def __init__(self, data, shape = None):
"""Constructor for tensor object.
data can be numpy.array or list.
shape can be numpy.array, list, tuple of integers"""
if(data.__class__ == list):
data = numpy.array(data);
if(shape != None):
if(len(shape) == 0):
raise ValueError("Second argument must be a row vector.");
if(shape.__class__ == numpy.ndarray):
if(shape.ndim != 2 and shape[0].size != 1):
raise ValueError("Second argument must be a row vector.");
shape = tuple(shape);
else:
shape = tuple(data.shape);
if (len(shape) == 0):
if (data.size != 0):
raise ValueError("Empty tensor cannot contain any elements");
elif (tools.prod(shape) != data.size):
raise ValueError("Size of data does not match specified size of tensor");
self.shape = shape;
self.data = data.reshape(self.shape, order='F');
def ttm(self, mat, dims = None, option = None):
""" computes the tensor times the given matrix.
arrs is a single 2-D matrix/array or a list of those matrices/arrays."""
if(dims == None):
dims = range(0,self.ndims());
#Handle when arrs is a list of arrays
if(mat.__class__ == list):
if(len(mat) == 0):
raise ValueError("the given list of arrays is empty!");
(dims,vidx) = tools.tt_dimscheck(dims, self.ndims(), len(mat));
Y = self.ttm(mat[vidx[0]],dims[0],option);
for i in range(1, len(dims)):
Y = Y.ttm(mat[vidx[i]],dims[i],option);
return Y;
if(mat.ndim != 2):
raise ValueError ("matrix in 2nd armuent must be a matrix!");
if(dims.__class__ == list):
if(len(dims) != 1):
raise ValueError("Error in number of elements in dims");
else:
dims = dims[0];
if(dims < 0 or dims > self.ndims()):
raise ValueError ("Dimension N must be between 1 and num of dimensions");
#Compute the product
N = self.ndims();
shp = self.shape;
order = []
order.extend([dims]);
order.extend(range(0,dims));
order.extend(range(dims+1,N));
newdata = self.permute(order).data;
newdata = newdata.reshape(shp[dims], tools.prod(shp)/shp[dims]);
if(option == None):
newdata = numpy.dot(mat, newdata);
p = mat.shape[0];
elif(option == 't'):
newdata = numpy.dot(mat.transpose(), newdata);
p = mat.shape[1];
else:
raise ValueError("Unknown option");
newshp = [p];
newshp.extend(tools.getelts(shp,range(0,dims)));
newshp.extend(tools.getelts(shp,range(dims+1,N)));
Y = tensor(newdata, newshp);
Y = Y.ipermute(order);
return Y;
def ttm(self, mat, dims = None, option = None):
""" computes the tensor times the given matrix.
arrs is a single 2-D matrix/array or a list of those matrices/arrays."""
if(dims == None):
dims = range(0,self.ndims());
#Handle when arrs is a list of arrays
if(mat.__class__ == list):
if(len(mat) == 0):
raise ValueError("the given list of arrays is empty!");
(dims,vidx) = tools.tt_dimscehck(dims, self.ndims(), len(mat));
Y = self.ttm(mat[vidx[0]],dims[0],option);
for i in range(1, len(dims)):
Y = Y.ttm(mat[vidx[i]],dims[i],option);
return Y;
if(mat.ndim != 2):
raise ValueError ("matrix in 2nd armuent must be a matrix!");
if(dims.__class__ == list):
if(len(dims) != 1):
raise ValueError("Error in number of elements in dims");
else:
dims = dims[0];
if(dims < 0 or dims > self.ndims()):
raise ValueError ("Dimension N must be between 1 and num of dimensions");
#Compute the product
N = self.ndims();
shp = self.shape;
order = []
order.extend([dims]);
order.extend(range(0,dims));
order.extend(range(dims+1,N));
newdata = self.permute(order).data;
newdata = newdata.reshape(shp[dims], tools.prod(shp)/shp[dims]);
if(option == None):
newdata = numpy.dot(mat, newdata);
p = mat.shape[0];
elif(option == 't'):
newdata = numpy.dot(mat.transpose(), newdata);
p = mat.shape[1];
else:
raise ValueError("Unknown option");
newshp = [p];
newshp.extend(tools.getelts(shp,range(0,dims)));
newshp.extend(tools.getelts(shp,range(dims+1,N)));
Y = tensor(newdata, newshp);
Y = Y.ipermute(order);
return Y;
def __init__(self, data, shape = None):
"""Constructor for tensor object.
dat can be numpy.array or list.
shape can be numpy.array, list, tuple of integers"""
if(data.__class__ == list):
data = numpy.array(data);
if(shape != None):
if(len(shape) == 0):
raise ValueError("Second argument must be a row vector.");
if(shape.__class__ == numpy.ndarray):
if(shape.ndim != 2 and shape[0].size != 1):
raise ValueError("Second argument must be a row vector.");
shape = tuple(shape);
else:
shape = tuple(data.shape);
if (len(shape) == 0):
if (data.size != 0):
raise ValueError("Empty tensor cannot contain any elements");
elif (tools.prod(shape) != data.size):
raise ValueError("Size of data does not match specified size of tensor");
self.shape = shape;
self.data = data.reshape(self.shape);
def gm(values):
"""Computes the geometric mean of a list of numbers.
>>> print gm([2, 8])
4.0
"""
assert len(values) >= 1
return round(tools.prod(values) ** (1/len(values)), 4)
def __call__(self, *args, **kwargs):
_, nd, dims, strs, offsets, contig = check_args(args,
collapse=False,
broadcast=False)
out = kwargs.pop('out', None)
if len(kwargs) != 0:
raise TypeError('Unexpected keyword argument: %s' %
kwargs.keys()[0])
n = prod(dims)
out_shape = tuple(d for i, d in enumerate(dims) if not self.redux[i])
gs = prod(out_shape)
if gs == 0:
gs = 1
n /= gs
if gs > self.context.maxgsize:
raise ValueError("Array to big to be reduced along the "
"selected axes")
if out is None:
out = gpuarray.empty(out_shape,
context=self.context,
dtype=self.dtype_out)
else:
if out.shape != out_shape or out.dtype != self.dtype_out:
raise TypeError(
"Out array is not of expected type "
"(expected %s %s, got %s %s)" %
(out_shape, self.dtype_out, out.shape, out.dtype))
#Don't compile and cache for nothing for big size
if self.init_local_size < n:
k, _, _, ls = self._get_basic_kernel(self.init_local_size, nd)
else:
k, _, _, ls = self._get_basic_kernel(n, nd)
kargs = [n, out]
kargs.extend(dims)
for i, arg in enumerate(args):
kargs.append(arg)
if isinstance(arg, gpuarray.GpuArray):
kargs.append(offsets[i])
kargs.extend(strs[i])
k(*kargs, ls=ls, gs=gs)
return out
def gm(values):
"""Computes the geometric mean of a list of numbers.
>>> gm([2, 8])
4.0
"""
assert len(values) >= 1
exp = 1.0 / len(values)
return round(tools.prod([val ** exp for val in values]), 4)
def __call__(self, *args, **kwargs):
_, nd, dims, strs, offsets, contig = check_args(args, collapse=False,
broadcast=False)
out = kwargs.pop('out', None)
if len(kwargs) != 0:
raise TypeError('Unexpected keyword argument: %s' %
kwargs.keys()[0])
n = prod(dims)
out_shape = tuple(d for i, d in enumerate(dims) if not self.redux[i])
gs = prod(out_shape)
if gs == 0:
gs = 1
n /= gs
if gs > self.context.maxgsize:
raise ValueError("Array to big to be reduced along the "
"selected axes")
if out is None:
out = gpuarray.empty(out_shape, context=self.context,
dtype=self.dtype_out)
else:
if out.shape != out_shape or out.dtype != self.dtype_out:
raise TypeError("Out array is not of expected type "
"(expected %s %s, got %s %s)" % (
out_shape, self.dtype_out, out.shape, out.dtype))
#Don't compile and cache for nothing for big size
if self.init_local_size < n:
k, _, _, ls = self._get_basic_kernel(self.init_local_size, nd)
else:
k, _, _, ls = self._get_basic_kernel(n, nd)
kargs = [n, out]
kargs.extend(dims)
for i, arg in enumerate(args):
kargs.append(arg)
if isinstance(arg, gpuarray.GpuArray):
kargs.append(offsets[i])
kargs.extend(strs[i])
k(*kargs, ls=ls, gs=gs)
return out
def ttv(self, vec, dims=None):
if (dims is None):
dims = range(0, self.ndims())
if (vec.__class__ == list):
if (len(vec) == 0):
raise ValueError("the given list of arrays is empty!")
(dims, vidx) = tools.tt_dimscheck(dims, self.ndims(), len(vec))
Y = self.ttv(vec[vidx[0]], dims[0])
for i in range(1, len(dims)):
Y = Y.ttv(vec[videx[i]], dims[i])
return Y
if (vec.ndim != 1):
raise ValueError("param in 2nd argment must be a vector")
if (dims.__class__ == list):
if (len(dims) != 1):
raise ValueError("Error in number of elements in dims")
else:
dims = dims[0]
if (dims < 0 or dims > self.ndims()):
raise ValueError(
"Dimension N must be between 1 and num of dimensions")
N = self.ndims()
shp = self.shape
order = []
order.extend(range(0, dims))
order.extend(range(dims + 1, N))
order.extend([dims])
newdata = self.permute(order).data
newdata = newdata.reshape(tools.prod(shp) / shp[dims], shp[dims])
newdata = numpy.dot(newdata, vec)
N -= 1
sz = numpy.array(self.shape)[order]
newshp = sz[:N]
# newshp = []
# newshp.extend(tools.getelts(shp, range(0, dims)))
# newshp.extend(tools.getelts(shp, range(dims+1, N)))
Y = tensor(newdata, newshp)
return Y
def ttm(self, mat, dims=None, option=None):
""" computes the tensor times the given matrix.
arrs is a single 2-D matrix/array or a list of those matrices/arrays."""
if (dims == None):
dims = range(0, self.ndims())
#Handle when arrs is a list of arrays
if (mat.__class__ == list):
if (len(mat) == 0):
raise ValueError("the given list of arrays is empty!")
(dims, vidx) = tools.tt_dimscehck(dims, self.ndims(), len(mat))
Y = self.ttm(mat[vidx[0]], dims[0], option)
for i in range(1, len(dims)):
Y = Y.ttm(mat[vidx[i]], dims[i], option)
return Y
if (mat.ndim != 2):
raise ValueError("matrix in 2nd armuent must be a matrix!")
if (dims.__class__ == list):
if (len(dims) != 1):
raise ValueError("Error in number of elements in dims")
else:
dims = dims[0]
if (dims < 0 or dims > self.ndims()):
raise ValueError(
"Dimension N must be between 1 and num of dimensions")
#Compute the product
N = self.ndims()
shp = self.shape
order = []
order.extend([dims])
order.extend(range(0, dims))
order.extend(range(dims + 1, N))
# 第一步把对应模I Permute到第一个,然后reshape成(I,I~)
# 第二步matrix * 上面得到的矩阵
newdata = self.permute(order).data
newdata = newdata.reshape(shp[dims],
tools.prod(shp) / shp[dims])
if (option == None):
newdata = numpy.dot(mat, newdata)
p = mat.shape[0]
elif (option == 't'):
newdata = numpy.dot(mat.transpose(), newdata)
p = mat.shape[1]
else:
raise ValueError("Unknown option")
newshp = [p]
newshp.extend(tools.getelts(shp, range(0, dims)))
newshp.extend(tools.getelts(shp, range(dims + 1, N)))
Y = tensor(newdata, newshp)
# 这里的ipermute很关键,按照相反permute,理解的不是很直观,有时间可以多想想
Y = Y.ipermute(order)
return Y
def gm_old(values):
return round(prod(values) ** (1 / len(values)), 2)
