def add_batch(self, X, T, wc=None):
"""Add a batch of training data to an iterative solution, weighted if neeed.
The batch is processed as a whole, the training data is splitted in `ELM.add_data()` method.
With parameters HH_out, HT_out, the output will be put into these matrices instead of model.
Args:
X (matrix): input data matrix size (N * `inputs`)
T (matrix): output data matrix size (N * `outputs`)
wc (vector): vector of weights for data samples, one weight per sample, size (N * 1)
HH_out, HT_out (matrix, optional): output matrices to add batch result into, always given together
"""
devH = self._project(X, dev=True)
T = np.array(T, order="C", dtype=self.precision)
devT = gpuarray.to_gpu(T)
if wc is not None: # apply weights if given
w = np.array(wc**0.5, dtype=self.precision)[:, None] # re-shape to column matrix
devWC = gpuarray.to_gpu(w)
misc.mult_matvec(devH, devWC, axis=0, out=devH)
misc.mult_matvec(devT, devWC, axis=0, out=devT)
if self.HH is None: # initialize space for self.HH, self.HT
self.HT = misc.zeros((self.L, self.outputs), dtype=self.precision)
self.HH = linalg.eye(self.L, dtype=self.precision)
self.HH *= self.norm
linalg.add_dot(devH, devT, self.HT, transa='T')
if self.precision is np.float64:
linalg.add_dot(devH, devH, self.HH, transa='T')
else:
cublas.cublasSsyrk(self.handle, 'L', 'N', self.L, X.shape[0], 1, devH.ptr, self.L, 1, self.HH.ptr, self.L)
def add_batch(self, X, T, wc=None):
"""Add a batch of training data to an iterative solution, weighted if neeed.
The batch is processed as a whole, the training data is splitted in `ELM.add_data()` method.
With parameters HH_out, HT_out, the output will be put into these matrices instead of model.
Args:
X (matrix): input data matrix size (N * `inputs`)
T (matrix): output data matrix size (N * `outputs`)
wc (vector): vector of weights for data samples, one weight per sample, size (N * 1)
HH_out, HT_out (matrix, optional): output matrices to add batch result into, always given together
"""
devH = self._project(X, dev=True)
T = np.array(T, order="C", dtype=self.precision)
devT = gpuarray.to_gpu(T)
if wc is not None: # apply weights if given
w = np.array(
wc**0.5,
dtype=self.precision)[:, None] # re-shape to column matrix
devWC = gpuarray.to_gpu(w)
misc.mult_matvec(devH, devWC, axis=0, out=devH)
misc.mult_matvec(devT, devWC, axis=0, out=devT)
if self.HH is None: # initialize space for self.HH, self.HT
self.HT = misc.zeros((self.L, self.outputs), dtype=self.precision)
self.HH = linalg.eye(self.L, dtype=self.precision)
self.HH *= self.norm
linalg.add_dot(devH, devT, self.HT, transa='T')
if self.precision is np.float64:
linalg.add_dot(devH, devH, self.HH, transa='T')
else:
cublas.cublasSsyrk(self.handle, 'L', 'N', self.L, X.shape[0], 1,
devH.ptr, self.L, 1, self.HH.ptr, self.L)
def sqsum_adddot2(a, b):
"""
Compute squared euclidean distance between two 2D arrays representing
n-dimensional points using GPU. This uses the GPUArray versions of the
input arrays to compute element-wise summations of squared sum of rows and
accumulates into the matrix-multiplication result residing on GPU.
The final result resides on GPU.
Parameters
----------
A : ndarray
2D NumPy array of float dtype representing n-dimensional points, with
each row being one point.
B : ndarray
2D NumPy array of float dtype representing n-dimensional points, with
each row being one point.
Returns
-------
out : GPUArray
This holds the euclidean distances.
"""
a_gpu = gpuarray.to_gpu(a)
b_gpu = gpuarray.to_gpu(b)
c_gpu = sq_sums(a_gpu, b_gpu)
return culinalg.add_dot(a_gpu, b_gpu, c_gpu, transb='T', alpha=-2.0)
def sqsum_adddot(a, b, method):
"""
Compute squared euclidean distance between two 2D arrays representing
n-dimensional points using GPU. This uses the input arrays themselves to
compute element-wise summations of squared sum of rows and accumulates into
the matrix-multiplication result residing on GPU.
The final result resides on GPU.
Parameters
----------
A : ndarray
2D NumPy array of float dtype representing n-dimensional points, with
each row being one point.
B : ndarray
2D NumPy array of float dtype representing n-dimensional points, with
each row being one point.
method : str
It can be 'add_togpu' or 'togpu_misc_add' or 'togpu_cuda_add'.
Refer to function "squared_sum" for more information.
Returns
-------
out : GPUArray
This holds the euclidean distances residing on GPU.
"""
a_gpu = gpuarray.to_gpu(a)
b_gpu = gpuarray.to_gpu(b)
c_gpu = squared_sum(a, b, method=method)
return culinalg.add_dot(a_gpu, b_gpu, c_gpu, transb='T', alpha=-2.0)
def get_distances_to_centers(self, data):
# make sure the array is c order
data = np.asarray(data, dtype=np.float32, order='C')
# ship to gpu
data_gpu = gpuarray.to_gpu(data)
# alloc space on gpu for distances
dists_shape = (data.shape[0], self.centers.shape[0])
dists_gpu = gpuarray.zeros(dists_shape, np.float32)
# calc data norms on gpu
data_norms = cumisc.sum(data_gpu**2, axis=1)
# calc distance on gpu
cumisc.add_matvec(dists_gpu, self.center_norms, 1, dists_gpu)
cumisc.add_matvec(dists_gpu, data_norms, 0, dists_gpu)
culinalg.add_dot(data_gpu, self.centers_gpu,
dists_gpu, transb='T', alpha=-2.0)
return dists_gpu
def multiplyConv2DHGradKLGPU(W, H, V, VLam, doDivision=True):
"""
Compute the 2D convolutional multiplicative update for H
under the Kullback-Liebler divergence, using skcuda
:param W: A TxNxK matrix of K sources over spatiotemporal spans NxT\
:param H: A FxKxM matrix of source activations for each submatrix of W\
over F transpositions over M time
:param VLam: Convolutional WH multiplication
:param doDivision: If true, return the factor Numerator/Denomenator\
otherwise, return (Numerator, Denomenator)
:returns Ratio: A FxKxM matrix of multiplicative updates for H\
or (RatioNum, RatioDenom) if doDivision = False
"""
HNums = gpuarray.zeros(H.shape, np.float32)
HDenoms = gpuarray.zeros((H.shape[0], H.shape[1]), np.float32)
thisVLam = VLam.copy()
ZerosToOnes(thisVLam)
VLamQuot = skcuda.misc.divide(V, thisVLam)
thisVLamQuot = VLamQuot.copy()
thisW = W.copy()
for t in range(W.shape[0]):
if t > 0:
z = gpuarray.zeros((V.shape[0], t), np.float32)
thisVLamQuot[:, 0:-t] = VLamQuot[:, t::]
thisVLamQuot[:, -t::] = z
for f in range(H.shape[0]):
if f > 0:
thisW[t, f::, :] = W[t, 0:-f, :]
thisW[t, 0:f, :] = gpuarray.zeros((f, W.shape[2]), np.float32)
linalg.add_dot(thisW[t, :, :],
thisVLamQuot,
HNums[f, :, :],
transa='T')
HDenoms[f, :] = skcuda.misc.add(HDenoms[f, :],
skcuda.misc.sum(thisW[t, :, :], 0))
HDenoms = TileHDenom(HDenoms[:, :, None], H.shape[2])
if doDivision:
return skcuda.misc.divide(HNums, HDenoms)
else:
return (HNums, HDenoms)
def _impl_add_dot_matrix_tests(self, dtype, transa, transb):
a = np.asarray(np.random.rand(4, 2), dtype)
if transa == 'n':
b = np.asarray(np.random.rand(2, 2), dtype)
else:
b = np.asarray(np.random.rand(4, 4), dtype)
a_gpu = gpuarray.to_gpu(a)
b_gpu = gpuarray.to_gpu(b)
aa = a if transa == 'n' else a.T
bb = b if transb == 'n' else b.T
c = np.asarray(np.random.rand(aa.shape[0], bb.shape[1]), dtype)
c_gpu = gpuarray.to_gpu(c)
c_gpu = linalg.add_dot(a_gpu, b_gpu, c_gpu, transa, transb)
assert np.allclose(c + np.dot(aa, bb), c_gpu.get())
a = a.astype(dtype, order="F", copy=True)
b = b.astype(dtype, order="F", copy=True)
c = c.astype(dtype, order="F", copy=True)
a_gpu = gpuarray.to_gpu(a)
b_gpu = gpuarray.to_gpu(b)
c_gpu = gpuarray.to_gpu(c)
c_gpu = linalg.add_dot(a_gpu, b_gpu, c_gpu, transa, transb)
assert np.allclose(c+np.dot(aa, bb), c_gpu.get())
def _impl_add_dot_matrix_tests(self, dtype, transa, transb):
a = np.asarray(np.random.rand(4, 2), dtype)
if transa == 'n':
b = np.asarray(np.random.rand(2, 2), dtype)
else:
b = np.asarray(np.random.rand(4, 4), dtype)
a_gpu = gpuarray.to_gpu(a)
b_gpu = gpuarray.to_gpu(b)
aa = a if transa == 'n' else a.T
bb = b if transb == 'n' else b.T
c = np.asarray(np.random.rand(aa.shape[0], bb.shape[1]), dtype)
c_gpu = gpuarray.to_gpu(c)
c_gpu = linalg.add_dot(a_gpu, b_gpu, c_gpu, transa, transb)
assert np.allclose(c + np.dot(aa, bb), c_gpu.get())
a = a.astype(dtype, order="F", copy=True)
b = b.astype(dtype, order="F", copy=True)
c = c.astype(dtype, order="F", copy=True)
a_gpu = gpuarray.to_gpu(a)
b_gpu = gpuarray.to_gpu(b)
c_gpu = gpuarray.to_gpu(c)
c_gpu = linalg.add_dot(a_gpu, b_gpu, c_gpu, transa, transb)
assert np.allclose(c+np.dot(aa, bb), c_gpu.get())
def multiplyConv2DHGradGPU(W, H, V, VLam, doDivision=True):
"""
Compute the 2D convolutional multiplicative update for H using skcuda
:param W: A TxNxK GPU array of K sources over spatiotemporal spans NxT\
:param H: A FxKxM GPU array of source activations for each submatrix of W\
over F transpositions over M time
:param V: An MxN GPU target array
:param VLam: An MxN GPU estimate array
:param doDivision: If true, return the factor Numerator/Denomenator\
otherwise, return (Numerator, Denomenator)
"""
thisV = V.copy()
thisVLam = VLam.copy()
thisW = W.copy()
HNums = gpuarray.zeros(H.shape, np.float32)
HDenoms = gpuarray.zeros(H.shape, np.float32)
for t in range(W.shape[0]):
if t > 0:
#thisV = shiftMatLRUD(V, dj=-t)
z = gpuarray.zeros((V.shape[0], t), np.float32)
thisV[:, 0:-t] = V[:, t::]
thisV[:, -t::] = z
thisVLam[:, 0:-t] = VLam[:, t::]
thisVLam[:, -t::] = z
for f in range(H.shape[0]):
if f > 0:
#thisW = shiftMatLRUD(W[t, :, :], di=f)
thisW[t, f::, :] = W[t, 0:-f, :]
thisW[t, 0:f, :] = gpuarray.zeros((f, W.shape[2]), np.float32)
linalg.add_dot(thisW[t, :, :], thisV, HNums[f, :, :], transa='T')
linalg.add_dot(thisW[t, :, :],
thisVLam,
HDenoms[f, :, :],
transa='T')
if doDivision:
return skcuda.misc.divide(HNums, HDenoms)
else:
return (HNums, HDenoms)
def multiplyConv2DWGradGPU(W, H, V, VLam, doDivision=True):
"""
Compute the 2D convolutional multiplicative update for W using skcuda
:param W: A TxNxK GPU array of K sources over spatiotemporal spans NxT\
:param H: A FxKxM GPU array of source activations for each submatrix of W\
over F transpositions over M time
:param V: An MxN GPU target array
:param VLam: An MxN GPU estimate array
:param doDivision: If true, return the factor Numerator/Denomenator\
otherwise, return (Numerator, Denomenator)
"""
thisV = V.copy()
thisVLam = VLam.copy()
thisH = H.copy()
WNums = gpuarray.zeros(W.shape, np.float32)
WDenoms = gpuarray.zeros(W.shape, np.float32)
for f in range(H.shape[0]):
if f > 0:
z = gpuarray.zeros((f, V.shape[1]), np.float32)
thisV[0:-f, :] = V[f::, :]
thisV[-f::, :] = z
thisVLam[0:-f, :] = VLam[f::, :]
thisVLam[-f::, :] = z
for t in range(W.shape[0]):
if t > 0:
thisH[f, :, t::] = H[f, :, 0:-t]
thisH[f, :, 0:t] = gpuarray.zeros((H.shape[1], t), np.float32)
linalg.add_dot(thisV, thisH[f, :, :], WNums[t, :, :], transb='T')
linalg.add_dot(thisVLam,
thisH[f, :, :],
WDenoms[t, :, :],
transb='T')
if doDivision:
return skcuda.misc.divide(WNums, WDenoms)
else:
return (WNums, WDenoms)
def _impl_test_dot_strided(self, dtype):
# n/n
a = np.asarray(np.random.rand(4, 10), dtype)
b = np.asarray(np.random.rand(2, 20), dtype)
c = np.zeros((4, 30), dtype)
a_gpu = gpuarray.to_gpu(a)
b_gpu = gpuarray.to_gpu(b)
c_gpu = gpuarray.to_gpu(c)
linalg.add_dot(a_gpu[:, 4:6], b_gpu[:, 2:8], c_gpu[:, 1:7], 'n', 'n')
res = c_gpu.get()
assert np.allclose(np.dot(a[:, 4:6], b[:, 2:8]), res[:, 1:7])
# t/n
a = np.asarray(np.random.rand(4, 10), dtype)
b = np.asarray(np.random.rand(4, 20), dtype)
c = np.zeros((2, 30), dtype)
a_gpu = gpuarray.to_gpu(a)
b_gpu = gpuarray.to_gpu(b)
c_gpu = gpuarray.to_gpu(c)
linalg.add_dot(a_gpu[:, 4:6], b_gpu[:, 2:8], c_gpu[:, 1:7], 't', 'n')
res = c_gpu.get()
assert np.allclose(np.dot(a[:, 4:6].T, b[:, 2:8]), res[:, 1:7])
# n/t
a = np.asarray(np.random.rand(4, 10), dtype)
b = np.asarray(np.random.rand(6, 20), dtype)
c = np.zeros((4, 30), dtype)
a_gpu = gpuarray.to_gpu(a)
b_gpu = gpuarray.to_gpu(b)
c_gpu = gpuarray.to_gpu(c)
linalg.add_dot(a_gpu[:, 4:10], b_gpu[:, 2:8], c_gpu[:, 1:7], 'n', 't')
res = c_gpu.get()
assert np.allclose(np.dot(a[:, 4:10], b[:, 2:8].T), res[:, 1:7])
# t/t
a = np.asarray(np.random.rand(6, 10), dtype)
b = np.asarray(np.random.rand(8, 20), dtype)
c = np.zeros((2, 30), dtype)
a_gpu = gpuarray.to_gpu(a)
b_gpu = gpuarray.to_gpu(b)
c_gpu = gpuarray.to_gpu(c)
linalg.add_dot(a_gpu[:, 4:6], b_gpu[:, 2:8], c_gpu[:, 1:9], 't', 't')
res = c_gpu.get()
assert np.allclose(np.dot(a[:, 4:6].T, b[:, 2:8].T), res[:, 1:9])
def _impl_test_dot_strided(self, dtype):
# n/n
a = np.asarray(np.random.rand(4, 10), dtype)
b = np.asarray(np.random.rand(2, 20), dtype)
c = np.zeros((4, 30), dtype)
a_gpu = gpuarray.to_gpu(a)
b_gpu = gpuarray.to_gpu(b)
c_gpu = gpuarray.to_gpu(c)
linalg.add_dot(a_gpu[:, 4:6], b_gpu[:, 2:8], c_gpu[:, 1:7], 'n', 'n')
res = c_gpu.get()
assert np.allclose(np.dot(a[:, 4:6], b[:, 2:8]), res[:, 1:7])
# t/n
a = np.asarray(np.random.rand(4, 10), dtype)
b = np.asarray(np.random.rand(4, 20), dtype)
c = np.zeros((2, 30), dtype)
a_gpu = gpuarray.to_gpu(a)
b_gpu = gpuarray.to_gpu(b)
c_gpu = gpuarray.to_gpu(c)
linalg.add_dot(a_gpu[:, 4:6], b_gpu[:, 2:8], c_gpu[:, 1:7], 't', 'n')
res = c_gpu.get()
assert np.allclose(np.dot(a[:, 4:6].T, b[:, 2:8]), res[:, 1:7])
# n/t
a = np.asarray(np.random.rand(4, 10), dtype)
b = np.asarray(np.random.rand(6, 20), dtype)
c = np.zeros((4, 30), dtype)
a_gpu = gpuarray.to_gpu(a)
b_gpu = gpuarray.to_gpu(b)
c_gpu = gpuarray.to_gpu(c)
linalg.add_dot(a_gpu[:, 4:10], b_gpu[:, 2:8], c_gpu[:, 1:7], 'n', 't')
res = c_gpu.get()
assert np.allclose(np.dot(a[:, 4:10], b[:, 2:8].T), res[:, 1:7])
# t/t
a = np.asarray(np.random.rand(6, 10), dtype)
b = np.asarray(np.random.rand(8, 20), dtype)
c = np.zeros((2, 30), dtype)
a_gpu = gpuarray.to_gpu(a)
b_gpu = gpuarray.to_gpu(b)
c_gpu = gpuarray.to_gpu(c)
linalg.add_dot(a_gpu[:, 4:6], b_gpu[:, 2:8], c_gpu[:, 1:9], 't', 't')
res = c_gpu.get()
assert np.allclose(np.dot(a[:, 4:6].T, b[:, 2:8].T), res[:, 1:9])
def dot_add_mm(self, a, b, out, transa=False, transb=False):
transa = 'T' if transa else 'N'
transb = 'T' if transb else 'N'
culinalg.add_dot(a, b, out, transa, transb)
def run(self, iterations):
for i in range(0,iterations):
# F = XG(G.T G)^-1
linalg.add_dot(self.G_gpu, self.G_gpu, self.GTG_gpu, transa="T",
beta=0.)
try:
self.GTGinv_gpu.set(np.linalg.inv(self.GTG_gpu.get()))
# linalg.pinv only worked with CULA
except LinAlgError:
self.GTGinv_gpu.set(np.linalg.iinv(self.GTG_gpu.get()))
linalg.add_dot(self.X_gpu, self.G_gpu, self.XG_gpu, beta=0.)
linalg.add_dot(self.XG_gpu, self.GTGinv_gpu, self.F_gpu, beta=0.)
# preparation and calculation of the matrix separations
linalg.add_dot(self.X_gpu, self.F_gpu, self.XTF_gpu, transa="T",
beta=0.)
linalg.add_dot(self.F_gpu, self.F_gpu, self.FTF_gpu, transa="T",
beta=0.)
self.matrix_separationXTF(self.XTF_gpu, self.XTFpos_gpu,
self.XTFneg_gpu,
block=(self.block_G, 1, 1),
grid=(self.grid_G, 1))
self.matrix_separationFTF(self.FTF_gpu, self.FTFpos_gpu,
self.FTFneg_gpu,
block=(self.block_FTF, 1, 1),
grid=(self.grid_FTF, 1))
# compute the G update
linalg.add_dot(self.G_gpu, self.FTFpos_gpu, self.GFTFpos_gpu,
beta=0.)
linalg.add_dot(self.G_gpu, self.FTFneg_gpu, self.GFTFneg_gpu,
beta=0.)
self.G_ew_update(self.G_gpu, self.XTFpos_gpu, self.GFTFneg_gpu,
self.XTFneg_gpu, self.GFTFpos_gpu,
block=(self.block_G, 1, 1),
grid=(self.grid_G, 1))
# test for convergence
if (i % self.niter_test_conv == 0) and self.checkConvergence():
print "NMF converged after %i iterations" % i
break
def run(self, iterations):
for i in range(0,iterations):
# update H
linalg.add_dot(self.W_gpu, self.X_gpu, self.WTX_gpu, transa="T",
beta=0.) # add_dot is faster than dot, dot calls ad
linalg.add_dot(self.W_gpu, self.W_gpu, self.WTW_gpu, transa="T",
beta=0.)
linalg.add_dot(self.WTW_gpu, self.H_gpu, self.WTWH_gpu, beta=0.)
self.update_H(self.H_gpu, self.WTX_gpu, self.WTWH_gpu,
block=(self.block_H, 1, 1), grid=(self.grid_H, 1))
# update W
linalg.add_dot(self.X_gpu, self.H_gpu, self.XHT_gpu, transb="T",
beta=0.)
linalg.add_dot(self.W_gpu, self.H_gpu, self.WH_gpu, beta=0.)
linalg.add_dot(self.WH_gpu, self.H_gpu, self.WHHT_gpu, transb="T", beta=0.)
self.update_W(self.W_gpu, self.XHT_gpu, self.WHHT_gpu,
block=(self.block_W, 1, 1), grid=(self.grid_W, 1))
# test for convergence
if (i % self.niter_test_conv == 0) and self.checkConvergence():
print "NMF converged after %i iterations" % i
break
def demosaick_gpu(img):
img = gp.to_gpu(img)
p2x = im2col(img, _i2c2)
cm.log(img + _eps, out=img)
p1x = im2col(img, _i2c1)
wA = p1x.shape[0]
wB = p2x.shape[0]
hA = p1x.shape[1]
hB = p2x.shape[1]
# Path 1
p1x = p1x.reshape([wA * hA, 576])
p1y = lg.dot(p1x, _wts.int1)
cm.exp(p1y, out=p1y)
p1y = p1y.reshape([wA * hA * 64, 3 * _ofac])
p1x = lg.dot(p1y, _wts.int2)
msc.add_matvec(p1x, _wts.int2b, out=p1x)
p1x = p1x.reshape([wA * hA * 64 * 3, _ofac])
# Path 2
# conv1
p2x = p2x.reshape([wB * hB, 64])
p2y = lg.dot(p2x, _wts.c1)
msc.add_matvec(p2y, _wts.c1b, out=p2y)
gp.maximum(p2y, 0., p2y)
p2y = p2y.reshape([wB, hB, _numsel])
# conv2
shI = [wB - 1, hB - 1, _numsel]
shM = [(wB - 1) * (hB - 1), _numsel]
p2x = gp.empty(shM, dtype=np.float32)
pTT = gp.empty(shI, dtype=np.float32)
pTT = pTT.reshape(shI)
pTT[...] = p2y[0:-1, 0:-1, :]
pTT = pTT.reshape(shM)
p2x = lg.dot(pTT, _wts.c200)
pTT = pTT.reshape(shI)
pTT[...] = p2y[0:-1, 1:, :]
pTT = pTT.reshape(shM)
lg.add_dot(pTT, _wts.c201, p2x)
pTT = pTT.reshape(shI)
pTT[...] = p2y[1:, 0:-1, :]
pTT = pTT.reshape(shM)
lg.add_dot(pTT, _wts.c210, p2x)
pTT = pTT.reshape(shI)
pTT[...] = p2y[1:, 1:, :]
pTT = pTT.reshape(shM)
lg.add_dot(pTT, _wts.c211, p2x)
msc.add_matvec(p2x, _wts.c2b, out=p2x)
gp.maximum(p2x, 0., p2x)
p2x = p2x.reshape(shI)
# conv 3
shI = [wB - 2, hB - 2, _numsel]
shM = [(wB - 2) * (hB - 2), _numsel]
p2y = gp.empty(shM, dtype=np.float32)
pTT = gp.empty(shI, dtype=np.float32)
pTT = pTT.reshape(shI)
pTT[...] = p2x[0:-1, 0:-1, :]
pTT = pTT.reshape(shM)
p2y = lg.dot(pTT, _wts.c300)
pTT = pTT.reshape(shI)
pTT[...] = p2x[0:-1, 1:, :]
pTT = pTT.reshape(shM)
lg.add_dot(pTT, _wts.c301, p2y)
pTT = pTT.reshape(shI)
pTT[...] = p2x[1:, 0:-1, :]
pTT = pTT.reshape(shM)
lg.add_dot(pTT, _wts.c310, p2y)
pTT = pTT.reshape(shI)
pTT[...] = p2x[1:, 1:, :]
pTT = pTT.reshape(shM)
lg.add_dot(pTT, _wts.c311, p2y)
msc.add_matvec(p2y, _wts.c3b, out=p2y)
gp.maximum(p2y, 0., p2y)
p2x = lg.dot(p2y, _wts.sout)
msc.add_matvec(p2x, _wts.soutb, out=p2x)
gp.maximum(p2x, 0., p2x)
p2x = p2x.reshape(p1x.shape)
# Combine
p1x *= p2x
p1 = msc.sum(p1x, axis=1)
gp.maximum(p1, 0., p1)
gp.minimum(p1, 1., p1)
p1 = p1.reshape([wA, hA, 64 * 3])
im = p2im(p1.get())
return im
def getnextz(\
z_last,za, c_obs_long,\
Nx,Ny,nx,ny,nXobs,\
Se_inv,Sa_inv,\
Arule, Brule, Crule, Drule,\
KAxrule, KAyrule, \
KBxrule, KByrule, \
KCxrule, KCyrule, \
KDxrule, KDyrule):
nx_last_long = cds.dot(Nx,z_last)
ny_last_long = cds.dot(Ny,z_last)
nx_last = matrix(reshape(nx_last_long,(ny-1,nx-1)));
ny_last = matrix(reshape(ny_last_long,(ny-1,nx-1)));
vbigK, bigc_last = getbigK(\
nx,ny,nXobs,
nx_last, ny_last,
Arule, Brule, Crule, Drule,
KAxrule, KAyrule,
KBxrule, KByrule,
KCxrule, KCyrule,
KDxrule, KDyrule)
NxNy = vstack((Nx,Ny))
KN = cds.dot(vbigK, NxNy)
nobs = nXobs*4 # int*int
bigc_last_long = matrix(reshape(cds.T(bigc_last),(nobs,1)))
delta_c = cds.substract(c_obs_long, bigc_last_long)
# # This is the Gauss method
# dz = linalg.inv(KN.T*KN)*KN.T*delta_c
# z_next = z_last + dz
# di2 = squeeze(dot(dz.T,dz))
# This is the optimal estimation method
# nasledujici dva radky by se daly mergnout
#term3
zz_gpu = gpuarray.to_gpu(z_last-za)
delta_c_gpu = gpuarray.to_gpu(delta_c)
KN_gpu = gpuarray.to_gpu(KN) # (important for term2)
temp0_gpu = linalg.add_dot(KN_gpu, zz_gpu, delta_c_gpu)
zz_gpu.gpudata.free()
del(zz_gpu)
Se_inv_gpu = gpuarray.to_gpu(Se_inv) # (important for term2)
temp1_gpu = linalg.dot(Se_inv_gpu, temp0_gpu)
#temp0_gpu.gpudata.free()
#del(temp0_gpu)
term3_gpu = linalg.dot(KN_gpu, temp1_gpu, transa="T")
term3 = term3_gpu.get()
temp1_gpu.gpudata.free()
del(temp1_gpu)
term3_gpu.gpudata.free()
del(term3_gpu)
# term2
temp2_gpu = linalg.dot(Se_inv_gpu, KN_gpu)
Se_inv_gpu.gpudata.free()
del(Se_inv_gpu)
Sa_inv_gpu = gpuarray.to_gpu(Sa_inv)
term2_gpu = linalg.add_dot(KN_gpu, temp2_gpu, Sa_inv_gpu, transa="T")
temp2_gpu.gpudata.free()
del(temp2_gpu)
KN_gpu.gpudata.free()
del(KN_gpu)
term2 = term2_gpu.get()
#term0 = cds.dot3(Se_inv, KN)
#term2 = Sa_inv+term0
#term3 = cds.dot2(KN, cds.dot(Se_inv, (delta_c+cds.dot(KN,(z_last-za)))))
z_next = za + np.linalg.solve(term2,term3) # same as term2\term3
dz_gpu = gpuarray.to_gpu(z_next-z_last)
temp3_gpu = linalg.dot(term2_gpu, dz_gpu)
term2_gpu.gpudata.free()
del(term2_gpu)
di2_gpu = linalg.dot(dz_gpu, temp3_gpu, transa='T')
#dz = z_next - z_last
#di2 = cds.dot3(term2, dz)
# Get out
return z_next, di2_gpu.get()
def dot_add_mm(self, a, b, out, transa=False, transb=False):
transa = 'T' if transa else 'N'
transb = 'T' if transb else 'N'
culinalg.add_dot(a, b, out, transa, transb)
def elmvis(Xraw,
A,
slowdown=10,
report=5,
maxtime=24*60*60,
tol=0,
batch=None,
maxiter=None,
maxupdate=None,
maxstall=None,
cossim=None,
silent=False):
"""ELMVIS+ function running in GPU memory.
"""
X = Xraw / np.linalg.norm(Xraw, axis=1)[:, None] # unit-length version of X
Xh = np.dot(A, X) # X_hat, predicted value of X
N, d = X.shape
I = np.arange(N) # index of samples
# set default values
if cossim is None: cossim = np.trace(X.T.dot(A).dot(X)) / N
if maxiter is None: maxiter = N*N*N
if maxupdate is None: maxupdate = N*N
if maxstall is None: maxstall = N*N
if not silent:
print "original similarity: ", cossim
# init GPU
dt = X.dtype.type
try:
linalg.init()
except ImportError as e:
print e
devA = gpuarray.to_gpu(A.astype(dt))
devX = gpuarray.to_gpu(X.astype(dt))
devXi1 = gpuarray.empty((d,), dtype=dt)
devXh = linalg.dot(devA, devX)
devAi = gpuarray.empty((N, 2), dtype=dt)
devDelta = gpuarray.empty((2, d), dtype=dt)
result = gpuarray.empty((d,), dtype=dt)
# swap kernel
kernel = """
__global__ void diff(%s *A, %s *Y, %s *AY, %s *result, long d, long N, long i1, long i2) {
long j = blockDim.x * blockIdx.x + threadIdx.x;
%s yi1 = Y[i1*d + j];
%s yi2 = Y[i2*d + j];
result[j] = (A[i1*N + i1] * (yi2 - yi1) + 2*AY[i1*d + j]) * (yi2 - yi1) +
(A[i2*N + i2] * (yi1 - yi2) + 2*(AY[i2*d + j] + A[i2*N + i1]*(yi2 - yi1))) * (yi1 - yi2);
}
"""
if dt is np.float64:
kernel = kernel % ("double", "double", "double", "double", "double", "double")
else:
kernel = kernel % ("float", "float", "float", "float", "float", "float")
mod_diff = SourceModule(kernel)
dev_diff = mod_diff.get_function("diff")
dev_diff.prepare("PPPPllll")
block = result._block
grid = (int(np.ceil(1.0 * result.shape[0] / block[0])), 1)
t0 = tlast = time()
stall = 0
iters = 0
updates = 0
updates_last = 0
iters_last = 0
ups_max = 0
while (iters < maxiter) and (stall < maxstall):
iters += 1
stall += 1
# get two different random numbers
i1, i2 = np.random.randint(0, N, size=2)
while i1 == i2:
i1, i2 = np.random.randint(0, N, size=2)
dev_diff.prepared_call(grid, block, devA.gpudata, devX.gpudata, devXh.gpudata, result.gpudata, d, N, i1, i2)
diff = np.sum(result.get())
if diff > tol:
stall = 0
devAi[:, 0] = devA[:, i1]
devAi[:, 1] = devA[:, i2]
devDelta[0, :] = devX[i1, :] - devX[i2, :]
devDelta[1, :] = devX[i2, :] - devX[i1, :]
linalg.add_dot(devAi, devDelta, devXh, alpha=-1)
tI = I[i1]
I[i1] = I[i2]
I[i2] = tI
devXi1[:] = devX[i1, :]
devX[i1] = devX[i2]
devX[i2] = devXi1
cossim += diff / N
updates += 1
if updates > maxupdate:
break
t = time()
if t - tlast > report:
ups = (updates-updates_last)*1.0/(t-tlast)
ips = (iters-iters_last)*1.0/(t-tlast)
if not silent:
print "%d iters | %d updates | %.0f iters/s | %.0f updates/s | cos similarity = %.4f" % (iters, updates, ips, ups, cossim)
updates_last = updates
iters_last = iters
tlast = t
ups_max = max(ups, ups_max)
if ups < ups_max/slowdown:
break
if t - t0 > maxtime:
break
ips = iters*1.0/(time()-t0)
ups = updates*1.0/(time()-t0)
Xraw[:] = Xraw[I]
cossim = np.trace(X.T.dot(A).dot(X)) / N
if not silent:
print "final similarity: ", cossim
info = {'cossim': cossim, 'iters': iters, 'updates': updates, 'ips': ips, 'ups': ups}
return I, info
def elmvis(Xraw,
A,
slowdown=10,
report=5,
maxtime=24 * 60 * 60,
tol=0,
batch=None,
maxiter=None,
maxupdate=None,
maxstall=None,
cossim=None,
silent=False):
"""ELMVIS+ function running in GPU memory.
"""
X = Xraw / np.linalg.norm(Xraw, axis=1)[:,
None] # unit-length version of X
Xh = np.dot(A, X) # X_hat, predicted value of X
N, d = X.shape
I = np.arange(N) # index of samples
# set default values
if cossim is None: cossim = np.trace(X.T.dot(A).dot(X)) / N
if maxiter is None: maxiter = N * N * N
if maxupdate is None: maxupdate = N * N
if maxstall is None: maxstall = N * N
if not silent:
print "original similarity: ", cossim
# init GPU
dt = X.dtype.type
try:
linalg.init()
except ImportError as e:
print e
devA = gpuarray.to_gpu(A.astype(dt))
devX = gpuarray.to_gpu(X.astype(dt))
devXi1 = gpuarray.empty((d, ), dtype=dt)
devXh = linalg.dot(devA, devX)
devAi = gpuarray.empty((N, 2), dtype=dt)
devDelta = gpuarray.empty((2, d), dtype=dt)
result = gpuarray.empty((d, ), dtype=dt)
# swap kernel
kernel = """
__global__ void diff(%s *A, %s *Y, %s *AY, %s *result, long d, long N, long i1, long i2) {
long j = blockDim.x * blockIdx.x + threadIdx.x;
%s yi1 = Y[i1*d + j];
%s yi2 = Y[i2*d + j];
result[j] = (A[i1*N + i1] * (yi2 - yi1) + 2*AY[i1*d + j]) * (yi2 - yi1) +
(A[i2*N + i2] * (yi1 - yi2) + 2*(AY[i2*d + j] + A[i2*N + i1]*(yi2 - yi1))) * (yi1 - yi2);
}
"""
if dt is np.float64:
kernel = kernel % ("double", "double", "double", "double", "double",
"double")
else:
kernel = kernel % ("float", "float", "float", "float", "float",
"float")
mod_diff = SourceModule(kernel)
dev_diff = mod_diff.get_function("diff")
dev_diff.prepare("PPPPllll")
block = result._block
grid = (int(np.ceil(1.0 * result.shape[0] / block[0])), 1)
t0 = tlast = time()
stall = 0
iters = 0
updates = 0
updates_last = 0
iters_last = 0
ups_max = 0
while (iters < maxiter) and (stall < maxstall):
iters += 1
stall += 1
# get two different random numbers
i1, i2 = np.random.randint(0, N, size=2)
while i1 == i2:
i1, i2 = np.random.randint(0, N, size=2)
dev_diff.prepared_call(grid, block, devA.gpudata, devX.gpudata,
devXh.gpudata, result.gpudata, d, N, i1, i2)
diff = np.sum(result.get())
if diff > tol:
stall = 0
devAi[:, 0] = devA[:, i1]
devAi[:, 1] = devA[:, i2]
devDelta[0, :] = devX[i1, :] - devX[i2, :]
devDelta[1, :] = devX[i2, :] - devX[i1, :]
linalg.add_dot(devAi, devDelta, devXh, alpha=-1)
tI = I[i1]
I[i1] = I[i2]
I[i2] = tI
devXi1[:] = devX[i1, :]
devX[i1] = devX[i2]
devX[i2] = devXi1
cossim += diff / N
updates += 1
if updates > maxupdate:
break
t = time()
if t - tlast > report:
ups = (updates - updates_last) * 1.0 / (t - tlast)
ips = (iters - iters_last) * 1.0 / (t - tlast)
if not silent:
print "%d iters | %d updates | %.0f iters/s | %.0f updates/s | cos similarity = %.4f" % (
iters, updates, ips, ups, cossim)
updates_last = updates
iters_last = iters
tlast = t
ups_max = max(ups, ups_max)
if ups < ups_max / slowdown:
break
if t - t0 > maxtime:
break
ips = iters * 1.0 / (time() - t0)
ups = updates * 1.0 / (time() - t0)
Xraw[:] = Xraw[I]
cossim = np.trace(X.T.dot(A).dot(X)) / N
if not silent:
print "final similarity: ", cossim
info = {
'cossim': cossim,
'iters': iters,
'updates': updates,
'ips': ips,
'ups': ups
}
return I, info