def main(args):
starttime = Timer.starttimer()
#Cache thrashing is common when working with large files
# we help alleviate misses by setting a larger than normal cache. 1GB
gdal.SetCacheMax(1073741824)
#Get stretch type
stretch = OptParse.argget_stretch(args)
#Get some info about the machine for mp
cores = args['ncores']
if cores is None:
cores = mp.cpu_count()
#Load the input dataset using the GdalIO class and get / set the output datatype.
dataset = OpenDataSet(args['input'])
raster = dataset.load()
xsize, ysize, nbands, projection, geotransform = dataset.info(raster)
#Get band information
bands = [raster.GetRasterBand(b) for b in range(1, nbands + 1)]
bandstats = [Stats.get_band_stats(b) for b in bands]
b = bands[0]
banddtype = b.DataType
blocksize = b.GetBlockSize()
xblocksize = blocksize[0]
yblocksize = blocksize[1]
output = create_output(args['outputformat'], args['output'], xsize, ysize,
len(bands), projection, geotransform,
gdal.GetDataTypeByName(args['dtype']))
#Intelligently segment the image based upon number of cores and intrinsic block size
if args['byline'] is True:
segments = segment_image(xsize, ysize, 1, ysize)
args['statsper'] = True
elif args['bycolumn'] is True:
segments = segment_image(xsize, ysize, xsize, 1)
args['statsper'] = True
elif args['horizontal_segments'] is not None or args[
'vertical_segments'] is not None:
#The user is defining the segmentation
segments = segment_image(xsize, ysize, args['vertical_segments'],
args['horizontal_segments'])
else:
segments = [(0, 0, xsize, ysize)]
carray_dtype = _gdal_to_ctypes[banddtype]
#Preallocate a sharedmem array of the correct size
ctypesxsize, ctypesysize = segments[0][2:]
if args['byline'] is True:
ctypesysize = cores
elif args['bycolumn'] is True:
ctypesxsize = cores
carray = mp.RawArray(carray_dtype, ctypesxsize * ctypesysize)
glb.sharedarray = np.frombuffer(carray,
dtype=_gdal_to_numpy[banddtype]).reshape(
ctypesysize, ctypesxsize)
pool = mp.Pool(processes=cores,
initializer=glb.init,
initargs=(glb.sharedarray, ))
#A conscious decision to iterate over the bands in serial - a IO bottleneck anyway
for j, band in enumerate(bands):
stats = bandstats[j]
bandmin = stats['minimum']
bandmax = stats['maximum']
ndv = stats['ndv']
userndv = args['ndv']
args.update(stats)
if args['byline'] is True:
for y in range(0, ysize, cores):
xstart, ystart, intervalx, intervaly = 0, y, xsize, cores
if ystart + intervaly > ysize:
intervaly = ysize - ystart
#print ystart, ystart + intervaly
#print y, ystart, ystart+ intervaly, intervaly
glb.sharedarray[:intervaly, :intervalx] = band.ReadAsArray(
xstart, ystart, intervalx, intervaly)
#If the input has an NDV - mask it.
if stats['ndv'] != None:
glb.sharedarray = np.ma.masked_equal(glb.sharedarray,
stats['ndv'],
copy=False)
mask = np.ma.getmask(glb.sharedarray)
#if args['statsper'] is True:
#args.update(Stats.get_array_stats(glb.sharedarray, stretch))
for i in range(cores):
res = pool.apply(stretch, args=(slice(i, i + 1), args))
if args['ndv'] != None:
glb.sharedarray[glb.sharedarray == ndv] = args['ndv']
output.GetRasterBand(j + 1).SetNoDataValue(float(userndv))
if args['scale'] is not None:
#Scale the data before writing to disk
scale(args['scale'][0], args['scale'][1], bandmin, bandmax)
output.GetRasterBand(j + 1).WriteArray(
glb.sharedarray[:intervaly, :intervalx], xstart, ystart)
if args['quiet']:
print "Processed {} or {} lines \r".format(y, ysize),
sys.stdout.flush()
elif args['bycolumn'] is True:
for x in range(0, xsize, cores):
xstart, ystart, intervalx, intervaly = x, 0, cores, ysize
if xstart + intervalx > xsize:
intervalx = xsize - xstart
glb.sharedarray[:intervaly, :intervalx] = band.ReadAsArray(
xstart, ystart, intervalx, intervaly)
#If the input has an NDV - mask it.
if stats['ndv'] != None:
glb.sharedarray = np.ma.masked_equal(glb.sharedarray,
stats['ndv'],
copy=False)
mask = np.ma.getmask(glb.sharedarray)
if args['statsper'] is True:
args.update(Stats.get_array_stats(glb.sharedarray,
stretch))
for i in range(cores):
res = pool.apply(stretch, args=(slice(i, i + 1), args))
if args['ndv'] != None:
glb.sharedarray[glb.sharedarray == ndv] = args['ndv']
output.GetRasterBand(j + 1).SetNoDataValue(float(userndv))
if args['scale'] is not None:
scale(args['scale'][0], args['scale'][1], bandmin, bandmax)
output.GetRasterBand(j + 1).WriteArray(
glb.sharedarray[:intervaly, :intervalx], xstart, ystart)
if args['quiet']:
print "Processed {} or {} lines \r".format(x, xsize),
sys.stdout.flush()
#If not processing line by line, distirbuted the block over availabel cores
else:
for i, chunk in enumerate(segments):
xstart, ystart, intervalx, intervaly = chunk
#Read the array into the buffer
glb.sharedarray[:intervaly, :intervalx] = band.ReadAsArray(
xstart, ystart, intervalx, intervaly)
#If the input has an NDV - mask it.
if stats['ndv'] != None:
glb.sharedarray = np.ma.masked_equal(glb.sharedarray,
stats['ndv'],
copy=False)
mask = np.ma.getmask(glb.sharedarray)
if args['statsper'] is True:
args.update(Stats.get_array_stats(glb.sharedarray,
stretch))
#Determine the decomposition for each core
step = intervaly // cores
starts = range(0, intervaly + 1, step)
stops = starts[1:]
stops.append(intervaly + 1)
offsets = zip(starts, stops)
for o in offsets:
res = pool.apply(stretch, args=(slice(o[0], o[1]), args))
if args['ndv'] != None:
glb.sharedarray[glb.sharedarray == ndv] = args['ndv']
output.GetRasterBand(j + 1).SetNoDataValue(float(userndv))
if args['scale'] is not None:
#Scale the data before writing to disk
scale(args['scale'][0], args['scale'][1], bandmin, bandmax)
output.GetRasterBand(j + 1).WriteArray(
glb.sharedarray[:intervaly, :intervalx], xstart, ystart)
Timer.totaltime(starttime)
#Close up
dataset = None
output = None
pool.close()
pool.join()
def main(args):
starttime = Timer.starttimer()
#Cache thrashing is common when working with large files
# we help alleviate misses by setting a larger than normal cache. 1GB
gdal.SetCacheMax(1073741824)
#Get stretch type
stretch = OptParse.argget_stretch(args)
#Get some info about the machine for mp
cores = args['ncores']
if cores is None:
cores = mp.cpu_count()
print "Loading the input dataset..."
#Load the input dataset using the GdalIO class and get / set the output datatype.
dataset = OpenDataSet(args['input'])
raster = dataset.load()
xsize, ysize, nbands, projection, geotransform = dataset.info(raster)
#Get band information
print "Computing band statistics..."
bands = [raster.GetRasterBand(b) for b in range(1, nbands + 1)]
bandstats = [Stats.get_band_stats(b) for b in bands]
b = bands[0]
banddtype = b.DataType
blocksize = b.GetBlockSize()
xblocksize = blocksize[0]
yblocksize = blocksize[1]
print "Creating output file with correct geotransformation (where applicable)..."
output = create_output(args['outputformat'],args['output'],
xsize, ysize, len(bands), projection,
geotransform, gdal.GetDataTypeByName(args['dtype']))
print "Computing image block offsets..."
#Intelligently segment the image based upon number of cores and intrinsic block size
if args['byline'] is True:
segments = segment_image(xsize, ysize, 1, ysize)
args['statsper'] = True
elif args['bycolumn'] is True:
segments = segment_image(xsize, ysize, xsize, 1)
args['statsper'] = True
elif args['horizontal_segments'] is not None or args['vertical_segments'] is not None:
#The user is defining the segmentation
segments = segment_image(xsize, ysize, args['vertical_segments'],args['horizontal_segments'])
else:
segments = [(0,0,xsize, ysize)]
carray_dtype = _gdal_to_ctypes[banddtype]
print "Allocating a shared memory space for processing..."
#Preallocate a sharedmem array of the correct size
ctypesxsize, ctypesysize= segments[0][2:]
if args['byline'] is True:
ctypesysize = cores
elif args['bycolumn'] is True:
ctypesxsize = cores
carray = mp.RawArray(carray_dtype, ctypesxsize * ctypesysize)
glb.sharedarray = np.frombuffer(carray,dtype=_gdal_to_numpy[banddtype]).reshape(ctypesysize, ctypesxsize)
pool = mp.Pool(processes=cores, initializer=glb.init, initargs=(glb.sharedarray, ))
#A conscious decision to iterate over the bands in serial - a IO bottleneck anyway
for j,band in enumerate(bands):
stats = bandstats[j]
args.update(stats)
if args['byline'] is True:
for y in range(0, ysize, cores):
xstart, ystart, intervalx, intervaly = 0, y, xsize, cores
if ystart + intervaly > ysize:
intervaly = ysize - ystart
#print ystart, ystart + intervaly
#print y, ystart, ystart+ intervaly, intervaly
glb.sharedarray[:intervaly, :intervalx] = band.ReadAsArray(xstart, ystart, intervalx, intervaly)
#If the input has an NDV - mask it.
if stats['ndv'] != None:
glb.sharedarray = np.ma.masked_equal(glb.sharedarray, stats['ndv'], copy=False)
mask = np.ma.getmask(glb.sharedarray)
#if args['statsper'] is True:
#args.update(Stats.get_array_stats(glb.sharedarray, stretch))
for i in range(cores):
res = pool.apply(stretch, args=(slice(i, i+1), args))
if args['ndv'] != None:
#glb.sharedarray[mask] = args['ndv']
output.GetRasterBand(j+1).SetNoDataValue(float(args['ndv']))
output.GetRasterBand(j+1).WriteArray(glb.sharedarray[:intervaly, :intervalx], xstart,ystart)
if args['quiet']:
print "Processed {} or {} lines \r".format(y, ysize),
sys.stdout.flush()
elif args['bycolumn'] is True:
for x in range(0, xsize, cores):
xstart, ystart, intervalx, intervaly = x, 0, cores, ysize
if xstart + intervalx > xsize:
intervalx = xsize - xstart
glb.sharedarray[:intervaly, :intervalx] = band.ReadAsArray(xstart, ystart, intervalx, intervaly)
#If the input has an NDV - mask it.
if stats['ndv'] != None:
glb.sharedarray = np.ma.masked_equal(glb.sharedarray, stats['ndv'], copy=False)
mask = np.ma.getmask(glb.sharedarray)
if args['statsper'] is True:
args.update(Stats.get_array_stats(glb.sharedarray, stretch))
for i in range(cores):
res = pool.apply(stretch, args=(slice(i, i+1), args))
if args['ndv'] != None:
glb.sharedarray[mask] = args['ndv']
output.GetRasterBand(j+1).SetNoDataValue(float(args['ndv']))
output.GetRasterBand(j+1).WriteArray(glb.sharedarray[:intervaly, :intervalx], xstart,ystart)
if args['quiet']:
print "Processed {} or {} lines \r".format(x, xsize),
sys.stdout.flush()
#If not processing line by line, distirbuted the block over availabel cores
else:
for i, chunk in enumerate(segments):
print i, len(segments)
xstart, ystart, intervalx, intervaly = chunk
#Read the array into the buffer
bandslice = band.ReadAsArray(xstart, ystart, intervalx, intervaly)
glb.sharedarray[:intervaly, :intervalx] = bandslice
#If the input has an NDV - mask it.
if stats['ndv'] != None:
if len(np.where(bandslice != stats['ndv'])[0]) == 0:
continue
glb.sharedarray = np.ma.masked_equal(glb.sharedarray, stats['ndv'], copy=False)
mask = np.ma.getmask(glb.sharedarray)
if args['statsper'] is True:
args.update(Stats.get_array_stats(glb.sharedarray, stretch))
#Determine the decomposition for each core
step = intervaly // cores
if step == 0:
step = 1
starts = range(0, intervaly+1, step)
stops = starts[1:]
stops.append(intervaly+1)
offsets = zip(starts, stops)
for o in offsets:
res = pool.apply(stretch, args=(slice(o[0], o[1]), args))
if args['ndv'] != None:
glb.sharedarray[mask] = args['ndv']
output.GetRasterBand(j+1).SetNoDataValue(float(args['ndv']))
output.GetRasterBand(j+1).WriteArray(glb.sharedarray[:intervaly, :intervalx], xstart,ystart)
Timer.totaltime(starttime)
#Close up
dataset = None
output = None
pool.close()
pool.join()
def main(options, args):
starttime = Timer.starttimer()
#Cache thrashing is common when working with large files, we help alleviate misses by setting a larger than normal cache. 1GB
gdal.SetCacheMax(1073741824)
#Check for input
if not args:
print "\nERROR: You must supply an input data set.\n"
sys.exit(0)
#Get stretch type
stretch = OptParse.get_stretch(options)
#Get some info about the machine for multiprocessing
cores = multiprocessing.cpu_count()
cores *= 2
print "Processing on %i cores." %cores
#Load the input dataset using the GdalIO class and get / set the output datatype.
dataset = GdalIO.GdalIO(args[0])
raster = dataset.load()
#Default is none, unless user specified
if options['dtype'] == None:
dtype = gdal.GetDataTypeName(raster.GetRasterBand(1).DataType)
else:
dtype=options['dtype']
#Create an output if the stretch is written to disk
xsize, ysize, bands, projection, geotransform = dataset.info(raster)
output = dataset.create_output("",options['output'],xsize,ysize,bands,projection, geotransform, gdal.GetDataTypeByName(dtype))
#Segment the image to handle either RAM constraints or selective processing
segments = Segment.segment_image(xsize,ysize,options['vint'], options['hint'])
for b in xrange(bands):
band = raster.GetRasterBand(b+1)
bandstats = Stats.get_band_stats(band)
for key in bandstats.iterkeys():
options[key] = bandstats[key]
#Get the size of the segments to be manipulated
piecenumber = 1
for chunk in segments:
print "Image segmented. Processing segment %i of %i" %(piecenumber, len(segments))
piecenumber += 1
(xstart, ystart, intervalx, intervaly) = chunk
array = band.ReadAsArray(xstart, ystart, intervalx, intervaly).astype(numpy.float32)
if options['ndv_band'] != None:
array = numpy.ma.masked_values(array, options['ndv_band'], copy=False)
elif options['ndv'] != None:
array = numpy.ma.masked_values(array, options['ndv'], copy=False)
if 'stretch' in stretch.__name__:
array = Stats.normalize(array, options['bandmin'], options['bandmax'], dtype)
#If the user wants to calc stats per segment:
if options['segment'] == True:
stats = Stats.get_array_stats(array, stretch)
for key in stats.iterkeys():
options[key] = stats[key]
#Otherwise use the stats per band for each segment
else:
options['mean'] = options['bandmean']
options['maximum'] = options['bandmax']
options['minimum'] = options['bandmin']
options['standard_deviation'] = options['bandstd']
y,x = array.shape
#Calculate the hist and cdf if we need it. This way we do not calc it per core.
if options['histequ_stretch'] == True:
cdf, bins = Stats.gethist_cdf(array,options['num_bins'])
options['cdf'] = cdf
options['bins'] = bins
#Fill the masked values with NaN to get to a shared array
if options['ndv'] != None:
array = array.filled(numpy.nan)
#Create an ctypes array
init(ArrayConvert.SharedMemArray(array))
step = y // cores
jobs = []
if step != 0:
for i in range(0,y,step):
p = multiprocessing.Process(target=stretch,args= (shared_arr,slice(i, i+step)),kwargs=options)
jobs.append(p)
for job in jobs:
job.start()
del job
for job in jobs:
job.join()
del job
#Return the array to the proper data range and write it out. Scale if that is what the user wants
if options['histequ_stretch'] or options['gamma_stretch']== True:
pass
elif 'filter' in stretch.__name__:
pass
else:
Stats.denorm(shared_arr.asarray(), dtype, kwargs=options)
if options['scale'] != None:
Stats.scale(shared_arr.asarray(), kwargs=options)
#If their are NaN in the array replace them with the dataset no data value
Stats.setnodata(shared_arr, options['ndv'])
#Write the output
output.GetRasterBand(b+1).WriteArray(shared_arr.asarray(), xstart,ystart)
#Manually cleanup to stop memory leaks.
del array, jobs, shared_arr.data
try:
del stats
except:
pass
del globals()['shared_arr']
gc.collect()
if options['ndv'] != None:
output.GetRasterBand(b+1).SetNoDataValue(float(options['ndv']))
elif options['ndv_band'] != None:
output.GetRasterBand(b+1).SetNoDataValue(float(options['ndv_band']))
if options['visualize'] == True:
Plot.show_hist(shared_arr.asarray())
Timer.totaltime(starttime)
#Close up
dataset = None
output = None
gc.collect()
def main(options, args):
starttime = Timer.starttimer()
#Cache thrashing is common when working with large files, we help alleviate misses by setting a larger than normal cache. 1GB
gdal.SetCacheMax(1073741824)
#Check for input
if not args:
print "\nERROR: You must supply an input data set.\n"
sys.exit(0)
#Get stretch type
stretch = OptParse.get_stretch(options)
#Get some info about the machine for multiprocessing
cores = multiprocessing.cpu_count()
cores *= 2
print "Processing on %i cores." % cores
#Load the input dataset using the GdalIO class and get / set the output datatype.
dataset = GdalIO.GdalIO(args[0])
raster = dataset.load()
#Default is none, unless user specified
if options['dtype'] == None:
dtype = gdal.GetDataTypeName(raster.GetRasterBand(1).DataType)
else:
dtype = options['dtype']
#Create an output if the stretch is written to disk
xsize, ysize, bands, projection, geotransform = dataset.info(raster)
output = dataset.create_output("", options['output'], xsize, ysize, bands,
projection, geotransform,
gdal.GetDataTypeByName(dtype))
#Segment the image to handle either RAM constraints or selective processing
segments = Segment.segment_image(xsize, ysize, options['vint'],
options['hint'])
for b in xrange(bands):
band = raster.GetRasterBand(b + 1)
bandstats = Stats.get_band_stats(band)
for key in bandstats.iterkeys():
options[key] = bandstats[key]
#Get the size of the segments to be manipulated
piecenumber = 1
for chunk in segments:
print "Image segmented. Processing segment %i of %i" % (
piecenumber, len(segments))
piecenumber += 1
(xstart, ystart, intervalx, intervaly) = chunk
array = band.ReadAsArray(xstart, ystart, intervalx,
intervaly).astype(numpy.float32)
if options['ndv_band'] != None:
array = numpy.ma.masked_values(array,
options['ndv_band'],
copy=False)
elif options['ndv'] != None:
array = numpy.ma.masked_values(array,
options['ndv'],
copy=False)
if 'stretch' in stretch.__name__:
array = Stats.normalize(array, options['bandmin'],
options['bandmax'], dtype)
#If the user wants to calc stats per segment:
if options['segment'] == True:
stats = Stats.get_array_stats(array, stretch)
for key in stats.iterkeys():
options[key] = stats[key]
#Otherwise use the stats per band for each segment
else:
options['mean'] = options['bandmean']
options['maximum'] = options['bandmax']
options['minimum'] = options['bandmin']
options['standard_deviation'] = options['bandstd']
y, x = array.shape
#Calculate the hist and cdf if we need it. This way we do not calc it per core.
if options['histequ_stretch'] == True:
cdf, bins = Stats.gethist_cdf(array, options['num_bins'])
options['cdf'] = cdf
options['bins'] = bins
#Fill the masked values with NaN to get to a shared array
if options['ndv'] != None:
array = array.filled(numpy.nan)
#Create an ctypes array
init(ArrayConvert.SharedMemArray(array))
step = y // cores
jobs = []
if step != 0:
for i in range(0, y, step):
p = multiprocessing.Process(target=stretch,
args=(shared_arr,
slice(i, i + step)),
kwargs=options)
jobs.append(p)
for job in jobs:
job.start()
del job
for job in jobs:
job.join()
del job
#Return the array to the proper data range and write it out. Scale if that is what the user wants
if options['histequ_stretch'] or options['gamma_stretch'] == True:
pass
elif 'filter' in stretch.__name__:
pass
else:
Stats.denorm(shared_arr.asarray(), dtype, kwargs=options)
if options['scale'] != None:
Stats.scale(shared_arr.asarray(), kwargs=options)
#If their are NaN in the array replace them with the dataset no data value
Stats.setnodata(shared_arr, options['ndv'])
#Write the output
output.GetRasterBand(b + 1).WriteArray(shared_arr.asarray(),
xstart, ystart)
#Manually cleanup to stop memory leaks.
del array, jobs, shared_arr.data
try:
del stats
except:
pass
del globals()['shared_arr']
gc.collect()
if options['ndv'] != None:
output.GetRasterBand(b + 1).SetNoDataValue(
float(options['ndv']))
elif options['ndv_band'] != None:
output.GetRasterBand(b + 1).SetNoDataValue(
float(options['ndv_band']))
if options['visualize'] == True:
Plot.show_hist(shared_arr.asarray())
Timer.totaltime(starttime)
#Close up
dataset = None
output = None
gc.collect()
