def test_init(self):
import numpy as np
import math
import sys
assert np.intp() == np.intp(0)
assert np.intp("123") == np.intp(123)
raises(TypeError, np.intp, None)
assert np.float64() == np.float64(0)
assert math.isnan(np.float64(None))
assert np.bool_() == np.bool_(False)
assert np.bool_("abc") == np.bool_(True)
assert np.bool_(None) == np.bool_(False)
assert np.complex_() == np.complex_(0)
# raises(TypeError, np.complex_, '1+2j')
assert math.isnan(np.complex_(None))
for c in ["i", "I", "l", "L", "q", "Q"]:
assert np.dtype(c).type().dtype.char == c
for c in ["l", "q"]:
assert np.dtype(c).type(sys.maxint) == sys.maxint
for c in ["L", "Q"]:
assert np.dtype(c).type(sys.maxint + 42) == sys.maxint + 42
assert np.float32(np.array([True, False])).dtype == np.float32
assert type(np.float32(np.array([True]))) is np.ndarray
assert type(np.float32(1.0)) is np.float32
a = np.array([True, False])
assert np.bool_(a) is a
def getrows(self, privateKey, senderID, mType, params, extra):
"""
Receive a single row or list of rows and store in a class property variable
Use **client.crow** or **client.rowlist** to access the indices of
**previously broadcasted** rows
"""
filen = params['url']
if mType == 'table.highlight.row':
idx = np.intp(params['row'])
print '[SAMP] Selected row %s from %s' % (idx,filen)
print '[SAMP] Row index stored in property -> crow'
self.crow = idx
elif mType == 'table.select.rowList':
idx = np.intp(params['row-list'])
print '[SAMP] Selected %s rows from %s' % (len(idx),filen)
print '[SAMP] List stored in property -> rowlist'
self.rowlist = idx
self.lastMessage = {'label':'Selected Rows',
'privateKey':privateKey,
'senderID': senderID,
'mType': mType,
'params': params,
'extra': extra }
def __init__(self, ldis, Nv, VX, VY, K, EToV):
l = self.ldis = ldis
self.dimensions = ldis.dimensions
self.Nv = Nv
self.VX = VX
self.K = K
va = np.intp(EToV[:, 0].T)
vb = np.intp(EToV[:, 1].T)
vc = np.intp(EToV[:, 2].T)
x = self.x = 0.5*(
-np.outer(VX[va], l.r+l.s, )
+np.outer(VX[vb], 1+l.r)
+np.outer(VX[vc], 1+l.s))
y = self.y = 0.5*(
-np.outer(VY[va], l.r+l.s)
+np.outer(VY[vb], 1+l.r)
+np.outer(VY[vc], 1+l.s))
self.rx, self.sx, self.ry, self.sy, self.J = GeometricFactors2D(x, y, l.Dr, l.Ds)
self.nx, self.ny, self.sJ = Normals2D(l, x, y, K)
self.Fscale = self.sJ/self.J[:, l.FmaskF]
# element-to-element, element-to-face connectivity
self.EToE, self.EToF = Connect2D(EToV)
self.mapM, self.mapP, self.vmapM, self.vmapP, self.vmapB, self.mapB = \
BuildMaps2D(l, l.Fmask, VX, VY, EToV, self.EToE, self.EToF, K, l.N, x, y)
def test_init(self):
import numpy as np
import math
import sys
assert np.intp() == np.intp(0)
assert np.intp('123') == np.intp(123)
raises(TypeError, np.intp, None)
assert np.float64() == np.float64(0)
assert math.isnan(np.float64(None))
assert np.bool_() == np.bool_(False)
assert np.bool_('abc') == np.bool_(True)
assert np.bool_(None) == np.bool_(False)
assert np.complex_() == np.complex_(0)
#raises(TypeError, np.complex_, '1+2j')
assert math.isnan(np.complex_(None))
for c in ['i', 'I', 'l', 'L', 'q', 'Q']:
assert np.dtype(c).type().dtype.char == c
for c in ['l', 'q']:
assert np.dtype(c).type(sys.maxint) == sys.maxint
for c in ['L', 'Q']:
assert np.dtype(c).type(sys.maxint + 42) == sys.maxint + 42
assert np.float32(np.array([True, False])).dtype == np.float32
assert type(np.float32(np.array([True]))) is np.ndarray
assert type(np.float32(1.0)) is np.float32
a = np.array([True, False])
assert np.bool_(a) is a
def execute(self, solver, stream=None):
slvr = solver
# The gaussian shape array can be empty if
# no gaussian sources were specified.
gauss = np.intp(0) if np.product(slvr.gauss_shape.shape) == 0 \
else slvr.gauss_shape
sersic = np.intp(0) if np.product(slvr.sersic_shape.shape) == 0 \
else slvr.sersic_shape
self.kernel(slvr.uvw, slvr.brightness, gauss, sersic,
slvr.wavelength, slvr.antenna1, slvr.antenna2,
slvr.jones_scalar,
slvr.flag, slvr.weight_vector,
slvr.model_vis, slvr.observed_vis, slvr.chi_sqrd_result,
**self.get_kernel_params(slvr))
# Call the pycuda reduction kernel.
# Divide by the single sigma squared value if a weight vector
# is not required. Otherwise the kernel will incorporate the
# individual sigma squared values into the sum
gpu_sum = gpuarray.sum(slvr.chi_sqrd_result).get()
if not self.weight_vector:
slvr.set_X2(gpu_sum/slvr.sigma_sqrd)
else:
slvr.set_X2(gpu_sum)
def from_range(cat_comp, lo, hi):
"""
Utility function to help construct the Roi from a range.
:param cat_comp: Anything understood by ._categorical_helper ... array, list or component
:param lo: lower bound of the range
:param hi: upper bound of the range
:return: CategoricalROI object
"""
# Convert lo and hi to integers. Note that if lo or hi are negative,
# which can happen if the user zoomed out, we need to reset the to zero
# otherwise they will have strange effects when slicing the categories.
# Note that we used ceil for lo, because if lo is 0.9 then we should
# only select 1 and above.
lo = np.intp(np.ceil(lo) if lo > 0 else 0)
hi = np.intp(np.ceil(hi) if hi > 0 else 0)
roi = CategoricalROI()
cat_data = cat_comp.categories
roi.update_categories(cat_data[lo:hi])
return roi
def Connect3D(EToV):
"""Build global connectivity arrays for grid based on
standard EToV input array from grid generator.
"""
EToV = EToV.astype(np.intp)
Nfaces = 4
# Find number of elements and vertices
K = EToV.shape[0]
#Nv = EToV.max()+1
# Create face to node connectivity matrix
#TotalFaces = Nfaces*K
# List of local face to local vertex connections
vn = np.int32([[0,1,2],[0,1,3],[1,2,3],[0,2,3]])
# Build global face to node connectivity
g_face_no = 0
vert_indices_to_face_numbers = {}
face_numbers = xrange(Nfaces)
for k in xrange(K):
for face in face_numbers:
vert_indices_to_face_numbers.setdefault(
frozenset(EToV[k,vn[face]]), []).append(g_face_no)
g_face_no += 1
faces1 = []
faces2 = []
# check this
for i in vert_indices_to_face_numbers.itervalues():
if len(i) == 2:
faces1.append(i[0])
faces2.append(i[1])
faces2.append(i[0])
faces1.append(i[1])
faces1 = np.intp(faces1)
faces2 = np.intp(faces2)
# Convert faceglobal number to element and face numbers
element1, face1 = divmod(faces1, Nfaces)
element2, face2 = divmod(faces2, Nfaces)
# Rearrange into Nelements x Nfaces sized arrays
ind = element1*Nfaces + face1
EToE = np.outer(np.arange(K), np.ones((1, Nfaces)))
EToF = np.outer(np.ones((K,1)), np.arange(Nfaces))
EToE = EToE.reshape(K*Nfaces)
EToF = EToF.reshape(K*Nfaces)
EToE[np.int32(ind)] = element2
EToF[np.int32(ind)] = face2
EToE = EToE.reshape(K, Nfaces)
EToF = EToF.reshape(K, Nfaces)
return EToE, EToF
def test_intp(self):
# Ticket #99
i_width = np.int_(0).nbytes*2 - 1
np.intp('0x' + 'f'*i_width, 16)
assert_raises(OverflowError, np.intp, '0x' + 'f'*(i_width+1), 16)
assert_raises(ValueError, np.intp, '0x1', 32)
assert_equal(255, np.intp('0xFF', 16))
def _accumDiffStims(self, d_resp_tmp, diffV1GausBuf, sizes, orderX,
orderY, orderT):
""" Gets the responses of the filters specified in d_v1popDirs by
interpolation.
This is basically what shSwts.m did in the original S&H code."""
# a useful list of factorials for computing the scaling factors for
# the derivatives
factorials = (1, 1, 2, 6)
# the scaling factor for this directional derivative
# similar to the binomial coefficients
scale = 6/factorials[orderX]/factorials[orderY]/factorials[orderT]
gdim = (int(iDivUp(sizes[0] * sizes[1], 256)), 1)
bdim = (256, 1, 1)
self.dev_accumDiffStims(
np.intp(d_resp_tmp),
np.intp(diffV1GausBuf),
np.int32(sizes[0] * sizes[1]),
np.int32(scale),
np.int32(orderX),
np.int32(orderY),
np.int32(orderT),
block=bdim, grid=gdim)
def __init__(self, code, point, struct_ptr):
self.code = cuda.to_device(code)
self.point = cuda.to_device(point)
self.code_shape, self.code_dtype = code.shape, code.dtype
self.point_shape, self.point_dtype = point.shape, point.dtype
cuda.memcpy_htod(int(struct_ptr), np.int32(code.size))
cuda.memcpy_htod(int(struct_ptr) + 8, np.intp(int(self.code)))
cuda.memcpy_htod(int(struct_ptr) + 8 + np.intp(0).nbytes, np.intp(int(self.point)))
def check_intp(self,level=rlevel):
"""Ticket #99"""
i_width = N.int_(0).nbytes*2 - 1
N.intp('0x' + 'f'*i_width,16)
self.failUnlessRaises(OverflowError,N.intp,'0x' + 'f'*(i_width+1),16)
self.failUnlessRaises(ValueError,N.intp,'0x1',32)
assert_equal(255,N.intp('0xFF',16))
assert_equal(1024,N.intp(1024))
def test_intp(self,level=rlevel):
"""Ticket #99"""
i_width = np.int_(0).nbytes*2 - 1
np.intp('0x' + 'f'*i_width,16)
self.assertRaises(OverflowError,np.intp,'0x' + 'f'*(i_width+1),16)
self.assertRaises(ValueError,np.intp,'0x1',32)
assert_equal(255,np.intp('0xFF',16))
assert_equal(1024,np.intp(1024))
def getSRT(self, gmap, store_srt=False):
"""
Computes the sample rank templates for the expression matrix(on instantiation) and
gmap
gmap is a 1d numpy array where gmap[2*i] and gmap[2*i +1] are
gene indices for comparison i
b_size is the block size to use in gpu computation
store_srt - determines what is returned,
False(default) - returns the srt numpy array (npairs,nsamp)
True - returns the srt_gpu object and the object's padded shape (npairs,nsamp)
"""
#the x coords in the gpu map to sample_ids
#the y coords to gmap
#sample blocks
b_size = self.b_size
exp = self.exp
g_y_sz = self.getGrid( exp.shape[1] )
#pair blocks
g_x_sz = self.getGrid( gmap.shape[0]/2 )
#put gene map on gpu
gmap_buffer = self.gmap_buffer= self.getBuff(gmap, 2*(g_x_sz*b_size), 1,np.int32)
gmap_gpu = np.intp(gmap_buffer.base.get_device_pointer()) #cuda.mem_alloc(gmap_buffer.nbytes)
#cuda.memcpy_htod(gmap_gpu,gmap_buffer)
#make room for srt
srt_shape = (g_x_sz*b_size , g_y_sz*b_size)
srt_buffer = self.srt_buffer = self.getBuff(np.zeros(srt_shape, dtype=np.int32),srt_shape[0],srt_shape[1], np.int32)
srt_gpu = np.intp(srt_buffer.base.get_device_pointer()) #cuda.mem_alloc(srt_shape[0]*srt_shape[1]*np.int32(1).nbytes)
srtKern = self.getsrtKern()
exp_gpu = self.exp_gpu
nsamp = np.uint32( g_y_sz * b_size )
ngenes = np.uint32( self.exp.shape[0] )
npairs = np.uint32( g_x_sz * b_size )
block = (b_size,b_size,1)
grid = (g_x_sz, g_y_sz)
srtKern(exp_gpu, nsamp, ngenes, gmap_gpu, npairs, srt_gpu, block=block, grid=grid)
#gmap_gpu.free()
if store_srt:
#this is in case we want to run further stuff without
#transferring back and forth
return (srt_gpu, npairs , nsamp)
else:
#srt_buffer = np.zeros(srt_shape, dtype=np.int32)
#cuda.memcpy_dtoh(srt_buffer, srt_gpu)
#srt_gpu.free()
return srt_buffer[:gmap.shape[0]/2,:self.exp.shape[1]]
def _default_norm(self, layer):
vals = np.sort(layer.ravel())
vals = vals[np.isfinite(vals)]
result = DS9Normalize()
result.stretch = 'arcsinh'
result.clip = True
if vals.size > 0:
result.vmin = vals[np.intp(.01 * vals.size)]
result.vmax = vals[np.intp(.99 * vals.size)]
return result
def _loadInput(self, stim):
logging.debug('loadInput')
# shortcuts
nrXY = self.nrX * self.nrY
nrXYD = self.nrX * self.nrY * self.nrDirs
# parse input
assert type(stim).__module__ == "numpy", "stim must be numpy array"
assert type(stim).__name__ == "ndarray", "stim must be numpy.ndarray"
assert stim.size > 0, "stim cannot be []"
stim = stim.astype(np.ubyte)
rows, cols = stim.shape
logging.debug("- stim shape={0}x{1}".format(rows, cols))
# shift d_stimBuf in time by 1 frame, from frame i to frame i-1
# write our own memcpy kernel... :-(
gdim = (int(iDivUp(nrXY, 128)), 1)
bdim = (128, 1, 1)
for i in xrange(1, self.nrT):
stimBufPt_dst = np.intp(self.d_stimBuf) + self.szXY * (i - 1)
stimBufPt_src = np.intp(self.d_stimBuf) + self.szXY * i
self.dev_memcpy_dtod(
stimBufPt_dst,
stimBufPt_src,
np.int32(nrXY),
block=bdim, grid=gdim)
# index into d_stimBuf array to place the new stim at the end
# (newest frame at pos: nrT-1)
d_stimBufPt = np.intp(self.d_stimBuf) + self.szXY * (self.nrT-1)
# \TODO implement RGB support
self.dev_split_gray(
d_stimBufPt,
cuda.In(stim),
np.int32(stim.size),
block=bdim, grid=gdim)
# create working copy of d_stimBuf
cuda.memcpy_dtod(self.d_scalingStimBuf, self.d_stimBuf,
self.szXY*self.nrT)
# reset V1complex responses to 0
# \FIXME not sure how to use memset...doesn't seem to give expected
# result
tmp = np.zeros(nrXYD).astype(np.float32)
cuda.memcpy_htod(self.d_respV1c, tmp)
# allocate d_resp, which will contain the response to all 28
# (nrFilters) space-time orientations at 3 (nrScales) scales for
# every pixel location (nrX*nrY)
tmp = np.zeros(nrXY*self.nrFilters*self.nrScales).astype(np.float32)
cuda.memcpy_htod(self.d_resp, tmp)
def readout(mesh, pos, mode="raise", period=None, transform=None, out=None):
""" CIC approximation, reading out mesh values at pos,
see document of paint.
"""
pos = numpy.array(pos)
if out is None:
out = numpy.zeros(len(pos), dtype='f8')
else:
out[:] = 0
chunksize = 1024 * 16 * 4
Ndim = pos.shape[-1]
Np = pos.shape[0]
if transform is None:
transform = lambda x: x
neighbours = ((numpy.arange(2 ** Ndim)[:, None] >> \
numpy.arange(Ndim)[None, :]) & 1)
for start in range(0, Np, chunksize):
chunk = slice(start, start+chunksize)
if mode == 'raise':
gridpos = transform(pos[chunk])
rmi_mode = 'raise'
intpos = numpy.intp(numpy.floor(gridpos))
elif mode == 'ignore':
gridpos = transform(pos[chunk])
rmi_mode = 'raise'
intpos = numpy.intp(numpy.floor(gridpos))
for i, neighbour in enumerate(neighbours):
neighbour = neighbour[None, :]
targetpos = intpos + neighbour
kernel = (1.0 - numpy.abs(gridpos - targetpos)).prod(axis=-1)
if period is not None:
period = numpy.int32(period)
numpy.remainder(targetpos, period, targetpos)
if mode == 'ignore':
# filter out those outside of the mesh
mask = (targetpos >= 0).all(axis=-1)
for d in range(Ndim):
mask &= (targetpos[..., d] < mesh.shape[d])
targetpos = targetpos[mask]
kernel = kernel[mask]
else:
mask = Ellipsis
if len(targetpos) > 0:
targetindex = numpy.ravel_multi_index(
targetpos.T, mesh.shape, mode=rmi_mode)
out[chunk][mask] += kernel * mesh.flat[targetindex]
return out
def getRT(self, s_map, srt_gpu, srt_nsamp, srt_npairs, npairs, store_rt=False):
"""
Computes the rank template
s_map(Sample Map) - an list of 1s and 0s of length nsamples where 1 means use this sample
to compute rank template
srt_gpu - cuda memory object containing srt(sample rank template) array on gpu
srt_nsamp, srt_npairs - shape(buffered) of srt_gpu object
npairs - true number of gene pairs being compared
b_size - size of the blocks for computation
store_rt - determines the RETURN value
False(default) = returns an numpy array shape(npairs) of the rank template
True = returns the rt_gpu object and the padded size of the rt_gpu objet (rt_obj, npairs_padded)
"""
b_size = self.b_size
s_map_buff = self.s_map_buff = cuda.pagelocked_zeros((int(srt_nsamp),), np.int32, mem_flags=cuda.host_alloc_flags.DEVICEMAP)
s_map_buff[:len(s_map)] = np.array(s_map,dtype=np.int32)
s_map_gpu = np.intp(s_map_buff.base.get_device_pointer())
#cuda.memcpy_htod(s_map_gpu, s_map_buff)
#sample blocks
g_y_sz = self.getGrid( srt_nsamp)
#pair blocks
g_x_sz = self.getGrid( srt_npairs )
block_rt_gpu = cuda.mem_alloc(int(g_y_sz*srt_npairs*(np.uint32(1).nbytes)) )
grid = (g_x_sz, g_y_sz)
func1,func2 = self.getrtKern(g_y_sz)
shared_size = b_size*b_size*np.uint32(1).nbytes
func1( srt_gpu, np.uint32(srt_nsamp), np.uint32(srt_npairs), s_map_gpu, block_rt_gpu, np.uint32(g_y_sz), block=(b_size,b_size,1), grid=grid, shared=shared_size)
rt_buffer =self.rt_buffer = cuda.pagelocked_zeros((int(srt_npairs),), np.int32, mem_flags=cuda.host_alloc_flags.DEVICEMAP)
rt_gpu = np.intp(rt_buffer.base.get_device_pointer())
func2( block_rt_gpu, rt_gpu, np.int32(s_map_buff.sum()), block=(b_size,1,1), grid=(g_x_sz,))
if store_rt:
#this is in case we want to run further stuff without
#transferring back and forth
return (rt_gpu, srt_npairs)
else:
#rt_buffer = np.zeros((srt_npairs ,), dtype=np.int32)
#cuda.memcpy_dtoh(rt_buffer, rt_gpu)
#rt_gpu.free()
return rt_buffer[:npairs]
def __init__(self, ldis, Nv, VX, VY, VZ, K, EToV):
l = self.ldis = ldis
self.dimensions = ldis.dimensions
self.Nv = Nv
self.VX = VX
self.VY = VY; # check ?
self.VZ = VZ
self.K = K
va = np.intp(EToV[:, 0].T)
vb = np.intp(EToV[:, 1].T)
vc = np.intp(EToV[:, 2].T)
vd = np.intp(EToV[:, 3].T)
x = self.x = 0.5*(
-np.outer(VX[va], 1+l.r+l.s+l.t, )
+np.outer(VX[vb], 1+l.r)
+np.outer(VX[vc], 1+l.s)
+np.outer(VX[vd], 1+l.t))
y = self.y = 0.5*(
-np.outer(VY[va], 1+l.r+l.s+l.t, )
+np.outer(VY[vb], 1+l.r)
+np.outer(VY[vc], 1+l.s)
+np.outer(VY[vd], 1+l.t))
z = self.z = 0.5*(
-np.outer(VZ[va], 1+l.r+l.s+l.t, )
+np.outer(VZ[vb], 1+l.r)
+np.outer(VZ[vc], 1+l.s)
+np.outer(VZ[vd], 1+l.t))
drst_dxyz = np.empty((3,3), dtype=object)
geo_fac = GeometricFactors3D(x, y, z, l.Dr, l.Ds, l.Dt)
self.J = geo_fac[-1]
drst_dxyz.reshape(-1)[:] = geo_fac[:-1]
self.drst_dxyz = drst_dxyz.T
self.nx, self.ny, self.nz, self.sJ = Normals3D(l, x, y, z, K)
self.Fscale = self.sJ/self.J[:, l.FmaskF]
# element-to-element, element-to-face connectivity
self.EToE, self.EToF = Connect3D(EToV)
self.mapM, self.mapP, self.vmapM, self.vmapP, self.vmapB, self.mapB = \
BuildMaps3D(l, l.Fmask, VX, VY, VZ, EToV, self.EToE, self.EToF, K, l.N, x, y, z)
self.bc = np.ones((K, l.Nfp*l.Nfaces))
self.bc.reshape(-1)[self.mapB] = -1
def coarsegrain(self,flowy,deltafy,Ny):
if not hasattr(self,'Moments'):
self.getMultipleMoments(msign='pn')
Nx = self.Length
flowx = self.FreqOffset
deltafx = self.Cadence
if ( (deltafx <= 0) | (deltafy <= 0) | (Ny <= 0) ):
raise ValueError, 'bad input argument'
if ( deltafy < deltafx ):
raise ValueError, 'deltaf coarse-grain < deltaf fine-grain'
if ( (flowy - 0.5*deltafy) < (flowx - 0.5*deltafx) ):
raise ValueError, 'desired coarse-grained start frequency is too low'
fhighx = flowx + (Nx-1)*deltafx
fhighy = flowy + (Ny-1)*deltafy
if ( (fhighy + 0.5*deltafy) > (fhighx + 0.5*deltafx) ):
raise ValueError, 'desired coarse-rained stop frequency is too high'
i = numpy.arange(Ny)
jlow = numpy.intp(
1 + numpy.floor((flowy + (i-0.5)*deltafy - flowx - 0.5*deltafx)/deltafx))
jhigh = numpy.intp(
1 + numpy.floor((flowy + (i+0.5)*deltafy - flowx - 0.5*deltafx)/deltafx))
index1 = jlow[0]
index2 = jhigh[-1]
fraclow = (flowx + (jlow+0.5)*deltafx - flowy - (i-0.5)*deltafy)/deltafx
frachigh = (flowy + (i+0.5)*deltafy - flowx - (jhigh-0.5)*deltafx)/deltafx
frac1 = fraclow[0]
frac2 = frachigh[-1]
jtemp = jlow + 1
coarseMoments = numpy.zeros( (numpy.shape(self.Moments)[0],Ny) , complex)
for lm in range(numpy.shape(self.Moments)[0]):
midsum = sumTerms(self.Moments[lm,:], jtemp, jhigh)
ya = (deltafx/deltafy)*(self.Moments[lm,:][jlow[:-1]]*fraclow[:-1] +
self.Moments[lm,:][jhigh[:-1]]*frachigh[:-1] +
midsum[:-1])
if (jhigh[-1] > Nx-1):
yb = (deltafx/deltafy)*(self.Moments[lm,:][jlow[-1]]*fraclow[-1] + midsum[-1])
else:
yb = (deltafx/deltafy)*(self.Moments[lm,:][jlow[-1]]*fraclow[-1] +
self.Moments[lm,:][jhigh[-1]]*frachigh[-1] +
midsum[-1])
coarseMoments[lm,:] = numpy.array( list(ya) + [yb] )
self.coarseMoments = coarseMoments
self.coarseFreqOffset = flowy
self.coarseCadence = deltafy
self.coarseLength = numpy.shape(self.coarseMoments)[1]
self.index1 = index1
self.index2 = index2
self.frac1 = frac1
self.frac2 = frac2
return coarseMoments
def rfGbmCombined(X, Y_casual, Y_registered, testSet_final): #creating models rf1 = randomForestModel() #train for casual rf2 = randomForestModel() #train for registered gbm1 = gradientDescentModel() #train for casual gbm2 = gradientDescentModel() #train for registered #fitting models # rf1.fit(train_X, train_Y[:, 0]) #train_Y[:, 0] - use 0th column of train_Y rf1.fit(X, Y_casual) plotFeatureImportance(rf1) rf2.fit(X, Y_registered) gbm1.fit(X, Y_casual) gbm2.fit(X, Y_registered) #prediction rf1_Y = np.exp(rf1.predict(testSet_final))-1 rf2_Y = np.exp(rf2.predict(testSet_final))-1 gbm1_Y = np.exp(gbm1.predict(testSet_final))-1 gbm2_Y = np.exp(gbm2.predict(testSet_final))-1 #Average the prediction from classifiers final_prediction = (rf1_Y + rf2_Y + gbm1_Y + gbm2_Y)/2 final_prediction = np.intp(np.around(final_prediction)) #round and convert to integer return final_prediction
def __new__(cls, x=0):
if isinstance(x, afnumpy.ndarray):
return x.astype(cls)
elif isinstance(x, numbers.Number):
return numpy.intp(x)
else:
return afnumpy.array(x).astype(cls)
def _build_arg_buf(args):
handlers = []
arg_data = []
format = ""
for i, arg in enumerate(args):
if isinstance(arg, np.number):
arg_data.append(arg)
format += arg.dtype.char
elif isinstance(arg, (DeviceAllocation, PooledDeviceAllocation)):
arg_data.append(int(arg))
format += "P"
elif isinstance(arg, ArgumentHandler):
handlers.append(arg)
arg_data.append(int(arg.get_device_alloc()))
format += "P"
elif isinstance(arg, np.ndarray):
arg_data.append(arg)
format += "%ds" % arg.nbytes
else:
try:
gpudata = np.intp(arg.gpudata)
except AttributeError:
raise TypeError("invalid type on parameter #%d (0-based)" % i)
else:
# for gpuarrays
arg_data.append(int(gpudata))
format += "P"
from pycuda._pvt_struct import pack
return handlers, pack(format, *arg_data)
def small_view_array(data):
"""
Same as small_view, except using a numpy array as input
"""
shp = data.shape
view = tuple([slice(None, None, np.intp(max(s / 50, 1))) for s in shp])
return np.asarray(data)[view]
def test_simple_kernel_2(self):
mod = SourceModule("""
__global__ void multiply_them(float *dest, float *a, float *b)
{
const int i = threadIdx.x;
dest[i] = a[i] * b[i];
}
""")
multiply_them = mod.get_function("multiply_them")
a = np.random.randn(400).astype(np.float32)
b = np.random.randn(400).astype(np.float32)
a_gpu = drv.to_device(a)
b_gpu = drv.to_device(b)
dest = np.zeros_like(a)
multiply_them(
drv.Out(dest), a_gpu, b_gpu,
block=(400, 1, 1))
assert la.norm(dest-a*b) == 0
drv.Context.synchronize()
# now try with offsets
dest = np.zeros_like(a)
multiply_them(
drv.Out(dest), np.intp(a_gpu)+a.itemsize, b_gpu,
block=(399, 1, 1))
assert la.norm((dest[:-1]-a[1:]*b[:-1])) == 0
def test_numpy(self):
"""NumPy objects get serialized to readable JSON."""
l = [
np.float32(12.5),
np.float64(2.0),
np.float16(0.5),
np.bool(True),
np.bool(False),
np.bool_(True),
np.unicode_("hello"),
np.byte(12),
np.short(12),
np.intc(-13),
np.int_(0),
np.longlong(100),
np.intp(7),
np.ubyte(12),
np.ushort(12),
np.uintc(13),
np.ulonglong(100),
np.uintp(7),
np.int8(1),
np.int16(3),
np.int32(4),
np.int64(5),
np.uint8(1),
np.uint16(3),
np.uint32(4),
np.uint64(5),
]
l2 = [l, np.array([1, 2, 3])]
roundtripped = loads(dumps(l2, cls=EliotJSONEncoder))
self.assertEqual([l, [1, 2, 3]], roundtripped)
def fast_limits(data, plo, phi):
"""
Quickly estimate percentiles in an array, using a downsampled version
Parameters
----------
data : `numpy.ndarray`
The array to estimate the percentiles for
plo, phi : float
The percentile values
Returns
-------
lo, hi : float
The percentile values
"""
shp = data.shape
view = tuple([slice(None, None, np.intp(max(s / 50, 1))) for s in shp])
values = np.asarray(data)[view]
if ~np.isfinite(values).any():
return (0.0, 1.0)
limits = (-np.inf, np.inf)
lo = _scoreatpercentile(values.flat, plo, limit=limits)
hi = _scoreatpercentile(values.flat, phi, limit=limits)
return lo, hi
def small_view(data, attribute):
"""
Extract a downsampled view from a dataset, for quick statistical summaries
"""
shp = data.shape
view = tuple([slice(None, None, np.intp(max(s / 50, 1))) for s in shp])
return data[attribute, view]
def coarsegrain(self,flowy,deltafy,Ny):
if not hasattr(self,'CAE'):
self.CrossSpectraDataAE()
Nx = self.FreqLength
flowx = self.FreqOffset
deltafx = self.FreqCadence
if ( (deltafx <= 0) | (deltafy <= 0) | (Ny <= 0) ):
raise ValueError, 'bad input argument'
if ( deltafy < deltafx ):
raise ValueError, 'deltaf coarse-grain < deltaf fine-grain'
if ( (flowy - 0.5*deltafy) < (flowx - 0.5*deltafx) ):
raise ValueError, 'desired coarse-grained start frequency is too low'
fhighx = flowx + (Nx-1)*deltafx
fhighy = flowy + (Ny-1)*deltafy
if ( (fhighy + 0.5*deltafy) > (fhighx + 0.5*deltafx) ):
raise ValueError, 'desired coarse-grained stop frequency is too high'
i = numpy.arange(Ny)
jlow = numpy.intp(
1 + numpy.floor((flowy + (i-0.5)*deltafy - flowx - 0.5*deltafx)/deltafx))
jhigh = numpy.intp(
1 + numpy.floor((flowy + (i+0.5)*deltafy - flowx - 0.5*deltafx)/deltafx))
index1 = jlow[0]
index2 = jhigh[-1]
fraclow = (flowx + (jlow+0.5)*deltafx - flowy - (i-0.5)*deltafy)/deltafx
frachigh = (flowy + (i+0.5)*deltafy - flowx - (jhigh-0.5)*deltafx)/deltafx
frac1 = fraclow[0]
frac2 = frachigh[-1]
jtemp = jlow + 1
midsum = sumTerms(self.CAE, jtemp, jhigh)
ya = (deltafx/deltafy)*(self.CAE[jlow[:-1]]*fraclow[:-1] +
self.CAE[jhigh[:-1]]*frachigh[:-1] +
midsum[:-1])
if (jhigh[-1] > Nx-1):
yb = (deltafx/deltafy)*(self.CAE[jlow[-1]]*fraclow[-1] + midsum[-1])
else:
yb = (deltafx/deltafy)*(self.CAE[jlow[-1]]*fraclow[-1] +
self.CAE[jhigh[-1]]*frachigh[-1] +
midsum[-1])
self.coarseCAE = numpy.array( list(ya) + [yb] )
self.coarseFreqOffset = flowy
self.coarseFreqCadence = deltafy
self.coarseFreqLength = len(self.coarseCAE)
self.index1 = index1
self.index2 = index2
self.frac1 = frac1
self.frac2 = frac2
return self.coarseCAE
def test_more_barycentric_transforms(self):
# Triangulate some "nasty" grids
eps = np.finfo(float).eps
npoints = {2: 70, 3: 11, 4: 5, 5: 3}
_is_32bit_platform = np.intp(0).itemsize < 8
for ndim in xrange(2, 6):
# Generate an uniform grid in n-d unit cube
x = np.linspace(0, 1, npoints[ndim])
grid = np.c_[list(map(np.ravel, np.broadcast_arrays(*np.ix_(*([x]*ndim)))))].T
err_msg = "ndim=%d" % ndim
# Check using regular grid
tri = qhull.Delaunay(grid)
self._check_barycentric_transforms(tri, err_msg=err_msg,
unit_cube=True)
# Check with eps-perturbations
np.random.seed(1234)
m = (np.random.rand(grid.shape[0]) < 0.2)
grid[m,:] += 2*eps*(np.random.rand(*grid[m,:].shape) - 0.5)
tri = qhull.Delaunay(grid)
self._check_barycentric_transforms(tri, err_msg=err_msg,
unit_cube=True,
unit_cube_tol=2*eps)
# Check with duplicated data
tri = qhull.Delaunay(np.r_[grid, grid])
self._check_barycentric_transforms(tri, err_msg=err_msg,
unit_cube=True,
unit_cube_tol=2*eps)
if not _is_32bit_platform:
# test numerically unstable, and reported to fail on 32-bit
# installs
# Check with larger perturbations
np.random.seed(4321)
m = (np.random.rand(grid.shape[0]) < 0.2)
grid[m,:] += 1000*eps*(np.random.rand(*grid[m,:].shape) - 0.5)
tri = qhull.Delaunay(grid)
self._check_barycentric_transforms(tri, err_msg=err_msg,
unit_cube=True,
unit_cube_tol=1500*eps)
# Check with yet larger perturbations
np.random.seed(4321)
m = (np.random.rand(grid.shape[0]) < 0.2)
grid[m,:] += 1e6*eps*(np.random.rand(*grid[m,:].shape) - 0.5)
tri = qhull.Delaunay(grid)
self._check_barycentric_transforms(tri, err_msg=err_msg,
unit_cube=True,
unit_cube_tol=1e7*eps)
def test_64bit_integer(self):
a = scipy.sparse.csr_matrix(array([[2**32+1, 2**32+1],
[-2**63+2, 2**63-2]],
dtype=np.int64))
if (np.intp(0).itemsize < 8):
assert_raises(OverflowError, mmwrite, self.fn, a)
else:
self.check_exact(a, (2, 2, 4, 'coordinate', 'integer', 'general'))
def max_pool_forward_naive(x, pool_param):
"""
A naive implementation of the forward pass for a max pooling layer.
Inputs:
- x: Input data, of shape (N, C, H, W)
- pool_param: dictionary with the following keys:
- 'pool_height': The height of each pooling region
- 'pool_width': The width of each pooling region
- 'stride': The distance between adjacent pooling regions
Returns a tuple of:
- out: Output data
- cache: (x, pool_param)
"""
out = None
###########################################################################
# TODO: Implement the max pooling forward pass #
###########################################################################
N, C, H, W = x.shape
HH, WW, stride = pool_param['pool_height'], pool_param[
'pool_width'], pool_param['stride']
HP = np.intp(1 + (H - HH) / stride)
WP = np.intp(1 + (W - WW) / stride)
hs = stride * np.arange(HP)
ws = stride * np.arange(WP)
out = np.zeros((N, C, HP, WP))
for i in range(HP):
for j in range(WP):
out[:, :, i, j] = np.amax(x[:, :, hs[i]:hs[i] + HH,
ws[j]:ws[j] + WW],
axis=(2, 3))
###########################################################################
# END OF YOUR CODE #
###########################################################################
cache = (x, pool_param)
return out, cache
def control_signed(a, b):
tp = self.get_numpy_signed_upcast(a, b)
if b >= 0:
return tp(a)**tp(b)
else:
inv = tp(a)**tp(-b)
if inv == 0:
# Overflow
return 0
return np.intp(1.0 / inv)
def _unique_internal(data):
if len(data.shape) != 1:
raise ValueError("_unique_internal currently "
"only supports 1D arrays")
# Handle the empty array case
if data.shape[0] == 0:
return (data,
np.empty((0,), dtype=np.intp),
np.empty((0,), dtype=np.intp),
np.empty((0,), dtype=np.intp))
# See numpy's unique1d
perm = np.argsort(data, kind='mergesort')
# Combine these arrays to save on allocations?
aux = np.empty_like(data)
mask = np.empty(aux.shape, dtype=np.bool_)
inv_idx = np.empty(mask.shape, dtype=np.intp)
# Hard code first iteration
p = perm[0]
aux[0] = data[p]
mask[0] = True
cumsum = 1
inv_idx[p] = cumsum - 1
counts = [np.intp(0)]
for i in range(1, aux.shape[0]):
p = perm[i]
aux[i] = data[p]
d = aux[i] != aux[i - 1]
mask[i] = d
cumsum += d
inv_idx[p] = cumsum - 1
if d:
counts.append(np.intp(i))
counts.append(aux.shape[0])
# (uniques, indices, inverse index, counts)
return aux[mask], perm[mask], inv_idx, np.diff(np.array(counts))
def _imhistmatch(I_inverse, I_counts, ref_unique, ref_counts):
I_unique = empty(len(I_counts), dtype=ref_unique.dtype)
ir = intp(0)
for iI in range(len(I_counts)):
while I_counts[iI] > ref_counts[ir]:
ir += 1
I_unique[iI] = ref_unique[ir]
return I_unique[I_inverse]
def __init__(self, array, struct_arr_ptr):
print "copying data to device"
self.data = cuda.to_device(array)
self.shape, self.dtype = array.shape, array.dtype
cuda.memcpy_htod(int(struct_arr_ptr),
numpy.getbuffer(numpy.int32(len(array[0]))))
cuda.memcpy_htod(int(struct_arr_ptr) + 8,
numpy.getbuffer(numpy.intp(int(self.data))))
def __init__(self, ar):
self.inputs = np.float32(ar[:, 0:4])
self.outputs = np.intp(ar[:, 4:7])
self.classes = np.array(
list(map(lambda x: output_to_class(x), self.outputs)))
self.count = self.outputs.size
self.bias = np.array([])
self.inputs_m = np.matrix(self.inputs)
self.outputs_m = np.matrix(self.outputs)
def get_rgba(ptr):
#global plotData_d, plot_rgba_d, colorMap_rgba_d
get_rgbaKernel(nCol,
minVar,
maxVar,
plotData_d,
np.intp(ptr),
colorMap_rgba_d,
background_d,
grid=grid2D_GL,
block=block2D_GL)
class DoubleOpStruct:
mem_size = 8 + numpy.intp(0).nbytes
def __init__(self, array, struct_arr_ptr):
self.data = cuda.to_device(array)
self.shape, self.dtype = array.shape, array.dtype
cuda.memcpy_htod(int(struct_arr_ptr), numpy.int32(array.size))
cuda.memcpy_htod(int(struct_arr_ptr) + 8, numpy.intp(int(self.data)))
def __str__(self):
return str(cuda.from_device(self.data, self.shape, self.dtype))
def Mapped(array):
'''Analog to pycuda.driver.InOut(), but indicates this array
is memory mapped to the device space and should not be copied.
To simplify coding, Mapped() will pass anything with a gpudata
member, like a gpuarray, through unchanged.
'''
if hasattr(array, 'gpudata'):
return array
else:
return np.intp(array.base.get_device_pointer())
def stars_mapping(P, stars):
'''
Applying affine mapping to stars
'''
stars = np.flip(stars, axis=1)
stars = np.hstack((stars, np.ones((3, 1))))
cords = np.matmul(P, stars.T).T
cords = np.flip(np.delete(cords, -1, axis=1), axis=1)
cords = np.intp(np.round(cords.ravel()))
return cords
def test_2d(self):
def pyfunc(arg):
return np.array(arg)
cfunc = nrtjit(pyfunc)
# A list of tuples
got = cfunc([(1, 2), (3, 4)])
self.assertPreciseEqual(got, np.intp([[1, 2], [3, 4]]))
got = cfunc([(1, 2.5), (3, 4.5)])
self.assertPreciseEqual(got, np.float64([[1, 2.5], [3, 4.5]]))
# A tuple of lists
got = cfunc(([1, 2], [3, 4]))
self.assertPreciseEqual(got, np.intp([[1, 2], [3, 4]]))
got = cfunc(([1, 2], [3.5, 4.5]))
self.assertPreciseEqual(got, np.float64([[1, 2], [3.5, 4.5]]))
# A tuple of tuples
got = cfunc(((1.5, 2), (3.5, 4.5)))
self.assertPreciseEqual(got, np.float64([[1.5, 2], [3.5, 4.5]]))
got = cfunc(((), ()))
self.assertPreciseEqual(got, np.float64(((), ())))
def rf(X, Y_casual, Y_registered, testSet_final):
rf1 = randomForestModel() #train for casual
rf2 = randomForestModel() #train for registered
rf1.fit(X, Y_casual)
rf2.fit(X, Y_registered)
rf1_Y = np.exp(rf1.predict(testSet_final)) - 1
rf2_Y = np.exp(rf2.predict(testSet_final)) - 1
final_prediction = (rf1_Y + rf2_Y)
final_prediction = np.intp(
np.around(final_prediction)) #round and convert to integer
return final_prediction
def gidx_frm_xycrd(xy_crd, xy_dim):
"""
Obtain the grid idex from grid xy_cord
return 'g_idx', starting from 0. Example for a 3*2 grid
1 | 3 | 5
---------
0 | 2 | 4
"""
g_idx = np.intp((xy_crd[0] - 1) * xy_dim[1] + xy_crd[1] - 1)
return g_idx
def _predict(self, xs):
ys0 = np.zeros(xs.shape[0])
ys1 = np.zeros(xs.shape[0])
for r in self.regs:
r0 = r.r0
r1 = r.r1
ys0 += np.exp(r0.predict(xs[:, r.select])) - 1
ys1 += np.exp(r1.predict(xs[:, r.select])) - 1
ys = np.intp(np.around((ys0 + ys1) * 1.0 / len(self.regs)))
ys[ys < 0] = 0
return ys
def check_read(self, example, a, info, dense, over32, over64):
with open(self.fn, 'w') as f:
f.write(example)
assert_equal(mminfo(self.fn), info)
if (over32 and (np.intp(0).itemsize < 8)) or over64:
assert_raises(OverflowError, mmread, self.fn)
else:
b = mmread(self.fn)
if not dense:
b = b.toarray()
assert_equal(a, b)
def __init__(self, offsets):
# Fix an issue where intp isn't always stored as intp
from numpy import intp
if offsets.dtype != intp:
if offsets.dtype.kind == 'i' and offsets.dtype.itemsize == intp(
0).itemsize:
offsets = offsets.view(intp)
else:
offsets = offsets.astype(intp)
super(Intensity,
self).__init__(abs(offsets.ravel()).max(), offsets.shape[1])
self.__offsets = offsets
def _assemble(self):
if self._assembled:
return
self._assembled = True
mod, sfun, vfun = Mat._lma2csr_cache.get(self.dtype,
(None, None, None))
if mod is None:
d = {'type': self.ctype}
src = _matrix_support_template.render(d).encode('ascii')
compiler_opts = [
'-m64', '-Xptxas', '-dlcm=ca', '-Xptxas=-v', '-O3',
'-use_fast_math', '-DNVCC'
]
mod = SourceModule(src, options=compiler_opts)
sfun = mod.get_function('__lma_to_csr')
vfun = mod.get_function('__lma_to_csr_vector')
sfun.prepare('PPPPPiPii')
vfun.prepare('PPPPPiiPiii')
Mat._lma2csr_cache[self.dtype] = mod, sfun, vfun
for rowmap, colmap in self.sparsity.maps:
assert rowmap.iterset is colmap.iterset
nelems = rowmap.iterset.size
nthread = 128
nblock = (nelems * rowmap.arity * colmap.arity) / nthread + 1
rowmap._to_device()
colmap._to_device()
offset = self._lmaoffset(rowmap.iterset) * self.dtype.itemsize
arglist = [
np.intp(self._lmadata.gpudata) + offset, self._csrdata.gpudata,
self._rowptr.gpudata, self._colidx.gpudata,
rowmap._device_values.gpudata,
np.int32(rowmap.arity)
]
if self._is_scalar_field:
arglist.extend([
colmap._device_values.gpudata,
np.int32(colmap.arity),
np.int32(nelems)
])
fun = sfun
else:
arglist.extend([
np.int32(self.dims[0]), colmap._device_values.gpudata,
np.int32(colmap.arity),
np.int32(self.dims[1]),
np.int32(nelems)
])
fun = vfun
_stream.synchronize()
fun.prepared_async_call((int(nblock), 1, 1), (nthread, 1, 1),
_stream, *arglist)
def computer_skin_probability(self):
'''
Every pixel has a mess of probability of being a skin pixel
according to their RGB values
'''
r = 1.0 * self.image[:, :, 0]
g = 1.0 * self.image[:, :, 1]
b = 1.0 * self.image[:, :, 2]
#print r , g , b
im = numpy.intp(1 + floor(r / 8.0) + floor(g / 8.0) * 32 +
floor(b / 8.0) * 32 * 32)
self.skinProb = skinData[im - 1]
def get_indexer(
self,
target: AnyArrayLike,
method: Optional[str] = None,
limit: Optional[int] = None,
tolerance: Optional[Any] = None,
) -> np.ndarray:
self._check_method(method)
if self.is_overlapping:
msg = (
"cannot handle overlapping indices; use "
"IntervalIndex.get_indexer_non_unique"
)
raise InvalidIndexError(msg)
target = ensure_index(target)
if isinstance(target, IntervalIndex):
# equal indexes -> 1:1 positional match
if self.equals(target):
return np.arange(len(self), dtype="intp")
# different closed or incompatible subtype -> no matches
common_subtype = find_common_type(
[self.dtype.subtype, target.dtype.subtype]
)
if self.closed != target.closed or is_object_dtype(common_subtype):
return np.repeat(np.intp(-1), len(target))
# non-overlapping -> at most one match per interval in target
# want exact matches -> need both left/right to match, so defer to
# left/right get_indexer, compare elementwise, equality -> match
left_indexer = self.left.get_indexer(target.left)
right_indexer = self.right.get_indexer(target.right)
indexer = np.where(left_indexer == right_indexer, left_indexer, -1)
elif not is_object_dtype(target):
# homogeneous scalar index: use IntervalTree
target = self._maybe_convert_i8(target)
indexer = self._engine.get_indexer(target.values)
else:
# heterogeneous scalar index: defer elementwise to get_loc
# (non-overlapping so get_loc guarantees scalar of KeyError)
indexer = []
for key in target:
try:
loc = self.get_loc(key)
except KeyError:
loc = -1
indexer.append(loc)
return ensure_platform_int(indexer)
def pytest_runtest_setup(item):
mark = _get_mark(item, "xslow")
if mark is not None:
try:
v = int(os.environ.get('SCIPY_XSLOW', '0'))
except ValueError:
v = False
if not v:
pytest.skip("very slow test; set environment variable SCIPY_XSLOW=1 to run it")
mark = _get_mark(item, 'xfail_on_32bit')
if mark is not None and np.intp(0).itemsize < 8:
pytest.xfail('Fails on our 32-bit test platform(s): %s' % (mark.args[0],))
def gbm(X, Y_casual, Y_registered, testSet_final):
gbm1 = gradientDescentModel() #train for casual
gbm2 = gradientDescentModel() #train for registered
gbm1.fit(X, Y_casual)
gbm2.fit(X, Y_registered)
gbm1_Y = np.exp(gbm1.predict(testSet_final)) - 1
gbm2_Y = np.exp(gbm2.predict(testSet_final)) - 1
final_prediction = (gbm1_Y + gbm2_Y)
final_prediction = np.intp(
np.around(final_prediction)) #round and convert to integer
return final_prediction
class YoloInfo:
mem_size = 8 * 4 + np.intp(0).nbytes
def __init__(self, n_classes, n_anchors, l_obj, l_noobj, anchors, ptr):
array = np.asarray(anchors, dtype=np.float32)
self.anchors = cuda.to_device(array)
cuda.memcpy_htod(int(ptr), np.getbuffer(np.int32(n_classes)))
cuda.memcpy_htod(int(ptr) + 8, np.getbuffer(np.int32(n_anchors)))
cuda.memcpy_htod(int(ptr) + 16, np.getbuffer(np.float32(l_obj)))
cuda.memcpy_htod(int(ptr) + 24, np.getbuffer(np.float32(l_noobj)))
cuda.memcpy_htod(
int(ptr) + 32, np.getbuffer(np.intp(int(self.anchors))))
def lasso(X, Y_casual, Y_registered, testSet_final):
alpha = 0.5
lasso1 = linear_model.Lasso(alpha=alpha)
lasso2 = linear_model.Lasso(alpha=alpha)
lasso1.fit(X, Y_casual)
lasso2.fit(X, Y_registered)
lasso1_Y = np.exp(lasso1.predict(testSet_final)) - 1
lasso2_Y = np.exp(lasso2.predict(testSet_final)) - 1
final_prediction = np.intp(np.around(lasso1_Y + lasso2_Y))
return final_prediction
def pytest_runtest_setup(item):
mark = _get_mark(item, "xslow")
if mark is not None:
try:
v = int(os.environ.get('SCIPY_XSLOW', '0'))
except ValueError:
v = False
if not v:
pytest.skip(
"very slow test; set environment variable SCIPY_XSLOW=1 to run it"
)
mark = _get_mark(item, 'xfail_on_32bit')
if mark is not None and np.intp(0).itemsize < 8:
pytest.xfail('Fails on our 32-bit test platform(s): %s' %
(mark.args[0], ))
# Older versions of threadpoolctl have an issue that may lead to this
# warning being emitted, see gh-14441
with npt.suppress_warnings() as sup:
sup.filter(pytest.PytestUnraisableExceptionWarning)
try:
from threadpoolctl import threadpool_limits
HAS_THREADPOOLCTL = True
except Exception: # observed in gh-14441: (ImportError, AttributeError)
# Optional dependency only. All exceptions are caught, for robustness
HAS_THREADPOOLCTL = False
if HAS_THREADPOOLCTL:
# Set the number of openmp threads based on the number of workers
# xdist is using to prevent oversubscription. Simplified version of what
# sklearn does (it can rely on threadpoolctl and its builtin OpenMP helper
# functions)
try:
xdist_worker_count = int(
os.environ['PYTEST_XDIST_WORKER_COUNT'])
except KeyError:
# raises when pytest-xdist is not installed
return
if not os.getenv('OMP_NUM_THREADS'):
max_openmp_threads = os.cpu_count(
) // 2 # use nr of physical cores
threads_per_worker = max(
max_openmp_threads // xdist_worker_count, 1)
try:
threadpool_limits(threads_per_worker, user_api='blas')
except Exception:
# May raise AttributeError for older versions of OpenBLAS.
# Catch any error for robustness.
return
def _empty(cls, shape: Shape, dtype: ExtensionDtype):
"""
Create an ExtensionArray with the given shape and dtype.
"""
obj = cls._from_sequence([], dtype=dtype)
taker = np.broadcast_to(np.intp(-1), shape)
result = obj.take(taker, allow_fill=True)
if not isinstance(result, cls) or dtype != result.dtype:
raise NotImplementedError(
f"Default 'empty' implementation is invalid for dtype='{dtype}'"
)
return result
class WordRep(object):
"""
This is used to store representation of word as a path from root to leaf.
length: length of the path from root to leaf.
code: sequence of {+1, -1} +1 if left child, -1 if right child.
point: sequence of indices into syn1 matrix.
"""
memsize = 8 + np.intp(0).nbytes + np.intp(0).nbytes
def __init__(self, code, point, struct_ptr):
self.code = cuda.to_device(code)
self.point = cuda.to_device(point)
self.code_shape, self.code_dtype = code.shape, code.dtype
self.point_shape, self.point_dtype = point.shape, point.dtype
cuda.memcpy_htod(int(struct_ptr), np.int32(code.size))
cuda.memcpy_htod(int(struct_ptr) + 8, np.intp(int(self.code)))
cuda.memcpy_htod(
int(struct_ptr) + 8 + np.intp(0).nbytes, np.intp(int(self.point)))
def __str__(self):
return "len: " + str(self.code_shape) + " code: " + str(cuda.from_device(self.code, self.code_shape, self.code_dtype)) + \
" point: " + str(cuda.from_device(self.point, self.point_shape, self.point_dtype))
def __init__(self, ptrList, struct_ptr, structtype='ARG'):
if structtype == 'ARG':
cuda.memcpy_htod(int(struct_ptr), np.getbuffer(np.int32(ptrList[0])))
struct_ptr = int (struct_ptr) + 4
cuda.memcpy_htod(int(struct_ptr), np.getbuffer(np.int32(ptrList[1])))
struct_ptr = int (struct_ptr) + 4
cuda.memcpy_htod(int(struct_ptr), np.getbuffer(np.int32(ptrList[2])))
struct_ptr = int (struct_ptr) + DOUBLE_SIZE
cuda.memcpy_htod(int(struct_ptr), np.getbuffer(np.int32(ptrList[3])))
struct_ptr = int (struct_ptr) + DOUBLE_SIZE
cuda.memcpy_htod(int(struct_ptr), np.getbuffer(np.int32(ptrList[4])))
struct_ptr = int (struct_ptr) + DOUBLE_SIZE
cuda.memcpy_htod(int(struct_ptr), np.getbuffer(np.int32(ptrList[5])))
struct_ptr = int (struct_ptr) + DOUBLE_SIZE
#print np.int32(ptrList[1])
for value in ptrList[6:]:
cuda.memcpy_htod(int(struct_ptr), np.getbuffer(np.intp(int(value))))
struct_ptr = int (struct_ptr) + np.intp(0).nbytes
else:
for value in ptrList:
cuda.memcpy_htod(int(struct_ptr), np.getbuffer(np.intp(int(value))))
struct_ptr = int (struct_ptr) + np.intp(0).nbytes
def __init__(self, array, ptr):
assert (len(array.shape) == 2)
if isinstance(array, gpuarray.GPUArray):
self.data = array.gpudata
else:
if array.dtype != np.float32:
array = array.astype(np.float32)
self.data = cuda.to_device(array)
self.shape = array.shape
self.dtype = array.dtype
cuda.memcpy_htod(int(ptr), np.getbuffer(np.int32(array.shape[0])))
cuda.memcpy_htod(int(ptr) + 8, np.getbuffer(np.int32(array.shape[1])))
cuda.memcpy_htod(int(ptr) + 16, np.getbuffer(np.intp(int(self.data))))
def elasticnet(X, Y_casual, Y_registered, testSet_final):
alpha = 0.001
l1_ratio = 0.1
glmnet1 = linear_model.ElasticNetCV()
glmnet2 = linear_model.ElasticNetCV()
glmnet1.fit(X, Y_casual)
glmnet2.fit(X, Y_registered)
glmnet1_Y = np.exp(glmnet1.predict(testSet_final)) - 1
glmnet2_Y = np.exp(glmnet2.predict(testSet_final)) - 1
final_prediction = np.intp(np.around(glmnet1_Y + glmnet2_Y))
return final_prediction
def _conv2D(self, d_idata, d_odata, sizes, d_filt, filtlen):
logging.debug("conv2D")
# convolve the first dimension
gdim = (int(iDivUp(sizes[0], self.CONV1_THREAD_SIZE - (filtlen - 1))),
sizes[1] * sizes[2])
bdim = (self.CONV1_THREAD_SIZE, 1, 1)
self.dev_conv1(d_idata,
d_odata,
np.int32(sizes[0]),
np.intp(d_filt),
np.int32(filtlen),
block=bdim,
grid=gdim)
szBytes = self.sizeofFloat * reduce(lambda x, y: x * y, sizes)
d_tmp = cuda.mem_alloc(szBytes)
cuda.memcpy_dtod(d_tmp, d_idata, szBytes)
cuda.memcpy_dtod(d_idata, d_odata, szBytes)
cuda.memcpy_dtod(d_odata, d_tmp, szBytes)
# convolve the second dimension
gdim = (int(iDivUp(sizes[0], self.CONVN_THREAD_SIZE1)),
int(
iDivUp(sizes[1], self.CONVN_THREAD_SIZE2 - (filtlen - 1)) *
sizes[2]))
bdim = (self.CONVN_THREAD_SIZE1, self.CONVN_THREAD_SIZE2, 1)
self.dev_convn(d_idata,
d_odata,
np.int32(sizes[0]),
np.int32(sizes[1]),
np.int32(sizes[0]),
np.int32(sizes[0] * sizes[1]),
np.int32(sizes[2]),
np.intp(d_filt),
np.int32(filtlen),
block=bdim,
grid=gdim)
