def _wmvnpdf(x, pi, mu, va, d=1, logged=False, use_gpu=True, **kwargs):
if len(x.shape) == 1:
x = x.reshape((1, x.shape))
if len(mu.shape) == 1:
mu = mu.reshape((1, mu.shape))
if len(va.shape) == 2:
va = va.reshape(1, va.shape[0], va.shape[1])
if len(va.shape) == 1:
va = va.reshape(va.shape[0], 1, 1)
if isinstance(pi, float) or isinstance(pi, int):
pi = array([pi])
elif isinstance(pi, ndarray):
if len(pi.shape) == 0:
pi = pi.reshape((1))
if has_gpu and use_gpu:
if 'device' in kwargs:
dev = kwargs['device']
else:
dev = 0
select_gpu(dev)
return mvnpdf_multi(x, mu, va, weights=pi, logged=logged, order='C').astype('float64')
else:
if logged:
return mvn_weighted_logged(x, mu, va, pi)
else:
return exp(mvn_weighted_logged(x, mu, va, pi))
def _mvnpdf(x, mu, va, logged=False, use_gpu=True, **kwargs):
if len(x.shape) == 1:
x = x.reshape((1, x.shape[0]))
if len(mu.shape) == 1:
mu = mu.reshape((1, mu.shape[0]))
if len(va.shape) == 2:
va = va.reshape(1, va.shape[0], va.shape[1])
if has_gpu and use_gpu:
if 'device' in kwargs:
dev = kwargs['device']
else:
dev = 0
select_gpu(dev)
return mvnpdf_multi(
x,
mu,
va,
weights=np.ones(mu.shape[0]),
logged=logged,
order='C').astype('float64')
else:
if logged:
return mvn_weighted_logged(x, mu, va, np.ones(mu.shape[0]))
else:
return np.exp(mvn_weighted_logged(x, mu, va, np.ones(mu.shape[0])))
def get_dp_labels_gpu(gpu_data, w, mu, sigma, relabel=False):
densities = gpustats.mvnpdf_multi(
gpu_data[0],
mu,
sigma,
weights=w.flatten(),
get=False,
logged=True,
order='C'
)
if relabel:
z = np.asarray(
cuda_functions.gpu_apply_row_max(densities)[1].get(),
dtype='i'
)
else:
z = None
labels = np.asarray(
gpu_sampler.sample_discrete(densities, logged=True),
dtype='i'
)
densities.gpudata.free()
del densities
return labels, z
def _mvnpdf(x, mu, va, n=1, logged=False, use_gpu=True, **kwargs):
if len(x.shape) == 1:
x = x.reshape((1, x.shape[0]))
if len(mu.shape) == 1:
mu = mu.reshape((1, mu.shape[0]))
if len(va.shape) == 2:
va = va.reshape(1, va.shape[0], va.shape[1])
if has_gpu and use_gpu:
if 'device' in kwargs:
dev = kwargs['device']
else:
dev = 0
select_gpu(dev)
return mvnpdf_multi(x,
mu,
va,
weights=ones(mu.shape[0]),
logged=logged,
order='C').astype('float64')
else:
if logged:
return mvn_weighted_logged(x, mu, va, ones(mu.shape[0]))
else:
return exp(mvn_weighted_logged(x, mu, va, ones(mu.shape[0])))
def get_hdp_labels_gpu(gpu_data, w, mu, sigma, relabel=False):
labels = []
z = []
for i, data_set in enumerate(gpu_data):
densities = gpustats.mvnpdf_multi(data_set,
mu,
sigma,
weights=w[i].flatten(),
get=False,
logged=True,
order='C')
if relabel:
z.append(
np.asarray(
cuda_functions.gpu_apply_row_max(densities)[1].get(),
dtype='i'))
else:
z.append(None)
labels.append(
np.asarray(gpu_sampler.sample_discrete(densities, logged=True),
dtype='i'))
densities.gpudata.free()
del densities
return labels, z
def _update_labels(self, mu, Sigma, weights):
# GPU business happens?
densities = gpustats.mvnpdf_multi(self.data, mu, Sigma, weights=weights)
densities = (densities.T / densities.sum(1)).T
# convert this to run in the GPU
return ffbs.f32_sample_discrete(densities)
def _compare_multi(n, k, p):
data, means, covs = _make_test_case(n, k, p)
# cpu in PyMC
py_result = python_mvnpdf(data, means, covs)
# gpu
result = gps.mvnpdf_multi(data, means, covs)
return result, py_result
def _compare_multi(n, k, p):
data, means, covs = _make_test_case(n, k, p)
# cpu in PyMC
pyresult = compat.python_mvnpdf(data, means, covs)
# gpu
result = gps.mvnpdf_multi(data, means, covs)
return result, pyresult
def get_bem_densities_gpu(gpu_data, w, mu, sigma):
densities = gpustats.mvnpdf_multi(gpu_data[0],
mu,
sigma,
weights=w.flatten(),
get=False,
logged=True,
order='C')
dens = np.asarray(densities.get(), dtype='d')
densities.gpudata.free()
return dens
def get_bem_densities_gpu(gpu_data, w, mu, sigma):
densities = gpustats.mvnpdf_multi(
gpu_data[0],
mu,
sigma,
weights=w.flatten(),
get=False,
logged=True,
order='C'
)
dens = np.asarray(densities.get(), dtype='d')
densities.gpudata.free()
return dens
def get_dp_labels_gpu(gpu_data, w, mu, sigma, relabel=False):
densities = gpustats.mvnpdf_multi(gpu_data[0],
mu,
sigma,
weights=w.flatten(),
get=False,
logged=True,
order='C')
if relabel:
z = np.asarray(cuda_functions.gpu_apply_row_max(densities)[1].get(),
dtype='i')
else:
z = None
labels = np.asarray(gpu_sampler.sample_discrete(densities, logged=True),
dtype='i')
densities.gpudata.free()
del densities
return labels, z
def get_hdp_labels_gpu(gpu_data, w, mu, sigma, relabel=False):
labels = []
z = []
for i, data_set in enumerate(gpu_data):
densities = gpustats.mvnpdf_multi(
data_set,
mu,
sigma,
weights=w[i].flatten(),
get=False,
logged=True,
order='C'
)
if relabel:
z.append(
np.asarray(
cuda_functions.gpu_apply_row_max(densities)[1].get(),
dtype='i'
)
)
else:
z.append(None)
labels.append(
np.asarray(
gpu_sampler.sample_discrete(densities, logged=True),
dtype='i'
)
)
densities.gpudata.free()
del densities
return labels, z
a_Sigma.append(Sigma)
for it in range(subtasknum):
# get task parameters
params = a_params[it]
dataind, ncomp, ttype, gid = params
nobs, ndim = alldata[dataind].shape
w = a_w[it]
mu = a_mu[it]
Sigma = a_Sigma[it]
if ttype > 0: # 1 -- just densities ... 2 -- relabel too
# do GPU work ...
densities = gpustats.mvnpdf_multi(gdata[dataind], mu, Sigma,
weights=w.flatten(),
get=False, logged=True,
order='C')
if ttype == 2: # identification!!
Z = np.asarray(
cufuncs.gpu_apply_row_max(densities)[1].get(),
dtype='i'
)
else:
Z = None
labs = np.asarray(
gsamp.sample_discrete(densities, logged=True),
dtype='i'
)
subresult = [
for it in range(subtasknum):
# get task parameters
params = a_params[it]
dataind, ncomp, ttype, gid = params
nobs, ndim = alldata[dataind].shape
w = a_w[it]
mu = a_mu[it]
Sigma = a_Sigma[it]
if ttype > 0: # 1 -- just densities ... 2 -- relabel too
# do GPU work ...
densities = gpustats.mvnpdf_multi(gdata[dataind],
mu,
Sigma,
weights=w.flatten(),
get=False,
logged=True,
order='C')
if ttype == 2: # identification!!
Z = np.asarray(
cufuncs.gpu_apply_row_max(densities)[1].get(),
dtype='i')
else:
Z = None
labs = np.asarray(gsamp.sample_discrete(densities,
logged=True),
dtype='i')
subresult = [
