cellmap_flow.serving.virtual_zarr

The zarr v2 array an inference server makes up, as data.

A server serves one model’s output as <dataset>/.zattrs, <dataset>/s0/.zarray and <dataset>/s0/<z>.<y>.<x>[.<c>] chunks, which neuroglancer reads as an OME-NGFF 0.4 group with one level. What those say, and which region each chunk key is, is plain arithmetic on the model’s geometry and the raw data’s; these functions are that arithmetic, with no Flask and no process-wide state, so the server’s routes only call them.

Every client reads this format, including dashboards older than the server, so it changes only additively.

Attributes

`logger`
`BLOSC`

Functions

`zattrs`(→ dict)	The group's OME-NGFF attributes: one level, s0, at the output voxel size.
`zarray`(→ dict)	The array's zarr v2 metadata. `dtype` is anything np.dtype takes;
`served_spatial_shape`(→ list)	Output voxels from `origin` to the end of the raw data, rounded up.
`chunk_roi`(→ funlib.geometry.Roi)	The region (nm) chunk `index` covers: `spatial_block` output voxels
`reorder_to_zarr_axes`(→ numpy.ndarray)	`data`, laid out as `model_axes`, in zarr's order: the spatial axes, then "c".
`chunk_encoder`(→ numcodecs.Blosc)	The codec for the chunks, as `BLOSC` describes it.

Module Contents

cellmap_flow.serving.virtual_zarr.logger

cellmap_flow.serving.virtual_zarr.BLOSC

cellmap_flow.serving.virtual_zarr.zattrs(axes, output_voxel_size, origin, has_channel: bool, name: str) → dict

The group’s OME-NGFF attributes: one level, s0, at the output voxel size.

axes are the spatial axes, origin the lower corner of output voxel 0 in nm; OME writes voxel 0’s centre, so neuroglancer draws voxel 0 at the origin. A channel axis comes last, with scale 1 and translation 0.

cellmap_flow.serving.virtual_zarr.zarray(shape, chunks, dtype) → dict: The array’s zarr v2 metadata. dtype is anything np.dtype takes; its typestr is written (“<f4”, “|u1”, …).

cellmap_flow.serving.virtual_zarr.served_spatial_shape(raw_offset, raw_shape, raw_voxel_size, origin, output_voxel_size) → list: Output voxels from origin to the end of the raw data, rounded up.

cellmap_flow.serving.virtual_zarr.chunk_roi(index, spatial_block, output_voxel_size, origin) → funlib.geometry.Roi: The region (nm) chunk index covers: spatial_block output voxels per chunk, the grid starting at origin.

cellmap_flow.serving.virtual_zarr.reorder_to_zarr_axes(data, model_axes, spatial_axes) → numpy.ndarray: data, laid out as model_axes, in zarr’s order: the spatial axes, then “c”.

cellmap_flow.serving.virtual_zarr.chunk_encoder() → numcodecs.Blosc: The codec for the chunks, as BLOSC describes it.