. When writing, write to the cloud then save the chunks you wrote to cache. If false, bypass cache completely. The cache is located at $HOME/.cloudvolume/cache.
    • path - Property that shows the current filesystem path to the cache
    • list - List files in cache
    • num_files - Number of files in cache at this mip level , use all_mips=True to get them all
    • num_bytes - Return the number of bytes in cache at this mip level, all_mips=True to get them all
    • flush - Delete the cache at this mip level, preserve=Bbox/slice to save a spatial region
    • flush_region - Delete a spatial region at this mip level
  • exists - Generate a report on which chunks within a bounding box exist.
  • delete - Delete the chunks within this bounding box.
  • transfer_to - Transfer data from a bounding box to another data storage location. Does not allocate memory and transfers in blocks, so can transfer large volumes of data. May be less efficient than a dedicated tool like gsutil or aws s3.
  • unlink_shared_memory - Delete shared memory associated with this instance (vol.shared_memory_id)
  • generate_shared_memory_location - Create a new unique shared memory identifier string. No side effects.
  • download_to_shared_memory - Instead of using ordinary numpy memory allocations, download to shared memory. Be careful, shared memory is like a file and doesn't disappear unless explicitly unlinked. (vol.unlink_shared_memory())
  • upload_from_shared_memory - Upload from a given shared memory block without making a copy.
  • download_point - Download the region around this mip 0 coordinate at a given mip level.

CloudVolume Properties

Accessed as vol.$PROPERTY like vol.mip. Parens next to each property mean (data type:default, writability). (r) means read only, (w) means write only, (rw) means read/write.

  • mip (uint:0, rw) - Read from and write to this mip level (0 is highest res). Each additional increment in the number is typically a 2x reduction in resolution.
  • bounded (bool:True, rw) - If a region outside of volume bounds is accessed throw an error if True or Fill the region with black (useful for e.g. marching cubes's 1px boundary) if False.
  • autocrop (bool:False, rw) - If bounded is False and this option is True, automatically crop requested uploads and downloads to the volume boundary.
  • fill_missing (bool:False, rw) - If a file inside volume bounds is unable to be fetched use a block of zeros if True, else throw an error.
  • delete_black_uploads (bool:False, rw) - If True, issue a DELETE http request instead of a PUT when an individual uploaded chunk is all zeros.
  • info (dict, rw) - Python dict representation of Neuroglancer info JSON file. You must call vol.commit_info() to save your changes to storage.
  • provenance (dict-like, rw) - Data layer provenance file representation. You must call vol.commit_provenance() to save your changes to storage.
  • available_mips (list of ints, r) - Query which mip levels are defined for reading and writing.
  • dataset_name (str, rw) - Which dataset (e.g. test_v0, snemi3d_v0) on S3, GS, or FS you're reading and writing to. Known as an "experiment" in BOSS terminology. Writing to this property triggers an info refresh.
  • layer (str, rw) - Which data layer (e.g. image, segmentation) on S3, GS, or FS you're reading and writing to. Known as a "channel" in BOSS terminology. Writing to this property triggers an info refresh.
  • base_cloudpath (str, r) - The cloud path to the dataset e.g. s3://bucket/dataset/
  • layer_cloudpath (str, r) - The cloud path to the data layer e.g. gs://bucket/dataset/image
  • info_cloudpath (str, r) - Generate the cloud path to this data layer's info file.
  • scales (dict, r) - Shortcut to the 'scales' property of the info object
  • scale (dict, rw)* - Shortcut to the working scale of the current mip level
  • shape (Vec4, r)* - Like numpy.ndarray.shape for the entire data layer.
  • volume_size (Vec3, r)* - Like shape, but omits channel (x,y,z only).
  • num_channels (int, r) - The number of channels, the last element of shape.
  • layer_type (str, r) - The neuroglancer info type, 'image' or 'segmentation'.
  • dtype (str, r) - The info data_type of the volume, e.g. uint8, uint32, etc. Similar to numpy.ndarray.dtype.
  • encoding (str, r) - The neuroglancer info encoding. e.g. 'raw', 'jpeg', 'npz'
  • resolution (Vec3, r)* - The 3D physical resolution of a voxel in nanometers at the working mip level.
  • downsample_ratio (Vec3, r) - Ratio of the current resolution to the highest resolution mip available.
  • chunk_size (Vec3, r)* - Size of the underlying chunks that constitute the volume in storage. e.g. Vec(64, 64, 64)
  • key (str, r)* - The 'directory' we're accessing the current working mip level from within the data layer. e.g. '6_6_30'
  • bounds (Bbox, r)* - A Bbox object that represents the bounds of the entire volume.
  • shared_memory_id (str, rw) - Shared memory location used for parallel operation or for output.

* These properties can also be accessed with a function named like vol.mip_$PROPERTY($MIP). By default they return the current mip level assigned to the CloudVolume, but any mip level can be accessed via the corresponding mip_ function. Example: vol.mip_resolution(2) would return the resolution of mip 2.

VolumeCutout Functions

When you download an image using CloudVolume it gives you a VolumeCutout. These are numpy.ndarray subclasses that support a few extra properties to help make book keeping easier. The major advantage is save_images() which can help you view your dataset as PNG slices.

  • dataset_name - The dataset this image came from.
  • layer - Which layer it came from.
  • mip - Which mip it came from
  • layer_type - "image" or "segmentation"
  • bounds - The bounding box of the cutout
  • num_channels - Alias for vol.shape[3]
  • save_images() - Save Z slice PNGs of the current image to ./saved_images for manual inspection
  • viewer() - Start a local web server (http://localhost:8080) that can view small volumes interactively. This was recently changed from view as view is a useful numpy method.

Viewing a Precomputed Volume on Disk

If you have Precomputed volume onto local disk and would like to point neuroglancer to it:

vol = CloudVolume(...)

You can then point any version of neuroglancer at it using precomputed://http://localhost:1337/NAME_OF_LAYER.


CloudVolume includes a built-in dependency free viewer for 3D volumetric datasets smaller than about 2GB uncompressed. It supports bool, uint8, uint16, uint32, float32, and float64 numpy data types for both images and segmentation and can render a composite overlay of image and segmentation.

You can launch a viewer using the .viewer() method of a VolumeCutout object or by using the view(...) or hyperview(...) functions that come with the cloudvolume module. This launches a web server on http://localhost:8080. You can read more on the wiki.

from cloudvolume import CloudVolume, view, hyperview

channel_vol = CloudVolume(...)
seg_vol = CloudVolume(...)
img = vol[...]
seg = vol[...]

img.viewer() # works on VolumeCutouts
seg.viewer() # segmentation type derived from info
view(img) # alternative for arbitrary numpy arrays
view(seg, segmentation=True)
hyperview(img, seg) # img and seg shape must match

>>> Viewer server listening to http://localhost:8080

There are also seperate viewers for skeleton and mesh objects that can be invoked by calling .viewer() on either object. However, skeletons depend on matplotlib and meshes depend on vtk and OpenGL to function.

pip install vtk matplotlib

Python 2.7 End of Life

Python 2.7 is no longer supported by CloudVolume. Updated versions of pip will download the last supported release 1.21.1. You can read more on the policy page:

Related Projects

  1. Igneous: Computational pipeline for visualizing neuroglancer volumes.
  2. CloudVolume.jl: CloudVolume in Julia
  3. fpzip: A Python Package for the C++ code by Lindstrom et al.
  4. compressed_segmentation: A Python Package wrapping the code for the compressed_segmentation format developed by Jeremy Maitin-Shepard and Stephen Plaza.
  5. Kimimaro: High performance skeletonization of densely labeled 3D volumes.
  6. compresso: High lossless compression of connectomics segmentation. Algorithm by and code derived from Matejek et al.
  7. zfpc: Optimized zfp multi-stream container for alignment vector fields (and similar floating point data).


Thank you to everyone that has contributed past or current to CloudVolume or the ecosystem it serves. We love you!

Jeremy Maitin-Shepard created Neuroglancer and defined the Precomputed format. Yann Leprince provided a pure Python codec for the compressed_segmentation format. Jeremy Maitin-Shepard and Stephen Plaza created C++ code defining the compression and decompression (respectively) protocol for compressed_segmentation. Peter Lindstrom et al. created the fpzip algorithm, and contributed a C++ implementation and advice. Nico Kemnitz adapted our data to fpzip using the "Kempression" protocol (we named it, not him). Dan Bumbarger contributed code and information helpful for getting CloudVolume working on Windows. Fredrik Kihlander's pure python implementation of murmurhash3 and Austin Appleby developed murmurhash3 which is necessary for the sharded format. Ben Falk advocated for and did the bulk of the work on brotli compression. Some of the ideas in CloudVolume are based on work by Jingpeng Wu in BigArrays.jl. Sven Dorkenwald, Manuel Castro, and Akhilesh Halageri contributed advice and code towards implementing the graphene interface. Oluwaseun Ogedengbe contributed documentation for the sharded format. Eric Perlman wrote the reader for Neuroglancer Multi-LOD meshes. Ignacio Tartavull and William Silversmith wrote the initial version of CloudVolume.

Project details

Release history Release notifications | RSS feed

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

cloud-volume-8.16.0.tar.gz (271.9 kB view hashes)

Uploaded source

Built Distribution

cloud_volume-8.16.0-py3-none-any.whl (234.8 kB view hashes)

Uploaded py3

Supported by

AWS AWS Cloud computing Datadog Datadog Monitoring Facebook / Instagram Facebook / Instagram PSF Sponsor Fastly Fastly CDN Google Google Object Storage and Download Analytics Huawei Huawei PSF Sponsor Microsoft Microsoft PSF Sponsor NVIDIA NVIDIA PSF Sponsor Pingdom Pingdom Monitoring Salesforce Salesforce PSF Sponsor Sentry Sentry Error logging StatusPage StatusPage Status page