def fromBinary(self, dataPath, ext='bin', confFilename='conf.json',
nkeys=None, nvalues=None, keyType=None, valueType=None,
newDtype='smallfloat', casting='safe', maxPartitionSize='32mb'):
"""
Load a Series object from a directory of binary files.
Parameters
----------
dataPath : string URI or local filesystem path
Specifies the directory or files to be loaded. May be formatted as a URI string with scheme (e.g. "file://",
"s3n://". If no scheme is present, will be interpreted as a path on the local filesystem. This path
must be valid on all workers. Datafile may also refer to a single file, or to a range of files specified
by a glob-style expression using a single wildcard character '*'.
newDtype : dtype or dtype specifier or string 'smallfloat' or None, optional, default 'smallfloat'
Numpy dtype of output series data. Most methods expect Series data to be floating-point. Input data will be
cast to the requested `newdtype` if not None - see Data `astype()` method.
casting : 'no'|'equiv'|'safe'|'same_kind'|'unsafe', optional, default 'safe'
Casting method to pass on to numpy's `astype()` method; see numpy documentation for details.
maxPartitionSize : str, optional, default = '32mb'
Maximum size of partitions as Java-style memory, will indirectly control the number of partitions
"""
paramsObj = self.__loadParametersAndDefaults(dataPath, confFilename, nkeys, nvalues, keyType, valueType)
self.__checkBinaryParametersAreSpecified(paramsObj)
dataPath = self.__normalizeDatafilePattern(dataPath, ext)
keyDtype = dtypeFunc(paramsObj.keytype)
valDtype = dtypeFunc(paramsObj.valuetype)
keySize = paramsObj.nkeys * keyDtype.itemsize
recordSize = keySize + paramsObj.nvalues * valDtype.itemsize
from thunder.utils.common import parseMemoryString
if isinstance(maxPartitionSize, basestring):
size = parseMemoryString(maxPartitionSize)
else:
raise Exception("Invalid size specification")
hadoopConf = {'recordLength': str(recordSize), 'mapred.max.split.size': str(size)}
lines = self.sc.newAPIHadoopFile(dataPath, 'thunder.util.io.hadoop.FixedLengthBinaryInputFormat',
'org.apache.hadoop.io.LongWritable',
'org.apache.hadoop.io.BytesWritable',
conf=hadoopConf)
data = lines.map(lambda (_, v):
(tuple(int(x) for x in frombuffer(buffer(v, 0, keySize), dtype=keyDtype)),
frombuffer(buffer(v, keySize), dtype=valDtype)))
return Series(data, dtype=str(valDtype), index=arange(paramsObj.nvalues)).astype(newDtype, casting)
def _calcSplitsForBlockSize(blockSize, elementSize, dims):
from thunder.utils.common import parseMemoryString
import bisect
if isinstance(blockSize, basestring):
blockSize = parseMemoryString(blockSize)
memSeq = _BlockMemoryAsReversedSequence(_normDimsToShapeTuple(dims))
tmpIdx = bisect.bisect_left(memSeq, blockSize / float(elementSize))
if tmpIdx == len(memSeq):
# handle case where requested block is bigger than the biggest image
# we can produce; just give back the biggest block size
tmpIdx -= 1
return memSeq.indToSub(tmpIdx)
def _calcSplitsForBlockSize(blockSize, elementSize, dims):
from thunder.utils.common import parseMemoryString
import bisect
if isinstance(blockSize, basestring):
blockSize = parseMemoryString(blockSize)
memSeq = _BlockMemoryAsReversedSequence(_normDimsToShapeTuple(dims))
tmpIdx = bisect.bisect_left(memSeq, blockSize / float(elementSize))
if tmpIdx == len(memSeq):
# handle case where requested block is bigger than the biggest image
# we can produce; just give back the biggest block size
tmpIdx -= 1
return memSeq.indToSub(tmpIdx)
def _scatterToBlocks(self, blockSize="150M", blocksPerDim=None, groupingDim=None):
if not groupingDim is None:
# get series from blocks defined by pivoting:
gd = self.__validateOrCalcGroupingDim(groupingDim=groupingDim)
blocksdata = self._toBlocksByImagePlanes(groupingDim=gd)
else:
# get series from blocks defined by splits
if not blocksPerDim:
# get splits from requested block size
blockSize = parseMemoryString(blockSize)
blocksPerDim = self.__calcBlocksPerDim(blockSize)
blocksdata = self._toBlocksBySplits(blocksPerDim)
return blocksdata
def _scatterToBlocks(self,
blockSize="150M",
blocksPerDim=None,
groupingDim=None):
if not groupingDim is None:
# get series from blocks defined by pivoting:
gd = self.__validateOrCalcGroupingDim(groupingDim=groupingDim)
blocksdata = self._toBlocksByImagePlanes(groupingDim=gd)
else:
# get series from blocks defined by splits
if not blocksPerDim:
# get splits from requested block size
blockSize = parseMemoryString(blockSize)
blocksPerDim = self.__calcBlocksPerDim(blockSize)
blocksdata = self._toBlocksBySplits(blocksPerDim)
return blocksdata
def _getSeriesBlocksFromMultiTif(self, dataPath, ext="tif", blockSize="150M",
newDtype='smallfloat', casting='safe', startIdx=None, stopIdx=None,
recursive=False):
import thunder.rdds.fileio.multitif as multitif
import itertools
from PIL import Image
import io
dataPath = self.__normalizeDatafilePattern(dataPath, ext)
blockSize = parseMemoryString(blockSize)
reader = getFileReaderForPath(dataPath)(awsCredentialsOverride=self.awsCredentialsOverride)
filenames = reader.list(dataPath, startIdx=startIdx, stopIdx=stopIdx, recursive=recursive)
if not filenames:
raise IOError("No files found for path '%s'" % dataPath)
ntimepoints = len(filenames)
doMinimizeReads = dataPath.lower().startswith("s3")
# check PIL version to see whether it is actually pillow or indeed old PIL and choose
# conversion function appropriately. See ImagesLoader.fromMultipageTif and common.pil_to_array
# for more explanation.
isPillow = hasattr(Image, "PILLOW_VERSION")
if isPillow:
conversionFcn = array # use numpy's array() function
else:
from thunder.utils.common import pil_to_array
conversionFcn = pil_to_array # use our modified version of matplotlib's pil_to_array
height, width, npages, dtype = SeriesLoader.__readMetadataFromFirstPageOfMultiTif(reader, filenames[0])
if dtype.startswith('int'):
raise ValueError('Signed integer tiff images are not supported in SeriesLoader (shuffle=False);' +
' please try loading as Images (shuffle=True)')
pixelBytesize = dtypeFunc(dtype).itemsize
if newDtype is None or str(newDtype) == '':
newDtype = str(dtype)
elif newDtype == 'smallfloat':
newDtype = str(smallestFloatType(dtype))
else:
newDtype = str(newDtype)
# intialize at one block per plane
bytesPerPlane = height * width * pixelBytesize * ntimepoints
bytesPerBlock = bytesPerPlane
blocksPerPlane = 1
# keep dividing while cutting our size in half still leaves us bigger than the requested size
# should end up no more than 2x blockSize.
while bytesPerBlock >= blockSize * 2:
bytesPerBlock /= 2
blocksPerPlane *= 2
blocklenPixels = max((height * width) / blocksPerPlane, 1) # integer division
while blocksPerPlane * blocklenPixels < height * width: # make sure we're reading the plane fully
blocksPerPlane += 1
# prevent bringing in self in closure:
awsCredentialsOverride = self.awsCredentialsOverride
# keys will be planeidx, blockidx:
keys = list(itertools.product(xrange(npages), xrange(blocksPerPlane)))
def readBlockFromTiff(planeIdxBlockIdx):
planeIdx, blockIdx = planeIdxBlockIdx
blocks = []
planeShape = None
blockStart = None
blockEnd = None
for fname in filenames:
reader_ = getFileReaderForPath(fname)(awsCredentialsOverride=awsCredentialsOverride)
fp = reader_.open(fname)
try:
if doMinimizeReads:
# use multitif module to generate a fake, in-memory one-page tif file
# the advantage of this is that it cuts way down on the many small reads
# that PIL/pillow will make otherwise, which would be a problem for s3
tiffParser_ = multitif.TiffParser(fp, debug=False)
tiffFilebuffer = multitif.packSinglePage(tiffParser_, pageIdx=planeIdx)
byteBuf = io.BytesIO(tiffFilebuffer)
try:
pilImg = Image.open(byteBuf)
ary = conversionFcn(pilImg).T
finally:
byteBuf.close()
del tiffFilebuffer, tiffParser_, pilImg, byteBuf
else:
# read tif using PIL directly
pilImg = Image.open(fp)
pilImg.seek(planeIdx)
ary = conversionFcn(pilImg).T
del pilImg
if not planeShape:
planeShape = ary.shape[:]
blockStart = blockIdx * blocklenPixels
blockEnd = min(blockStart+blocklenPixels, planeShape[0]*planeShape[1])
blocks.append(ary.ravel(order='C')[blockStart:blockEnd])
del ary
finally:
fp.close()
buf = vstack(blocks).T # dimensions are now linindex x time (images)
del blocks
buf = buf.astype(newDtype, casting=casting, copy=False)
# append subscript keys based on dimensions
linearIdx = arange(blockStart, blockEnd) # zero-based
seriesKeys = zip(*map(tuple, unravel_index(linearIdx, planeShape, order='C')))
# add plane index to end of keys
if npages > 1:
seriesKeys = [tuple(list(keys_)[::-1]+[planeIdx]) for keys_ in seriesKeys]
else:
seriesKeys = [tuple(list(keys_)[::-1]) for keys_ in seriesKeys]
return zip(seriesKeys, buf)
# map over blocks
rdd = self.sc.parallelize(keys, len(keys)).flatMap(readBlockFromTiff)
if npages > 1:
dims = (npages, width, height)
else:
dims = (width, height)
metadata = (dims, ntimepoints, newDtype)
return rdd, metadata
def _getSeriesBlocksFromStack(self, dataPath, dims, ext="stack", blockSize="150M", dtype='int16',
newDtype='smallfloat', casting='safe', startIdx=None, stopIdx=None, recursive=False):
"""Create an RDD of <string blocklabel, (int k-tuple indices, array of datatype values)>
Parameters
----------
dataPath: string URI or local filesystem path
Specifies the directory or files to be loaded. May be formatted as a URI string with scheme (e.g. "file://",
"s3n://". If no scheme is present, will be interpreted as a path on the local filesystem. This path
must be valid on all workers. Datafile may also refer to a single file, or to a range of files specified
by a glob-style expression using a single wildcard character '*'.
dims: tuple of positive int
Dimensions of input image data, ordered with the fastest-changing dimension first.
dtype: dtype or dtype specifier, optional, default 'int16'
Numpy dtype of input stack data
newDtype: floating-point dtype or dtype specifier or string 'smallfloat' or None, optional, default 'smallfloat'
Numpy dtype of output series data. Series data must be floating-point. Input data will be cast to the
requested `newdtype` - see numpy `astype()` method.
casting: 'no'|'equiv'|'safe'|'same_kind'|'unsafe', optional, default 'safe'
Casting method to pass on to numpy's `astype()` method; see numpy documentation for details.
recursive: boolean, default False
If true, will recursively descend directories rooted at dataPath, loading all files in the tree that
have an extension matching 'ext'. Recursive loading is currently only implemented for local filesystems
(not s3).
Returns
---------
pair of (RDD, ntimepoints)
RDD: sequence of keys, values pairs
(call using flatMap)
RDD Key: tuple of int
zero-based indicies of position within original image volume
RDD Value: numpy array of datatype
series of values at position across loaded image volumes
ntimepoints: int
number of time points in returned series, determined from number of stack files found at dataPath
newDtype: string
string representation of numpy data type of returned blocks
"""
dataPath = self.__normalizeDatafilePattern(dataPath, ext)
blockSize = parseMemoryString(blockSize)
totalDim = reduce(lambda x_, y_: x_*y_, dims)
dtype = dtypeFunc(dtype)
if newDtype is None or newDtype == '':
newDtype = str(dtype)
elif newDtype == 'smallfloat':
newDtype = str(smallestFloatType(dtype))
else:
newDtype = str(newDtype)
reader = getFileReaderForPath(dataPath)(awsCredentialsOverride=self.awsCredentialsOverride)
filenames = reader.list(dataPath, startIdx=startIdx, stopIdx=stopIdx, recursive=recursive)
if not filenames:
raise IOError("No files found for path '%s'" % dataPath)
dataSize = totalDim * len(filenames) * dtype.itemsize
nblocks = max(dataSize / blockSize, 1) # integer division
if len(dims) >= 3:
# for 3D stacks, do calculations to ensure that
# different planes appear in distinct files
blocksPerPlane = max(nblocks / dims[-1], 1)
pixPerPlane = reduce(lambda x_, y_: x_*y_, dims[:-1]) # all but last dimension
# get the greatest number of blocks in a plane (up to as many as requested) that still divide the plane
# evenly. This will always be at least one.
kUpdated = [x for x in range(1, blocksPerPlane+1) if not pixPerPlane % x][-1]
nblocks = kUpdated * dims[-1]
blockSizePerStack = (totalDim / nblocks) * dtype.itemsize
else:
# otherwise just round to make contents divide into nearly even blocks
blockSizePerStack = int(math.ceil(totalDim / float(nblocks)))
nblocks = int(math.ceil(totalDim / float(blockSizePerStack)))
blockSizePerStack *= dtype.itemsize
fileSize = totalDim * dtype.itemsize
def readBlock(blockNum):
# copy size out from closure; will modify later:
blockSizePerStack_ = blockSizePerStack
# get start position for this block
position = blockNum * blockSizePerStack_
# adjust if at end of file
if (position + blockSizePerStack_) > fileSize:
blockSizePerStack_ = int(fileSize - position)
# loop over files, loading one block from each
bufs = []
for fname in filenames:
buf = reader.read(fname, startOffset=position, size=blockSizePerStack_)
bufs.append(frombuffer(buf, dtype=dtype))
buf = vstack(bufs).T # dimensions are now linindex x time (images)
del bufs
buf = buf.astype(newDtype, casting=casting, copy=False)
# append subscript keys based on dimensions
itemPosition = position / dtype.itemsize
itemBlocksize = blockSizePerStack_ / dtype.itemsize
linearIdx = arange(itemPosition, itemPosition + itemBlocksize) # zero-based
keys = zip(*map(tuple, unravel_index(linearIdx, dims, order='F')))
return zip(keys, buf)
# map over blocks
return (self.sc.parallelize(range(0, nblocks), nblocks).flatMap(lambda bn: readBlock(bn)),
len(filenames), newDtype)
def fromBinary(self,
dataPath,
ext='bin',
confFilename='conf.json',
nkeys=None,
nvalues=None,
keyType=None,
valueType=None,
newDtype='smallfloat',
casting='safe',
maxPartitionSize='32mb'):
"""
Load a Series object from a directory of binary files.
Parameters
----------
dataPath : string URI or local filesystem path
Specifies the directory or files to be loaded. May be formatted as a URI string with scheme (e.g. "file://",
"s3n://". If no scheme is present, will be interpreted as a path on the local filesystem. This path
must be valid on all workers. Datafile may also refer to a single file, or to a range of files specified
by a glob-style expression using a single wildcard character '*'.
newDtype : dtype or dtype specifier or string 'smallfloat' or None, optional, default 'smallfloat'
Numpy dtype of output series data. Most methods expect Series data to be floating-point. Input data will be
cast to the requested `newdtype` if not None - see Data `astype()` method.
casting : 'no'|'equiv'|'safe'|'same_kind'|'unsafe', optional, default 'safe'
Casting method to pass on to numpy's `astype()` method; see numpy documentation for details.
maxPartitionSize : str, optional, default = '32mb'
Maximum size of partitions as Java-style memory, will indirectly control the number of partitions
"""
paramsObj = self.__loadParametersAndDefaults(dataPath, confFilename,
nkeys, nvalues, keyType,
valueType)
self.__checkBinaryParametersAreSpecified(paramsObj)
dataPath = self.__normalizeDatafilePattern(dataPath, ext)
keyDtype = dtypeFunc(paramsObj.keytype)
valDtype = dtypeFunc(paramsObj.valuetype)
keySize = paramsObj.nkeys * keyDtype.itemsize
recordSize = keySize + paramsObj.nvalues * valDtype.itemsize
from thunder.utils.common import parseMemoryString
if isinstance(maxPartitionSize, basestring):
size = parseMemoryString(maxPartitionSize)
else:
raise Exception("Invalid size specification")
hadoopConf = {
'recordLength': str(recordSize),
'mapred.max.split.size': str(size)
}
lines = self.sc.newAPIHadoopFile(
dataPath,
'thunder.util.io.hadoop.FixedLengthBinaryInputFormat',
'org.apache.hadoop.io.LongWritable',
'org.apache.hadoop.io.BytesWritable',
conf=hadoopConf)
data = lines.map(lambda (_, v): (tuple(
int(x) for x in frombuffer(buffer(v, 0, keySize), dtype=keyDtype)
), frombuffer(buffer(v, keySize), dtype=valDtype)))
return Series(data,
dtype=str(valDtype),
index=arange(paramsObj.nvalues)).astype(
newDtype, casting)
def _getSeriesBlocksFromMultiTif(self,
datapath,
ext="tif",
blockSize="150M",
newdtype='smallfloat',
casting='safe',
startidx=None,
stopidx=None):
import thunder.rdds.fileio.multitif as multitif
import itertools
from PIL import Image
from thunder.utils.common import pil_to_array, smallest_float_type
import io
datapath = self.__normalizeDatafilePattern(datapath, ext)
blockSize = parseMemoryString(blockSize)
reader = getFileReaderForPath(datapath)()
filenames = reader.list(datapath)
if not filenames:
raise IOError("No files found for path '%s'" % datapath)
filenames = selectByStartAndStopIndices(filenames, startidx, stopidx)
ntimepoints = len(filenames)
minimize_reads = datapath.lower().startswith("s3")
height, width, npages, datatype = SeriesLoader.__readMetadataFromFirstPageOfMultiTif(
reader, filenames[0])
pixelbytesize = dtypefunc(datatype).itemsize
if newdtype is None or str(newdtype) == '':
newdtype = str(datatype)
elif newdtype == 'smallfloat':
newdtype = str(smallest_float_type(datatype))
else:
newdtype = str(newdtype)
# intialize at one block per plane
bytesperplane = height * width * pixelbytesize * ntimepoints
bytesperblock = bytesperplane
blocksperplane = 1
# keep dividing while cutting our size in half still leaves us bigger than the requested size
# should end up no more than 2x blockSize.
while bytesperblock >= blockSize * 2:
bytesperblock /= 2
blocksperplane *= 2
blocklenPixels = max((height * width) / blocksperplane,
1) # integer division
while blocksperplane * blocklenPixels < height * width: # make sure we're reading the plane fully
blocksperplane += 1
# keys will be planeidx, blockidx:
keys = list(itertools.product(xrange(npages), xrange(blocksperplane)))
def readblockfromtif(pidxbidx_):
planeidx, blockidx = pidxbidx_
blocks = []
planeshape = None
blockstart = None
blockend = None
for fname in filenames:
reader_ = getFileReaderForPath(fname)()
fp = reader_.open(fname)
try:
if minimize_reads:
# use multitif module to generate a fake, in-memory one-page tif file
# the advantage of this is that it cuts way down on the many small reads
# that PIL/pillow will make otherwise, which would be a problem for s3
tiffparser_ = multitif.TiffParser(fp, debug=False)
tiffilebuffer = multitif.packSinglePage(
tiffparser_, page_idx=planeidx)
bytebuf = io.BytesIO(tiffilebuffer)
try:
pilimg = Image.open(bytebuf)
ary = pil_to_array(pilimg).T
finally:
bytebuf.close()
del tiffilebuffer, tiffparser_, pilimg, bytebuf
else:
# read tif using PIL directly
pilimg = Image.open(fp)
pilimg.seek(planeidx)
ary = pil_to_array(pilimg).T
del pilimg
if not planeshape:
planeshape = ary.shape[:]
blockstart = blockidx * blocklenPixels
blockend = min(blockstart + blocklenPixels,
planeshape[0] * planeshape[1])
blocks.append(ary.ravel(order='C')[blockstart:blockend])
del ary
finally:
fp.close()
buf = vstack(
blocks).T # dimensions are now linindex x time (images)
del blocks
buf = buf.astype(newdtype, casting=casting, copy=False)
# append subscript keys based on dimensions
linindx = arange(blockstart, blockend) # zero-based
serieskeys = zip(
*map(tuple, unravel_index(linindx, planeshape, order='C')))
# add plane index to end of keys
serieskeys = [
tuple(list(keys_)[::-1] + [planeidx]) for keys_ in serieskeys
]
return zip(serieskeys, buf)
# map over blocks
rdd = self.sc.parallelize(keys, len(keys)).flatMap(readblockfromtif)
dims = (npages, width, height)
metadata = (dims, ntimepoints, newdtype)
return rdd, metadata
def _getSeriesBlocksFromMultiTif(self, datapath, ext="tif", blockSize="150M",
newdtype='smallfloat', casting='safe', startidx=None, stopidx=None):
import thunder.rdds.fileio.multitif as multitif
import itertools
from PIL import Image
import io
datapath = self.__normalizeDatafilePattern(datapath, ext)
blockSize = parseMemoryString(blockSize)
reader = getFileReaderForPath(datapath)()
filenames = reader.list(datapath)
if not filenames:
raise IOError("No files found for path '%s'" % datapath)
filenames = selectByStartAndStopIndices(filenames, startidx, stopidx)
ntimepoints = len(filenames)
minimize_reads = datapath.lower().startswith("s3")
height, width, npages, datatype = SeriesLoader.__readMetadataFromFirstPageOfMultiTif(reader, filenames[0])
pixelbytesize = dtypefunc(datatype).itemsize
if newdtype is None or str(newdtype) == '':
newdtype = str(datatype)
elif newdtype == 'smallfloat':
newdtype = str(smallest_float_type(datatype))
else:
newdtype = str(newdtype)
# intialize at one block per plane
bytesperplane = height * width * pixelbytesize * ntimepoints
bytesperblock = bytesperplane
blocksperplane = 1
# keep dividing while cutting our size in half still leaves us bigger than the requested size
# should end up no more than 2x blockSize.
while bytesperblock >= blockSize * 2:
bytesperblock /= 2
blocksperplane *= 2
blocklenPixels = max((height * width) / blocksperplane, 1) # integer division
while blocksperplane * blocklenPixels < height * width: # make sure we're reading the plane fully
blocksperplane += 1
# keys will be planeidx, blockidx:
keys = list(itertools.product(xrange(npages), xrange(blocksperplane)))
def readblockfromtif(pidxbidx_):
planeidx, blockidx = pidxbidx_
blocks = []
planeshape = None
blockstart = None
blockend = None
for fname in filenames:
reader_ = getFileReaderForPath(fname)()
fp = reader_.open(fname)
try:
if minimize_reads:
# use multitif module to generate a fake, in-memory one-page tif file
# the advantage of this is that it cuts way down on the many small reads
# that PIL/pillow will make otherwise, which would be a problem for s3
tiffparser_ = multitif.TiffParser(fp, debug=False)
tiffilebuffer = multitif.packSinglePage(tiffparser_, page_idx=planeidx)
bytebuf = io.BytesIO(tiffilebuffer)
try:
pilimg = Image.open(bytebuf)
ary = array(pilimg).T
finally:
bytebuf.close()
del tiffilebuffer, tiffparser_, pilimg, bytebuf
else:
# read tif using PIL directly
pilimg = Image.open(fp)
pilimg.seek(planeidx)
ary = array(pilimg).T
del pilimg
if not planeshape:
planeshape = ary.shape[:]
blockstart = blockidx * blocklenPixels
blockend = min(blockstart+blocklenPixels, planeshape[0]*planeshape[1])
blocks.append(ary.ravel(order='C')[blockstart:blockend])
del ary
finally:
fp.close()
buf = vstack(blocks).T # dimensions are now linindex x time (images)
del blocks
buf = buf.astype(newdtype, casting=casting, copy=False)
# append subscript keys based on dimensions
linindx = arange(blockstart, blockend) # zero-based
serieskeys = zip(*map(tuple, unravel_index(linindx, planeshape, order='C')))
# add plane index to end of keys
serieskeys = [tuple(list(keys_)[::-1]+[planeidx]) for keys_ in serieskeys]
return zip(serieskeys, buf)
# map over blocks
rdd = self.sc.parallelize(keys, len(keys)).flatMap(readblockfromtif)
dims = (npages, width, height)
metadata = (dims, ntimepoints, newdtype)
return rdd, metadata
