def _testDS(self,
op,
array1=numpy.array([[1., 0], [3, 0], [0, 6]]),
array2=numpy.asarray([[0, 2.], [0, 4], [5, 0]])):
for mtype in _mtypes:
a = mtype(array1)
aR = as_sparse_variable(a)
self.assertFalse(aR.data is a)
self.assertTrue(_is_sparse(a))
self.assertTrue(_is_sparse_variable(aR))
b = numpy.asarray(array2)
bR = tensor.as_tensor_variable(b)
self.assertFalse(bR.data is b)
self.assertTrue(_is_dense(b))
self.assertTrue(_is_dense_variable(bR))
apb = op(aR, bR)
self.assertTrue(apb.type.dtype == aR.type.dtype, apb.type.dtype)
self.assertTrue(apb.type.dtype == bR.type.dtype, apb.type.dtype)
val = eval_outputs([apb])
self.assertTrue(val.shape == (3, 2))
if op is add:
self.assertTrue(_is_dense_variable(apb))
self.assertTrue(numpy.all(val == (a + b)))
ans = numpy.array([[1., 2], [3, 4], [5, 6]])
self.assertTrue(numpy.all(val == ans))
elif op is mul:
self.assertTrue(_is_sparse_variable(apb))
ans = numpy.array([[1, 0], [9, 0], [0, 36]])
self.assertTrue(numpy.all(val.todense() == (a.multiply(b))))
self.assertTrue(numpy.all(val.todense() == ans))
def test_basicSS(self):
for mtype in _mtypes:
x = as_sparse_variable(mtype((500, 3)))
x.data[(10, 1)] = 1
x.data[(20, 2)] = 2
self.assertTrue(_is_sparse_variable(x))
xT = x.T
self.assertTrue(_is_sparse_variable(xT))
zop = true_dot(x, xT)
self.assertTrue(_is_sparse_variable(zop))
z = eval_outputs([zop])
self.assertTrue(_is_sparse(z))
self.assertTrue(z.shape == (500, 500))
self.assertTrue(type(z) is mtype)
w = mtype((500, 500))
w[(10, 10)] = 1
w[(20, 20)] = 4
self.assertTrue(z.shape == w.shape)
self.assertTrue(type(z) == type(w))
self.assertTrue(z.dtype == w.dtype)
#self.assertTrue(z == w)
self.assertTrue(abs(z - w).nnz == 0)
z = z.todense()
w = w.todense()
self.assertTrue((z == w).all() == True)
def test_basicSD(self):
for mtype in _mtypes:
x = as_sparse_variable(mtype((500,3)))
x.data[(10, 1)] = 1
x.data[(20, 2)] = 2
self.assertTrue(_is_sparse_variable(x))
y = tensor.as_tensor_variable([[1., 2], [3, 4], [2, 1]])
self.assertTrue(_is_dense_variable(y))
zop = true_dot(x,y)
self.assertTrue(_is_sparse_variable(zop))
z = eval_outputs([zop])
self.assertTrue(_is_sparse(z))
self.assertTrue(z.shape == (500,2))
self.assertTrue(type(z) is mtype)
w = mtype((500,2))
w[(10, 0)] = 3.
w[(20, 0)] = 4
w[(10, 1)] = 4
w[(20, 1)] = 2
self.assertTrue(z.shape == w.shape)
self.assertTrue(type(z) == type(w))
self.assertTrue(z.dtype == w.dtype)
#self.assertTrue(z == w)
self.assertTrue(abs(z-w).nnz == 0)
z = z.todense()
w = w.todense()
self.assertTrue((z == w).all() == True)
def test_basicSD(self):
for mtype in _mtypes:
x = as_sparse_variable(mtype((500, 3)))
x.data[(10, 1)] = 1
x.data[(20, 2)] = 2
self.assertTrue(_is_sparse_variable(x))
y = tensor.as_tensor_variable([[1., 2], [3, 4], [2, 1]])
self.assertTrue(_is_dense_variable(y))
zop = true_dot(x, y)
self.assertTrue(_is_sparse_variable(zop))
z = eval_outputs([zop])
self.assertTrue(_is_sparse(z))
self.assertTrue(z.shape == (500, 2))
self.assertTrue(type(z) is mtype)
w = mtype((500, 2))
w[(10, 0)] = 3.
w[(20, 0)] = 4
w[(10, 1)] = 4
w[(20, 1)] = 2
self.assertTrue(z.shape == w.shape)
self.assertTrue(type(z) == type(w))
self.assertTrue(z.dtype == w.dtype)
#self.assertTrue(z == w)
self.assertTrue(abs(z - w).nnz == 0)
z = z.todense()
w = w.todense()
self.assertTrue((z == w).all() == True)
def test_basicSS(self):
for mtype in _mtypes:
x = as_sparse_variable(mtype((500,3)))
x.data[(10, 1)] = 1
x.data[(20, 2)] = 2
self.assertTrue(_is_sparse_variable(x))
xT = x.T
self.assertTrue(_is_sparse_variable(xT))
zop = true_dot(x,xT)
self.assertTrue(_is_sparse_variable(zop))
z = eval_outputs([zop])
self.assertTrue(_is_sparse(z))
self.assertTrue(z.shape == (500,500))
self.assertTrue(type(z) is mtype)
w = mtype((500,500))
w[(10, 10)] = 1
w[(20, 20)] = 4
self.assertTrue(z.shape == w.shape)
self.assertTrue(type(z) == type(w))
self.assertTrue(z.dtype == w.dtype)
#self.assertTrue(z == w)
self.assertTrue(abs(z-w).nnz == 0)
z = z.todense()
w = w.todense()
self.assertTrue((z == w).all() == True)
def _testDS(
self, op, array1=numpy.array([[1.0, 0], [3, 0], [0, 6]]), array2=numpy.asarray([[0, 2.0], [0, 4], [5, 0]])
):
for mtype in _mtypes:
a = mtype(array1)
aR = as_sparse_variable(a)
self.assertFalse(aR.data is a)
self.assertTrue(_is_sparse(a))
self.assertTrue(_is_sparse_variable(aR))
b = numpy.asarray(array2)
bR = tensor.as_tensor_variable(b)
self.assertFalse(bR.data is b)
self.assertTrue(_is_dense(b))
self.assertTrue(_is_dense_variable(bR))
apb = op(aR, bR)
self.assertTrue(apb.type.dtype == aR.type.dtype, apb.type.dtype)
self.assertTrue(apb.type.dtype == bR.type.dtype, apb.type.dtype)
val = eval_outputs([apb])
self.assertTrue(val.shape == (3, 2))
if op is add:
self.assertTrue(_is_dense_variable(apb))
self.assertTrue(numpy.all(val == (a + b)))
ans = numpy.array([[1.0, 2], [3, 4], [5, 6]])
self.assertTrue(numpy.all(val == ans))
elif op is mul:
self.assertTrue(_is_sparse_variable(apb))
ans = numpy.array([[1, 0], [9, 0], [0, 36]])
self.assertTrue(numpy.all(val.todense() == (a.multiply(b))))
self.assertTrue(numpy.all(val.todense() == ans))
def grad(self, inp, grads):
x, y = inp
gz, = grads
assert _is_sparse_variable(gz)
assert _is_sparse_variable(x)
rval = [true_dot(gz, y.T), true_dot(x.T, gz)]
if _is_dense_variable(y):
if self.grad_preserves_dense:
rval[1] = dense_from_sparse(rval[1])
return rval
def grad(self, inp, grads):
x, y = inp
gz, = grads
assert _is_sparse_variable(gz)
assert _is_sparse_variable(x)
rval = [true_dot(gz, y.T), true_dot(x.T, gz)]
if _is_dense_variable(y):
if self.grad_preserves_dense:
rval[1] = dense_from_sparse(rval[1])
return rval
class SamplingDot(gof.op.Op):
"""
Operand for calculating the dot product DOT(X, Y) = Z when you
only want to calculate a subset of Z. It is equivalent to P o (X
. Y) where o is the element-wise product, X and Y operands of the
dot product and P is a matrix that contains 1 when the
corresponding element of Z should be calculated and 0 when it
shouldn't. Note that SamplingDot has a different interface than
DOT because SamplingDot requires X to be a MxK matrix while Y is a
NxK matrix instead of the usual KxN matrix.
It will work if the pattern is not binary value, but if the
pattern doesn't have a high sparsity proportion it will be slower
then a more optimized dot followed by a normal elemwise
multiplication.
"""
def __eq__(self, other):
return type(self) == type(other)
def __hash__(self):
return hash(type(self))
def __str__(self):
return 'SamplingDot'
def make_node(self, x, y, p):
x = tensor.as_tensor_variable(x)
y = tensor.as_tensor_variable(y)
if not _is_sparse_variable(p):
raise TypeError(p)
#TODO: use it.
dtype_out = scalar.upcast(x.type.dtype, y.type.dtype, p.type.dtype)
return gof.Apply(self, [x, y, p], [p.type()])
def perform(self, node, (x, y, p), (out, )):
if _is_sparse_variable(x):
raise TypeError(x)
if _is_sparse_variable(y):
raise TypeError(y)
if not _is_sparse(p):
raise TypeError(p)
rval = p.__class__(p.multiply(numpy.dot(x, y.T)))
out[0] = rval
def local_structured_add_s_v(node):
if node.op == structured_add_s_v:
x, y = node.inputs
x_is_sparse_variable = _is_sparse_variable(x)
#y_is_sparse_variable = _is_sparse_variable(y)
if x_is_sparse_variable:
svar = x
dvar = y
else:
svar = y
dvar = x
if dvar.type.ndim != 1:
return False
elif svar.type.format == 'csr':
CSx = CSR
structured_add_s_v_csx = structured_add_s_v_csr
else:
return False
s_val, s_ind, s_ptr, s_shape = csm_properties(svar)
c_data = structured_add_s_v_csx(s_val, s_ind, s_ptr, dvar)
return [CSx(c_data, s_ind, s_ptr, s_shape)]
return False
def local_mul_s_v(node):
if node.op == mul_s_v:
x, y = node.inputs
x_is_sparse_variable = _is_sparse_variable(x)
if x_is_sparse_variable:
svar = x
dvar = y
else:
svar = y
dvar = x
if dvar.type.ndim != 1:
return False
elif svar.type.format == 'csr':
CSx = CSR
mul_s_v_csx = mul_s_v_csr
else:
return False
s_val, s_ind, s_ptr, s_shape = csm_properties(svar)
c_data = mul_s_v_csx(s_val, s_ind, s_ptr, dvar)
return [CSx(c_data, s_ind, s_ptr, s_shape)]
return False
def local_structured_add_s_v(node):
if node.op == structured_add_s_v:
x, y = node.inputs
x_is_sparse_variable = _is_sparse_variable(x)
#y_is_sparse_variable = _is_sparse_variable(y)
if x_is_sparse_variable:
svar = x
dvar = y
else:
svar = y
dvar = x
if dvar.type.ndim != 1:
return False
elif svar.type.format == 'csr':
CSx = CSR
structured_add_s_v_csx = structured_add_s_v_csr
else:
raise NotImplemented()
s_val, s_ind, s_ptr, s_shape = csm_properties(svar)
c_data = structured_add_s_v_csx(s_val, s_ind, s_ptr, dvar)
return [CSx(c_data, s_ind, s_ptr, s_shape)]
return False
def local_mul_s_d(node):
if node.op == mul_s_d:
x, y = node.inputs
x_is_sparse_variable = _is_sparse_variable(x)
# y_is_sparse_variable = _is_sparse_variable(y)
if x_is_sparse_variable:
svar = x
dvar = y
else:
svar = y
dvar = x
if dvar.type.ndim != 2:
return False
if svar.type.format == 'csc':
CSx = CSC
mul_s_d_csx = mul_s_d_csc
elif svar.type.format == 'csr':
CSx = CSR
mul_s_d_csx = mul_s_d_csr
else:
raise NotImplemented()
c_data = mul_s_d_csx(csm_data(svar), csm_indices(svar),
csm_indptr(svar), dvar)
return [CSx(c_data, csm_indices(svar), csm_indptr(svar),
csm_shape(svar))]
return False
def local_mul_s_d(node):
if node.op == mul_s_d:
x, y = node.inputs
x_is_sparse_variable = _is_sparse_variable(x)
# y_is_sparse_variable = _is_sparse_variable(y)
if x_is_sparse_variable:
svar = x
dvar = y
else:
svar = y
dvar = x
if dvar.type.ndim != 2:
return False
if svar.type.format == 'csc':
CSx = CSC
mul_s_d_csx = mul_s_d_csc
elif svar.type.format == 'csr':
CSx = CSR
mul_s_d_csx = mul_s_d_csr
else:
raise NotImplemented()
c_data = mul_s_d_csx(csm_data(svar), csm_indices(svar),
csm_indptr(svar), dvar)
return [
CSx(c_data, csm_indices(svar), csm_indptr(svar), csm_shape(svar))
]
return False
def true_dot(x, y, grad_preserves_dense=True):
"""
@todo: Maybe the triple-transposition formulation (when x is dense)
is slow. See if there is a direct way to do this.
"""
if hasattr(x, 'getnnz'): x = as_sparse_variable(x)
if hasattr(y, 'getnnz'): y = as_sparse_variable(y)
x_is_sparse_variable = _is_sparse_variable(x)
y_is_sparse_variable = _is_sparse_variable(y)
if not x_is_sparse_variable and not y_is_sparse_variable:
raise TypeError()
if x_is_sparse_variable:
return TrueDot(grad_preserves_dense)(x, y)
else:
assert y_is_sparse_variable
return transpose(TrueDot(grad_preserves_dense)(y.T, x.T))
def true_dot(x, y, grad_preserves_dense=True):
"""
@todo: Maybe the triple-transposition formulation (when x is dense)
is slow. See if there is a direct way to do this.
"""
if hasattr(x, 'getnnz'): x = as_sparse_variable(x)
if hasattr(y, 'getnnz'): y = as_sparse_variable(y)
x_is_sparse_variable = _is_sparse_variable(x)
y_is_sparse_variable = _is_sparse_variable(y)
if not x_is_sparse_variable and not y_is_sparse_variable:
raise TypeError()
if x_is_sparse_variable:
return TrueDot(grad_preserves_dense)(x, y)
else:
assert y_is_sparse_variable
return transpose(TrueDot(grad_preserves_dense)(y.T, x.T))
def make_node(self, x, y, p):
x = tensor.as_tensor_variable(x)
y = tensor.as_tensor_variable(y)
if not _is_sparse_variable(p):
raise TypeError(p)
#TODO: use it.
dtype_out = scalar.upcast(x.type.dtype, y.type.dtype, p.type.dtype)
return gof.Apply(self, [x, y, p], [p.type()])
def make_node(self, x, y, p):
x = tensor.as_tensor_variable(x)
y = tensor.as_tensor_variable(y)
if not _is_sparse_variable(p):
raise TypeError(p)
#TODO: use it.
dtype_out = scalar.upcast(x.type.dtype, y.type.dtype, p.type.dtype)
return gof.Apply(self, [x, y, p], [p.type()])
def test_basicDS(self):
for mtype in _mtypes:
x = as_sparse_variable(mtype((500, 3)))
x.data[(10, 1)] = 1
x.data[(20, 2)] = 2
self.assertTrue(_is_sparse_variable(x))
y = tensor.as_tensor_variable([[1., 2], [3, 4], [2, 1]])
self.assertTrue(_is_dense_variable(y))
x.data = x.data.T
y.data = y.data.T
zop = true_dot(y, x)
zop = transpose(true_dot(y, x))
self.assertTrue(_is_sparse_variable(zop))
z = eval_outputs([zop])
self.assertTrue(_is_sparse(z))
self.assertTrue(z.shape == (500, 2))
# self.assertTrue(type(z) is mtype)
w = mtype((500, 2))
w[(10, 0)] = 3.
w[(20, 0)] = 4
w[(10, 1)] = 4
w[(20, 1)] = 2
self.assertTrue(z.shape == w.shape)
# Type should switch from csr to csc and vice-versa, so don't perform this test
#self.assertTrue(type(z) == type(w))
self.assertTrue(z.dtype == w.dtype)
# Type should switch from csr to csc and vice-versa, so don't perform this test
#self.assertTrue(z == w)
self.assertTrue(abs(z - w).nnz == 0)
z = z.todense()
w = w.todense()
self.assertTrue((z == w).all() == True)
def test_basicDS(self):
for mtype in _mtypes:
x = as_sparse_variable(mtype((500,3)))
x.data[(10, 1)] = 1
x.data[(20, 2)] = 2
self.assertTrue(_is_sparse_variable(x))
y = tensor.as_tensor_variable([[1., 2], [3, 4], [2, 1]])
self.assertTrue(_is_dense_variable(y))
x.data = x.data.T
y.data = y.data.T
zop = true_dot(y, x)
zop = transpose(true_dot(y, x))
self.assertTrue(_is_sparse_variable(zop))
z = eval_outputs([zop])
self.assertTrue(_is_sparse(z))
self.assertTrue(z.shape == (500,2))
# self.assertTrue(type(z) is mtype)
w = mtype((500,2))
w[(10, 0)] = 3.
w[(20, 0)] = 4
w[(10, 1)] = 4
w[(20, 1)] = 2
self.assertTrue(z.shape == w.shape)
# Type should switch from csr to csc and vice-versa, so don't perform this test
#self.assertTrue(type(z) == type(w))
self.assertTrue(z.dtype == w.dtype)
# Type should switch from csr to csc and vice-versa, so don't perform this test
#self.assertTrue(z == w)
self.assertTrue(abs(z-w).nnz == 0)
z = z.todense()
w = w.todense()
self.assertTrue((z == w).all() == True)
def make_node(self, x, y):
"""
:note: Because of trickiness of implementing, we assume that the left argument x is SparseVariable (not dense)
"""
if x.type.dtype != y.type.dtype:
raise NotImplementedError()
if not _is_sparse_variable(x):
raise TypeError(x)
# These are the conversions performed by scipy.sparse.dot
if x.type.format == "csc" or x.type.format == "coo":
myformat = "csc"
elif x.type.format == "csr":
myformat = "csr"
else:
raise NotImplementedError()
inputs = [x, y] # Need to convert? e.g. assparse
outputs = [SparseType(dtype = x.type.dtype, format = myformat).make_variable()]
return gof.Apply(self, inputs, outputs)
def make_node(self, x, y):
"""
:note: Because of trickiness of implementing, we assume that the left argument x is SparseVariable (not dense)
"""
if x.type.dtype != y.type.dtype:
raise NotImplementedError()
if not _is_sparse_variable(x):
raise TypeError(x)
# These are the conversions performed by scipy.sparse.dot
if x.type.format == "csc" or x.type.format == "coo":
myformat = "csc"
elif x.type.format == "csr":
myformat = "csr"
else:
raise NotImplementedError()
inputs = [x, y] # Need to convert? e.g. assparse
outputs = [
SparseType(dtype=x.type.dtype, format=myformat).make_variable()
]
return gof.Apply(self, inputs, outputs)
assert y.type.ndim == 1
if x.type.dtype != y.type.dtype:
raise NotImplementedError()
return gof.Apply(self, [x, y], [
SparseType(dtype=x.type.dtype,
format=x.type.format).make_variable()
])
def perform(self, node, (x, y), (out, )):
assert _is_sparse(x) and not _is_sparse(y)
assert x.shape[1] == y.shape[0]
out[0] = x.__class__(x.toarray() * y)
def grad(self, (x, y), (gz, )):
assert _is_sparse_variable(x) and _is_dense_variable(y)
assert _is_sparse_variable(gz)
return mul_s_v(gz, y), sp_sum(x * gz, axis=0, sparse_grad=True)
mul_s_v = MulSV()
class MulSVCSR(gof.Op):
def __eq__(self, other):
return (type(self) == type(other))
def __hash__(self):
return hash(type(self))
def make_node(self, a_data, a_indices, a_indptr, b):
assert y.type.ndim == 1
if x.type.dtype != y.type.dtype:
raise NotImplementedError()
return gof.Apply(self,
[x, y],
[SparseType(dtype=x.type.dtype,
format=x.type.format).make_variable()])
def perform(self, node, (x, y), (out, )):
assert _is_sparse(x) and not _is_sparse(y)
assert x.shape[1] == y.shape[0]
out[0] = x.__class__(x + (x.toarray() != 0) * y)
def grad(self, (x, y), (gz,)):
assert _is_sparse_variable(x) and _is_sparse_variable(y)
assert _is_sparse_variable(gz)
return gz, gz
structured_add_s_v = StructuredAddSV()
class StrucutedAddSVCSR(gof.Op):
def __eq__(self, other):
return (type(self) == type(other))
def __hash__(self):
return hash(type(self))
def make_node(self, a_data, a_indices, a_indptr, b):
assert b.type.ndim == 1
return gof.Apply(self, [a_data, a_indices, a_indptr, b],
assert y.type.ndim == 1
if x.type.dtype != y.type.dtype:
raise NotImplementedError()
return gof.Apply(self,
[x, y],
[SparseType(dtype=x.type.dtype,
format=x.type.format).make_variable()])
def perform(self, node, (x, y), (out, )):
assert _is_sparse(x) and not _is_sparse(y)
assert x.shape[1] == y.shape[0]
out[0] = x.__class__(x.toarray() * y)
def grad(self, (x, y), (gz,)):
assert _is_sparse_variable(x) and _is_dense_variable(y)
assert _is_sparse_variable(gz)
return mul_s_v(gz, y), sp_sum(x * gz, axis=0, sparse_grad=True)
mul_s_v = MulSV()
class MulSVCSR(gof.Op):
def __eq__(self, other):
return (type(self) == type(other))
def __hash__(self):
return hash(type(self))
def make_node(self, a_data, a_indices, a_indptr, b):
assert b.type.ndim == 1
return gof.Apply(self, [a_data, a_indices, a_indptr, b],