def computeSampleDistribution(self, points):
''' compute sample frequency distributions histogram for NOMINAL/ORDINAL or pdf for INTERVAL/RATIO
'''
print 'computeSampleDistribution() called'
xmin = self.min
xmax = self.max
#print xmin, xmax
vals = util.extractCovariatesAtPoints([self], points)[0]
if self.__measurement_level in [conf.MSR_LEVEL_NOMINAL, conf.MSR_LEVEL_ORDINAL]:
y1, x = np.histogram(vals, range = (xmin-0.5, xmax+0.5), bins = int(xmax-xmin)+1, density = True)
#print np.unique(vals)
#print x, y1
if np.std(points.weights) != 0: # unequal weights
y2, x = np.histogram(vals, weights = points.weights, bins = int(xmax-xmin)+1, density = True)
else:
y2 = np.copy(y1)
else:
x = np.linspace(xmin, xmax, conf.N_INTERVALS)
gkde = gaussian_kde.gaussian_kde(vals)
y1 = gkde.evaluate(x)
if np.std(points.weights) != 0: # unequal weights
gkde = gaussian_kde.gaussian_kde(vals, weights=points.weights)
y2 = gkde.evaluate(x)
else:
y2 = np.copy(y1)
self.density_sample = y1
self.density_sample_weighted = y2
def __simLoc2SamplesV0(self, loc_ev, evs, SD_evs):
''' compute similarity between a location to N samples
return: a vector of similarity values, each to a sample
'''
try:
sample_evs = util.extractCovariatesAtPoints(self.__envrasters, self.__soilsamples)
sample_evs = np.array(sample_evs).T
# number of samples
N = np.shape(sample_evs)[0]
# similarity btw a loc to N samples
sim = np.zeros(N)
# compute similarities
t0 = time.time()
for i in range(N):
sim[i] = self.__simLoc2SampleV0(loc_ev, sample_evs[i], evs, SD_evs)
conf.TIME_KEEPING_DICT['parts']['compute'].append(time.time()-t0)
return sim
except Exception as e:
raise
def predict_opencl_atom(self, X = None, predict_class = False, single_cpu = conf.SINGLE_CPU, opencl_config = conf.OPENCL_CONFIG):
''' PyOpenCL implementation of the iPSM approach
return: a vector of predictions, eacn for a row in X
'''
print 'predict_opencl_atom() was called'
try:
t0 = time.time()
##### prepare data
# covariates values over the whole study area
r_evs = np.int32(self.__envrasters[0].getData().size)
c_evs = np.int32(len(self.__envrasters))
Std_evs = np.zeros(len(self.__envrasters))
AVG_evs = np.zeros(len(self.__envrasters))
for i in range(len(self.__envrasters)):
Std_evs[i] = self.__envrasters[i].std
AVG_evs[i] = self.__envrasters[i].mean
SD_evs = Std_evs.reshape(c_evs).astype(np.float32)
# covariates values at prediction locations
if X is None: # if X is not provided, make prediction for the whole study area
X = []
for raster in self.__envrasters:
X.append(raster.getData())
X = np.array(X).T
r, c = np.shape(X)
nrows_X = np.int32(r)
ncols_X = np.int32(c)
X = X.reshape(nrows_X*ncols_X).astype(np.float32)
MSRLEVES = self.__msrInts.reshape(c_evs).astype(np.int32)
#print MSRLEVES, MSRLEVES.shape
if not self.__samples_stats_collected:
# covariates values at sample locations
if self.__soilsamples.covariates_at_points is None:
samples_X = util.extractCovariatesAtPoints(self.__envrasters, self.__soilsamples)
else:
samples_X = self.__soilsamples.covariates_at_points.T#[0:c_evs].T ## prone to bug
nrows_samples = np.int32(samples_X.shape[1])
self.__nrows_samples = nrows_samples
samples_SD_evs = np.zeros((nrows_samples, c_evs))
for i in range(nrows_samples):
delta = samples_X[:,i].T - AVG_evs
tmp = Std_evs**2 + delta**2
samples_SD_evs[i] = np.sqrt(tmp)
#print '\nsamples_SD_evs:', samples_SD_evs, '\n'
self.__samples_SD_evs = np.array(samples_SD_evs).reshape(nrows_samples*c_evs).astype(np.float32)
self.__samples_X = np.array(samples_X).T.reshape(nrows_samples*c_evs).astype(np.float32)
#print 'samples_X:', samples_X.shape, samples_X.min()
# sample weights
self.__sample_weights = self.__soilsamples.weights.reshape(nrows_samples).astype(np.float32)
#print 'sample_weights:', sample_weights.shape, sample_weights.min()
# sample attributes
self.__sample_attributes = self.__soilsamples.attributes.reshape(nrows_samples).astype(np.float32)
self.__samples_stats_collected = True
# hold predictions for instances in X
X_predictions = np.zeros(nrows_X).astype(np.float32)
# hold prediction uncertainties for instances in X
X_uncertainties = np.zeros(nrows_X).astype(np.float32)
print 'preparation on HOST took', time.time() - t0, 's'
##### config computing platform and device
for platform in cl.get_platforms():
#print platform.name
if platform.name == conf.OPENCL_CONFIG['Platform']:
PLATFORM = platform
# Print each device per-platform
for device in platform.get_devices():
#print device.name
if device.name == conf.OPENCL_CONFIG['Device']:
DEVICE = device
break
print DEVICE.name, 'on', PLATFORM.name
# opencl context
ctx = cl.Context([DEVICE])
# opencl command queue
queue = cl.CommandQueue(ctx)
##### allocate memory space on device
mf = cl.mem_flags
t0 = time.time()
#evs_g = cl.Buffer(ctx, mf.READ_ONLY | mf.COPY_HOST_PTR, hostbuf=evs)
SD_evs_g = cl.Buffer(ctx, mf.READ_ONLY | mf.COPY_HOST_PTR, hostbuf=SD_evs)
X_g = cl.Buffer(ctx, mf.READ_ONLY | mf.COPY_HOST_PTR, hostbuf=X)
MSRLEVES_g = cl.Buffer(ctx, mf.READ_ONLY | mf.COPY_HOST_PTR, hostbuf=MSRLEVES)
sample_X_g = cl.Buffer(ctx, mf.READ_ONLY | mf.COPY_HOST_PTR, hostbuf=self.__samples_X)
## added 09/06/2017
samples_SD_evs_g = cl.Buffer(ctx, mf.READ_ONLY | mf.COPY_HOST_PTR, hostbuf=self.__samples_SD_evs)
sample_weights_g = cl.Buffer(ctx, mf.READ_ONLY | mf.COPY_HOST_PTR, hostbuf=self.__sample_weights)
sample_attributes_g = cl.Buffer(ctx, mf.READ_ONLY | mf.COPY_HOST_PTR, hostbuf=self.__sample_attributes)
X_predictions_g = cl.Buffer(ctx, mf.WRITE_ONLY, X_predictions.nbytes)
X_uncertainties_g = cl.Buffer(ctx, mf.WRITE_ONLY, X_uncertainties.nbytes)
queue.finish()
t1 = time.time()
conf.TIME_KEEPING_DICT['parts']['data_transfer'].append(t1-t0)
print 'allocation and copy from HOST to DEVICE took', t1 - t0, 's'
X = None
##### build opencl kernel from code in the file
f = open(conf.iPSM_KERNEL_FN, 'r')
fstr = "".join(f.readlines())
fstr = fstr.replace("#define N_SAMPLES 100", "#define N_SAMPLES " + str(self.__nrows_samples))
prg = cl.Program(ctx, fstr).build()
##### opencl computation
threshold = np.float32(self.__uncthreshold)
if predict_class:
mode = np.int32(1)
else:
mode = np.int32(0)
print X_predictions.shape
if not single_cpu:
t0 = time.time()
completeEvent = \
prg.iPSM_Predict(queue, X_predictions.shape, None, nrows_X, ncols_X, self.__nrows_samples, mode, \
threshold, MSRLEVES_g, samples_SD_evs_g, SD_evs_g, X_g, sample_X_g, sample_weights_g, sample_attributes_g, \
X_predictions_g, X_uncertainties_g)
queue.finish()
t1 = time.time()
conf.TIME_KEEPING_DICT['parts']['compute'].append(t1-t0)
print 'kernel took', t1 - t0, 's'
#print queue.finish()
else:
print 'SINGLE_CPU iPSM.predict_opencl() called'
t0 = time.time()
completeEvent = \
prg.iPSM_Predict_Sequential(queue, (1,), (1,), nrows_X, ncols_X, self.__nrows_samples, mode, \
threshold, MSRLEVES_g, samples_SD_evs_g, SD_evs_g, X_g, sample_X_g, sample_weights_g, sample_attributes_g, \
X_predictions_g, X_uncertainties_g)
queue.finish()
t1 = time.time()
conf.TIME_KEEPING_DICT['parts']['compute'].append(t1-t0)
print 'kernel took', t1 - t0, 's'
#print queue.finish()
#### wait until completions
events = [completeEvent]
queue.finish()
print 'up to events finished kernel took', time.time() - t0, 's'
#print queue.finish()
##### copy result data
t0 = time.time()
cl.enqueue_copy(queue, X_predictions, X_predictions_g, wait_for = events)#.wait()
#print queue.finish()
cl.enqueue_copy(queue, X_uncertainties, X_uncertainties_g)
queue.finish()
t1 = time.time()
conf.TIME_KEEPING_DICT['parts']['data_transfer'].append(t1-t0)
print 'copy from DEVICE to HOST took', t1 - t0, 's'
y = np.vstack((X_predictions, X_uncertainties)).T
#print y
return y
except Exception as e:
raise
def __simLocs2Samples(self, X, parallel = True, nprocess = conf.N_PROCESS):
''' compute similarity between locations to predict and samples
return: a matrix of similarity values, each row is a location, each column is a sample
'''
## this import is necessary [on Windows]:
# http://stackoverflow.com/questions/28445373/python-import-numpy-as-np-from-outer-code-gets-lost-within-my-own-user-defined
import numpy as np
import raster, points, util, conf
def simLoc2SamplesV0(loc_ev, datapkg): # this function is needed for parallel computing using multiprocessing
import conf
# unpack data in datapkg
t0 = time.time()
sample_evs = datapkg[0]
evs = datapkg[1]
SD_evs = datapkg[2]
conf.TIME_KEEPING_DICT['parts']['data_transfer'].append(time.time()-t0)
# number of environmental variables
M = SD_evs.size
# number of samples
N = np.shape(sample_evs)[0]
sim = np.zeros(N)
t0 = time.time()
for i in range(N): # for each sample
sim0 = np.zeros(M)
sample_ev = sample_evs[i]
for j in range(M): # for each environmental variable
evi = loc_ev[j]
evj= sample_ev[j]
msrlevel = self.__envrasters[j].getMsrLevel()
if msrlevel == conf.MSR_LEVEL_NOMINAL or msrlevel == conf.MSR_LEVEL_ORDINAL:
if evi == evj:
sim_i = 1.0
else:
sim_i = 0.0
else:
SD_ev = SD_evs[j]
ev = evs[:,j]
SD_evj = np.sqrt(np.mean((ev - evj) ** 2))
sim_i = np.exp(-0.5 * (evi - evj) ** 2 / (SD_ev ** 2 / SD_evj) ** 2)
sim0[j] = sim_i
sim[i] = np.min(sim0) ## limiting factor
conf.TIME_KEEPING_DICT['parts']['compute'].append(time.time()-t0)
return sim
def simLoc2Samples(loc_ev, datapkg): # this function is needed for parallel computing using multiprocessing
import conf ## IMPORTANT - makes **conf.MSR_LEVELS** visible
# unpack data in datapkg
t0 = time.time()
sample_evs = datapkg[0]
REVS = datapkg[1]
SD_evs = datapkg[2]
AVG_evs = datapkg[3]
SUM_DIF_SQ_AVG = datapkg[4]
# Guiming 3/31/2019
MSRLEVES = datapkg[5]
conf.TIME_KEEPING_DICT['parts']['data_transfer'].append(time.time()-t0)
# number of environmental variables
M = SD_evs.size
# number of samples
N = np.shape(sample_evs)[0]
sim = np.zeros(N)
t0 = time.time()
for i in range(N): # for each sample
sim0 = np.zeros(M)
sample_ev = sample_evs[i]
for j in range(M): # for each environmental variable
evi = loc_ev[j]
evj= sample_ev[j]
# Guiming 3/31/2019 - SAVES MEM, NO NEED TO DISPATCH self.__envrasters TO EACH THREAD
msrlevel = MSRLEVES[j]
## this line below does not work without ** import conf ** at the begining of this function
if msrlevel == conf.MSR_LEVEL_NOMINAL or msrlevel == conf.MSR_LEVEL_ORDINAL:
#if msrlevel == 'nominal' or msrlevel == 'ordinal':
if evi == evj:
sim_i = 1.0
else:
sim_i = 0.0
else:
SD_ev = SD_evs[j]
delta = sample_ev[j] - AVG_evs[j]
tmp = SUM_DIF_SQ_AVG[j] + REVS * delta**2
SD_evj = np.sqrt(tmp/REVS)
sim_i = np.exp(-0.5 * (evi - evj) ** 2 / (SD_ev ** 2 / SD_evj) ** 2)
sim0[j] = sim_i
sim[i] = np.min(sim0) ## limiting factor
conf.TIME_KEEPING_DICT['parts']['compute'].append(time.time()-t0)
return sim
try:
# do dimension match check here
if np.shape(X)[1] != len(self.__envrasters):
print 'dimension mismatch in computing similarity in iPSM'
sys.exit(1)
msr_levels = []
if conf.NAIVE:
evs = np.zeros((self.__envrasters[0].getData().size, len(self.__envrasters)))
SD_evs = np.zeros(len(self.__envrasters))
AVG_evs = np.zeros(len(self.__envrasters))
for i in range(len(self.__envrasters)):
if conf.NAIVE:
evs[:, i] = self.__envrasters[i].getData().T
msr_levels.append(self.__envrasters[i].getMsrLevel())
SD_evs[i] = self.__envrasters[i].std
AVG_evs[i] = self.__envrasters[i].mean
NROWS = np.shape(X)[0]
REVS = self.__envrasters[0].getData().size
SUM_DIF_SQ_AVG = REVS * SD_evs**2
samples_evs = util.extractCovariatesAtPoints(self.__envrasters, self.__soilsamples)
samples_evs = np.array(samples_evs).T
if not parallel:
sim = np.zeros((NROWS, self.__soilsamples.size))
for i in range(NROWS):
if conf.NAIVE: ## naive implementaton
sim[i,:] = self.__simLoc2SamplesV0(X[i], evs, SD_evs)
else: ## with optimizations
sim[i,:] = self.__simLoc2Samples(X[i], samples_evs, REVS, SD_evs, AVG_evs, SUM_DIF_SQ_AVG)
else:
datapkg = []
for i in range(NROWS):
if conf.NAIVE: ## naive implementaton
datapkg.append([samples_evs, evs, SD_evs])
else:
# Guiming 3/31/2019
datapkg.append([samples_evs, REVS, SD_evs, AVG_evs, SUM_DIF_SQ_AVG, msr_levels])
#print 'n process', nprocess
pool = Pool(nprocess)
t0 = time.time()
if conf.NAIVE: ## naive implementaton
sim = np.array(pool.map(simLoc2SamplesV0, X, datapkg))
else:
sim = np.array(pool.map(simLoc2Samples, X, datapkg))
conf.TIME_KEEPING_DICT['parts']['compute'].append(time.time()-t0)
pool.clear()
return sim
except Exception as e:
raise
def predict_opencl_atom(self,
X=None,
predict_class=True,
single_cpu=conf.SINGLE_CPU,
opencl_config=conf.OPENCL_CONFIG):
''' PyOpenCL implementation of the iPSM approach
return: a vector of predictions, eacn for a row in X
'''
try:
t0 = time.time()
##### prepare data
# covariates values over the whole study area
r_evs = np.int32(self.__envrasters[0].getData().size)
c_evs = np.int32(len(self.__envrasters))
evs = np.zeros((r_evs, c_evs))
for i in range(len(self.__envrasters)):
evs[:, i] = self.__envrasters[i].getData().T
# standard deviation of each variable (over the whole study area)
Std_evs = np.std(evs, axis=0) ## added on Feb 26 2018
SD_evs = Std_evs.reshape(c_evs).astype(np.float32)
#print 'SD_evs', SD_evs.shape, SD_evs
# covariates values at prediction locations
if X is None: # if X is not provided, make prediction for the whole study area
X = []
for raster in self.__envrasters:
X.append(raster.getData())
X = np.array(X).T
r, c = np.shape(X)
nrows_X = np.int32(r)
ncols_X = np.int32(c)
X = X.reshape(nrows_X * ncols_X).astype(np.float32)
#print X, X.shape, nrows_X, ncols_X
MSRLEVES = self.__msrInts.reshape(c_evs).astype(np.int32)
#print MSRLEVES, MSRLEVES.shape
#t0 = time.time()
# covariates values at sample locations
if self.__soilsamples.covariates_at_points is None:
samples_X = util.extractCovariatesAtPoints(
self.__envrasters, self.__soilsamples)
else:
samples_X = self.__soilsamples.covariates_at_points[
0:c_evs].T ## prone to bug
nrows_samples = np.int32(samples_X.shape[1])
#print samples_X.shape
#print 'prepare samples took', time.time() - t0, 's'
samples_SD_evs = np.zeros((nrows_samples, c_evs))
AVG_evs = np.mean(evs, axis=0)
SUM_DIF_SQ_AVG = r_evs * Std_evs**2
#SUM_DIF_AVG = np.sum(evs - AVG_evs, axis = 0) ## == 0.0!!
#print 'SUM_DIF_AVG', SUM_DIF_AVG
for i in range(nrows_samples):
delta = samples_X[:, i].T - AVG_evs
tmp = SUM_DIF_SQ_AVG + r_evs * delta**2
samples_SD_evs[i] = np.sqrt(tmp / r_evs)
samples_SD_evs = np.array(samples_SD_evs).reshape(
nrows_samples * c_evs).astype(np.float32)
samples_X = np.array(samples_X).T.reshape(nrows_samples *
c_evs).astype(np.float32)
#print 'samples_X:', samples_X.shape, samples_X.min()
# sample weights
sample_weights = self.__soilsamples.weights.reshape(
nrows_samples).astype(np.float32)
#print 'sample_weights:', sample_weights.shape, sample_weights.min()
# sample attributes
sample_attributes = self.__soilsamples.attributes.reshape(
nrows_samples).astype(np.float32)
#print 'sample_attributes:', sample_attributes.shape, sample_attributes.min()
# hold predictions for instances in X
X_predictions = np.zeros(nrows_X).astype(np.float32)
# hold prediction uncertainties for instances in X
X_uncertainties = np.zeros(nrows_X).astype(np.float32)
print 'preparation on HOST took', time.time() - t0, 's'
##### config computing platform and device
for platform in cl.get_platforms():
#print platform.name
if platform.name == conf.OPENCL_CONFIG['Platform']:
PLATFORM = platform
'''if os.environ['COMPUTERNAME'] == 'DU-7CQTHQ2' and 'NVIDIA CUDA' in platform.name:
print '!!!'
#for device in platform.get_devices():
# if device.name == conf.OPENCL_CONFIG['Device']:
DEVICE = platform.get_devices()[0]
break
else:'''
# Print each device per-platform
for device in platform.get_devices():
#print device.name
if device.name == conf.OPENCL_CONFIG['Device']:
DEVICE = device
break
# opencl context
ctx = cl.Context([DEVICE])
# opencl command queue
queue = cl.CommandQueue(ctx)
##### allocate memory space on device
mf = cl.mem_flags
t0 = time.time()
#evs_g = cl.Buffer(ctx, mf.READ_ONLY | mf.COPY_HOST_PTR, hostbuf=evs)
SD_evs_g = cl.Buffer(ctx,
mf.READ_ONLY | mf.COPY_HOST_PTR,
hostbuf=SD_evs)
X_g = cl.Buffer(ctx, mf.READ_ONLY | mf.COPY_HOST_PTR, hostbuf=X)
MSRLEVES_g = cl.Buffer(ctx,
mf.READ_ONLY | mf.COPY_HOST_PTR,
hostbuf=MSRLEVES)
sample_X_g = cl.Buffer(ctx,
mf.READ_ONLY | mf.COPY_HOST_PTR,
hostbuf=samples_X)
## added 09/06/2017
samples_SD_evs_g = cl.Buffer(ctx,
mf.READ_ONLY | mf.COPY_HOST_PTR,
hostbuf=samples_SD_evs)
sample_weights_g = cl.Buffer(ctx,
mf.READ_ONLY | mf.COPY_HOST_PTR,
hostbuf=sample_weights)
sample_attributes_g = cl.Buffer(ctx,
mf.READ_ONLY | mf.COPY_HOST_PTR,
hostbuf=sample_attributes)
X_predictions_g = cl.Buffer(ctx, mf.WRITE_ONLY,
X_predictions.nbytes)
X_uncertainties_g = cl.Buffer(ctx, mf.WRITE_ONLY,
X_uncertainties.nbytes)
queue.finish()
print 'allocation and copy from HOST to DEVICE took', time.time(
) - t0, 's'
##### build opencl kernel from code in the file
f = open(conf.iPSM_KERNEL_FN, 'r')
fstr = "".join(f.readlines())
fstr = fstr.replace("#define N_SAMPLES 100",
"#define N_SAMPLES " + str(nrows_samples))
prg = cl.Program(ctx, fstr).build()
##### opencl computation
threshold = np.float32(self.__uncthreshold)
if predict_class:
mode = np.int32(1)
else:
mode = np.int32(0)
print X_predictions.shape
## improved version, 09/06/2017
if not single_cpu:
t0 = time.time()
completeEvent = \
prg.iPSM_Predict(queue, X_predictions.shape, None, r_evs, nrows_X, ncols_X, nrows_samples, mode, \
threshold, MSRLEVES_g, samples_SD_evs_g, SD_evs_g, X_g, sample_X_g, sample_weights_g, sample_attributes_g, \
X_predictions_g, X_uncertainties_g)
queue.finish()
print 'kernel took', time.time() - t0, 's'
#print queue.finish()
## added on Oct. 7, 2018 [sequential version - CPU]
else:
print 'SINGLE_CPU iPSM.predict_opencl() called'
t0 = time.time()
completeEvent = \
prg.iPSM_Predict_Sequential(queue, (1,), (1,), r_evs, nrows_X, ncols_X, nrows_samples, mode, \
threshold, MSRLEVES_g, samples_SD_evs_g, SD_evs_g, X_g, sample_X_g, sample_weights_g, sample_attributes_g, \
X_predictions_g, X_uncertainties_g)
queue.finish()
print 'kernel took', time.time() - t0, 's'
#print queue.finish()
#### wait until completions
events = [completeEvent]
queue.finish()
print 'up to events finished kernel took', time.time() - t0, 's'
#print queue.finish()
##### copy result data
t0 = time.time()
cl.enqueue_copy(queue,
X_predictions,
X_predictions_g,
wait_for=events) #.wait()
#print queue.finish()
cl.enqueue_copy(queue, X_uncertainties, X_uncertainties_g)
queue.finish()
print 'copy from DEVICE to HOST took', time.time() - t0, 's'
y = np.vstack((X_predictions, X_uncertainties)).T
#print y
return y
except Exception as e:
raise
