def test_quantize():
qpath = 'file:///tmp/removeme/quantized/'
delete_layer()
delete_layer(qpath)
cf, _ = create_layer(size=(256,256,128,3), offset=(0,0,0), layer_type="affinities")
cv = CloudVolume(cf.cloudpath)
shape = (128, 128, 64)
slices = np.s_[ :shape[0], :shape[1], :shape[2], :1 ]
data = cv[slices]
data *= 255.0
data = data.astype(np.uint8)
task = partial(QuantizeTask,
source_layer_path=cf.cloudpath,
dest_layer_path=qpath,
shape=shape,
offset=(0,0,0),
mip=0,
)
info = create_quantized_affinity_info(
cf.cloudpath, qpath, shape,
mip=0, chunk_size=[64,64,64], encoding='raw'
)
qcv = CloudVolume(qpath, info=info)
qcv.commit_info()
create_downsample_scales(qpath, mip=0, ds_shape=shape)
task()
qcv.mip = 0
qdata = qcv[slices]
assert np.all(data.shape == qdata.shape)
assert np.all(data == qdata)
assert data.dtype == np.uint8
def create_transfer_tasks(
src_layer_path: str,
dest_layer_path: str,
chunk_size: ShapeType = None,
shape: ShapeType = None,
fill_missing: bool = False,
translate: ShapeType = None,
bounds: Optional[Bbox] = None,
mip: int = 0,
preserve_chunk_size: bool = True,
encoding=None,
skip_downsamples: bool = False,
delete_black_uploads: bool = False,
background_color: int = 0,
agglomerate: bool = False,
timestamp: Optional[int] = None,
compress: Union[str, bool] = 'gzip',
factor: ShapeType = None,
sparse: bool = False,
dest_voxel_offset: ShapeType = None,
memory_target: int = MEMORY_TARGET,
max_mips: int = 5,
clean_info: bool = False,
no_src_update: bool = False,
bounds_mip: int = 0,
) -> Iterator:
"""
Transfer data to a new data layer. You can use this operation
to make changes to the dataset representation as well. For
example, you can change the chunk size, compression, bounds,
and offset.
Downsamples will be automatically generated while transferring
unless skip_downsamples is set. The number of downsamples will
be determined by the chunk size and the task shape.
bounds: Bbox specified in terms of the destination image and its
highest resolution.
translate: Vec3 pointing from source bounds to dest bounds
and is in terms of the highest resolution of the source image.
This allows you to compensate for differing voxel offsets
or enables you to move part of the image to a new location.
dest_voxel_offset: When creating a new image, move the
global coordinate origin to this point. This is commonly
used to "zero" a newly aligned image (e.g. (0,0,0))
background_color: Designates which color should be considered background.
chunk_size: (overrides preserve_chunk_size) force chunk size for new layers to be this.
clean_info: scrub additional fields from the info file that might interfere
with later processing (e.g. mesh and skeleton related info).
compress: None, 'gzip', or 'br' Determines which compression algorithm to use
for new uploaded files.
delete_black_uploads: issue delete commands instead of upload chunks
that are all background.
encoding: "raw", "jpeg", "compressed_segmentation", "compresso", "fpzip", or "kempressed"
depending on which kind of data you're dealing with. raw works for everything (no compression)
but you might get better compression with another encoding. You can think of encoding as the
image type-specific first stage of compression and the "compress" flag as the data
agnostic second stage compressor. For example, compressed_segmentation and gzip work
well together, but not jpeg and gzip.
factor: (overrides axis) can manually specify what each downsampling round is
supposed to do: e.g. (2,2,1), (2,2,2), etc
fill_missing: Treat missing image tiles as zeroed for both src and dest.
max_mips: (pairs with memory_target) maximum number of downsamples to generate even
if the memory budget is large enough for more.
memory_target: given a task size in bytes, pick the task shape that will produce the
maximum number of downsamples. Only works for (2,2,1) or (2,2,2).
no_src_update: don't update the source's provenance file
preserve_chunk_size: if true, maintain chunk size of starting mip, else, find the closest
evenly divisible chunk size to 64,64,64 for this shape and use that. The latter can be
useful when mip 0 uses huge chunks and you want to simply visualize the upper mips.
shape: (overrides memory_target) The 3d size of each task. Choose a shape that meets
the following criteria unless you're doing something out of the ordinary.
(a) 2^n multiple of destination chunk size (b) doesn't consume too much memory
(c) n is related to the downsample factor for each axis, so for a factor of (2,2,1) (default)
z only needs to be a single chunk, but x and y should be 2, 4, 8,or 16 times the chunk size.
Remember to multiply 4/3 * shape.x * shape.y * shape.z * data_type to estimate how much memory
each task will require. If downsamples are off, you can skip the 4/3. In the future, if chunk
sizes match we might be able to do a simple file transfer. The problem can be formulated as
producing the largest number of downsamples within a given memory target.
EXAMPLE: destination is uint64 with chunk size (128, 128, 64) with a memory target of
at most 3GB per task and a downsample factor of (2,2,1).
The largest number of downsamples is 4 using 2048 * 2048 * 64 sized tasks which will
use 2.9 GB of memory. The next size up would use 11.5GB and is too big.
sparse: When downsampling segmentation, if true, don't count black pixels when computing
the mode. Useful for e.g. synapses and point labels.
agglomerate: (graphene only) remap the watershed layer to a proofread segmentation.
timestamp: (graphene only) integer UNIX timestamp indicating the proofreading state
to represent.
"""
src_vol = CloudVolume(src_layer_path, mip=mip)
if dest_voxel_offset:
dest_voxel_offset = Vec(*dest_voxel_offset, dtype=int)
else:
dest_voxel_offset = src_vol.voxel_offset.clone()
if factor is None:
factor = (2, 2, 1)
if skip_downsamples:
factor = (1, 1, 1)
if not chunk_size:
chunk_size = src_vol.info['scales'][mip]['chunk_sizes'][0]
chunk_size = Vec(*chunk_size)
try:
dest_vol = CloudVolume(dest_layer_path, mip=mip)
except cloudvolume.exceptions.InfoUnavailableError:
info = copy.deepcopy(src_vol.info)
dest_vol = CloudVolume(dest_layer_path, info=info, mip=mip)
dest_vol.commit_info()
if dest_voxel_offset is not None:
dest_vol.scale["voxel_offset"] = dest_voxel_offset
# If translate is not set, but dest_voxel_offset is then it should naturally be
# only be the difference between datasets.
if translate is None:
translate = dest_vol.voxel_offset - src_vol.voxel_offset # vector pointing from src to dest
else:
translate = Vec(*translate) // src_vol.downsample_ratio
if encoding is not None:
dest_vol.info['scales'][mip]['encoding'] = encoding
if encoding == 'compressed_segmentation' and 'compressed_segmentation_block_size' not in dest_vol.info[
'scales'][mip]:
dest_vol.info['scales'][mip][
'compressed_segmentation_block_size'] = (8, 8, 8)
dest_vol.info['scales'] = dest_vol.info['scales'][:mip + 1]
dest_vol.info['scales'][mip]['chunk_sizes'] = [chunk_size.tolist()]
if clean_info:
dest_vol.info = clean_xfer_info(dest_vol.info)
dest_vol.commit_info()
if shape is None:
if memory_target is not None:
shape = downsample_scales.downsample_shape_from_memory_target(
np.dtype(src_vol.dtype).itemsize, dest_vol.chunk_size.x,
dest_vol.chunk_size.y, dest_vol.chunk_size.z, factor,
memory_target, max_mips)
else:
raise ValueError(
"Either shape or memory_target must be specified.")
shape = Vec(*shape)
if factor[2] == 1:
shape.z = int(dest_vol.chunk_size.z *
round(shape.z / dest_vol.chunk_size.z))
if not skip_downsamples:
downsample_scales.create_downsample_scales(
dest_layer_path,
mip=mip,
ds_shape=shape,
preserve_chunk_size=preserve_chunk_size,
encoding=encoding)
dest_bounds = get_bounds(dest_vol,
bounds,
mip,
bounds_mip=bounds_mip,
chunk_size=chunk_size)
class TransferTaskIterator(FinelyDividedTaskIterator):
def task(self, shape, offset):
return partial(
TransferTask,
src_path=src_layer_path,
dest_path=dest_layer_path,
shape=shape.clone(),
offset=offset.clone(),
fill_missing=fill_missing,
translate=translate,
mip=mip,
skip_downsamples=skip_downsamples,
delete_black_uploads=bool(delete_black_uploads),
background_color=background_color,
agglomerate=agglomerate,
timestamp=timestamp,
compress=compress,
factor=factor,
sparse=sparse,
)
def on_finish(self):
job_details = {
'method': {
'task':
'TransferTask',
'src':
src_layer_path,
'dest':
dest_layer_path,
'shape':
list(map(int, shape)),
'fill_missing':
fill_missing,
'translate':
list(map(int, translate)),
'skip_downsamples':
skip_downsamples,
'delete_black_uploads':
bool(delete_black_uploads),
'background_color':
background_color,
'bounds':
[dest_bounds.minpt.tolist(),
dest_bounds.maxpt.tolist()],
'mip':
mip,
'agglomerate':
bool(agglomerate),
'timestamp':
timestamp,
'compress':
compress,
'encoding':
encoding,
'memory_target':
memory_target,
'factor': (tuple(factor) if factor else None),
'sparse':
bool(sparse),
},
'by': operator_contact(),
'date': strftime('%Y-%m-%d %H:%M %Z'),
}
dest_vol = CloudVolume(dest_layer_path)
dest_vol.provenance.sources = [src_layer_path]
dest_vol.provenance.processing.append(job_details)
dest_vol.commit_provenance()
if not no_src_update and src_vol.meta.path.protocol in ('gs', 's3',
'file'):
src_vol.provenance.processing.append(job_details)
src_vol.commit_provenance()
return TransferTaskIterator(dest_bounds, shape)
def create_contrast_normalization_tasks(src_path,
dest_path,
levels_path=None,
shape=None,
mip=0,
clip_fraction=0.01,
fill_missing=False,
translate=(0, 0, 0),
minval=None,
maxval=None,
bounds=None,
bounds_mip=0):
"""
Use the output of luminence levels to contrast
correct the image by stretching the histogram
to cover the full range of the data type.
"""
srcvol = CloudVolume(src_path, mip=mip)
try:
dvol = CloudVolume(dest_path, mip=mip)
except Exception: # no info file
info = copy.deepcopy(srcvol.info)
dvol = CloudVolume(dest_path, mip=mip, info=info)
dvol.info['scales'] = dvol.info['scales'][:mip + 1]
dvol.commit_info()
if bounds is None:
bounds = srcvol.bounds.clone()
if shape is None:
shape = Bbox((0, 0, 0), (2048, 2048, 64))
shape = shape.shrink_to_chunk_size(dvol.underlying).size3()
shape = Vec.clamp(shape, (1, 1, 1), bounds.size3())
shape = Vec(*shape)
downsample_scales.create_downsample_scales(dest_path,
mip=mip,
ds_shape=shape,
preserve_chunk_size=True)
dvol.refresh_info()
bounds = get_bounds(srcvol, bounds, mip, bounds_mip=bounds_mip)
class ContrastNormalizationTaskIterator(FinelyDividedTaskIterator):
def task(self, shape, offset):
return ContrastNormalizationTask(
src_path=src_path,
dest_path=dest_path,
levels_path=levels_path,
shape=shape.clone(),
offset=offset.clone(),
clip_fraction=clip_fraction,
mip=mip,
fill_missing=fill_missing,
translate=translate,
minval=minval,
maxval=maxval,
)
def on_finish(self):
dvol.provenance.processing.append({
'method': {
'task': 'ContrastNormalizationTask',
'src_path': src_path,
'dest_path': dest_path,
'shape': Vec(*shape).tolist(),
'clip_fraction': clip_fraction,
'mip': mip,
'translate': Vec(*translate).tolist(),
'minval': minval,
'maxval': maxval,
'bounds': [bounds.minpt.tolist(),
bounds.maxpt.tolist()],
},
'by':
operator_contact(),
'date':
strftime('%Y-%m-%d %H:%M %Z'),
})
dvol.commit_provenance()
return ContrastNormalizationTaskIterator(bounds, shape)
def create_downsampling_tasks(layer_path,
mip=0,
fill_missing=False,
axis='z',
num_mips=5,
preserve_chunk_size=True,
sparse=False,
bounds=None,
chunk_size=None,
encoding=None,
delete_black_uploads=False,
background_color=0,
dest_path=None,
compress=None,
factor=None,
bounds_mip=0):
"""
mip: Download this mip level, writes to mip levels greater than this one.
fill_missing: interpret missing chunks as black instead of issuing an EmptyVolumeException
axis: technically 'x' and 'y' are supported, but no one uses them.
num_mips: download a block chunk * 2**num_mips size and generate num_mips mips. If you have
memory problems, try reducing this number.
preserve_chunk_size: if true, maintain chunk size of starting mip, else, find the closest
evenly divisible chunk size to 64,64,64 for this shape and use that. The latter can be
useful when mip 0 uses huge chunks and you want to simply visualize the upper mips.
chunk_size: (overrides preserve_chunk_size) force chunk size for new layers to be this.
sparse: When downsampling segmentation, if true, don't count black pixels when computing
the mode. Useful for e.g. synapses and point labels.
bounds: By default, downsample everything, but you can specify restricted bounding boxes
instead. The bounding box will be expanded to the nearest chunk. Bbox is specifed in mip 0
coordinates.
delete_black_uploads: issue delete commands instead of upload chunks
that are all background.
background_color: Designates which color should be considered background.
dest_path: (optional) instead of writing downsamples to the existing
volume, write them somewhere else. This can be useful e.g. if someone
doesn't want you to touch the existing info file.
compress: None, 'gzip', or 'br' Determines which compression algorithm to use
for new uploaded files.
factor: (overrides axis) can manually specify what each downsampling round is
supposed to do: e.g. (2,2,1), (2,2,2), etc
"""
def ds_shape(mip, chunk_size=None, factor=None):
if chunk_size:
shape = Vec(*chunk_size)
else:
shape = vol.meta.chunk_size(mip)[:3]
if factor is None:
factor = downsample_scales.axis_to_factor(axis)
shape.x *= factor[0]**num_mips
shape.y *= factor[1]**num_mips
shape.z *= factor[2]**num_mips
return shape
vol = CloudVolume(layer_path, mip=mip)
shape = ds_shape(mip, chunk_size, factor)
vol = downsample_scales.create_downsample_scales(
layer_path,
mip,
shape,
preserve_chunk_size=preserve_chunk_size,
chunk_size=chunk_size,
encoding=encoding,
factor=factor)
if not preserve_chunk_size or chunk_size:
shape = ds_shape(mip + 1, chunk_size, factor)
bounds = get_bounds(
vol,
bounds,
mip,
bounds_mip=bounds_mip,
chunk_size=vol.chunk_size,
)
class DownsampleTaskIterator(FinelyDividedTaskIterator):
def task(self, shape, offset):
return partial(
DownsampleTask,
layer_path=layer_path,
mip=vol.mip,
shape=shape.clone(),
offset=offset.clone(),
axis=axis,
fill_missing=fill_missing,
sparse=sparse,
delete_black_uploads=delete_black_uploads,
background_color=background_color,
dest_path=dest_path,
compress=compress,
factor=factor,
)
def on_finish(self):
vol.provenance.processing.append({
'method': {
'task':
'DownsampleTask',
'mip':
mip,
'num_mips':
num_mips,
'shape':
shape.tolist(),
'axis':
axis,
'method':
'downsample_with_averaging' if vol.layer_type == 'image'
else 'downsample_segmentation',
'sparse':
sparse,
'bounds':
str(bounds),
'chunk_size': (list(chunk_size) if chunk_size else None),
'preserve_chunk_size':
preserve_chunk_size,
'encoding':
encoding,
'fill_missing':
bool(fill_missing),
'delete_black_uploads':
bool(delete_black_uploads),
'background_color':
background_color,
'dest_path':
dest_path,
'compress':
compress,
'factor': (tuple(factor) if factor else None),
},
'by':
operator_contact(),
'date':
strftime('%Y-%m-%d %H:%M %Z'),
})
vol.commit_provenance()
return DownsampleTaskIterator(bounds, shape)