yield vect
def randn(self):
_randn = (np.array(np.random.randn(*self.shape)).astype(self.dtype)
+ 1.0j * np.array(np.random.randn(*self.shape)).astype(self.dtype))
return pycuda.gpuarray.to_gpu(_randn)
def _inner_prod(self, x, y):
return pycuda.gpuarray.dot(x.ravel().conj(), y.ravel()).real
def _covector(self, x):
return x.conj()
VSpace.register(GPUArray,
lambda x: ComplexGPUArrayVSpace(x)
if cumisc.iscomplextype(x.dtype)
else GPUArrayVSpace(x))
# for type_ in (float, np.float64, np.float32, np.float16):
# GPUArrayVSpace.register(type_)
# for type_ in (complex, np.complex64, np.complex128):
# ComplexGPUArrayVSpace.register(type_)
### GPU VJPS ###
GPU_DIFFERENTIABLE_FUNCTIONS = (
# skcuda.linalg
1.0j * np.ones(self.shape, dtype=self.dtype))
def standard_basis(self):
for idxs in np.ndindex(*self.shape):
for v in [1.0, 1.0j]:
vect = np.zeros(self.shape, dtype=self.dtype)
vect[idxs] = v
yield vect
def randn(self):
return (
np.array(np.random.randn(*self.shape)).astype(self.dtype) +
1.0j * np.array(np.random.randn(*self.shape)).astype(self.dtype))
def _inner_prod(self, x, y):
return np.real(np.dot(np.conj(np.ravel(x)), np.ravel(y)))
def _covector(self, x):
return np.conj(x)
VSpace.register(
np.ndarray, lambda x: ComplexArrayVSpace(x)
if np.iscomplexobj(x) else ArrayVSpace(x))
for type_ in [float, np.float64, np.float32, np.float16]:
ArrayVSpace.register(type_)
for type_ in [complex, np.complex64, np.complex128]:
ComplexArrayVSpace.register(type_)
from autograd.extend import VSpace
class SparseArrayVSpace(VSpace):
def __init__(self, value):
self.t = type(value)
self.value = self.t(value)
self.shape = value.shape
self.dtype = value.dtype
@property
def size(self):
return self.value.size
@property
def ndim(self):
return len(self.shape)
def randn(self):
a = cp.sparse.random(m=self.shape[0], n=self.shape[1])
return self.t(a)
def zeros(self):
return self.t(self.shape)
VSpace.register(cp.sparse.csc_matrix, lambda x: SparseArrayVSpace(x))
VSpace.register(cp.sparse.csr_matrix, lambda x: SparseArrayVSpace(x))
VSpace.register(cp.sparse.coo_matrix, lambda x: SparseArrayVSpace(x))
VSpace.register(cp.sparse.dia_matrix, lambda x: SparseArrayVSpace(x))
return ( np.ones(self.shape, dtype=self.dtype)
+ 1.0j * np.ones(self.shape, dtype=self.dtype))
def standard_basis(self):
for idxs in np.ndindex(*self.shape):
for v in [1.0, 1.0j]:
vect = np.zeros(self.shape, dtype=self.dtype)
vect[idxs] = v
yield vect
def randn(self):
return ( np.array(np.random.randn(*self.shape)).astype(self.dtype)
+ 1.0j * np.array(np.random.randn(*self.shape)).astype(self.dtype))
def _inner_prod(self, x, y):
return np.real(np.dot(np.conj(np.ravel(x)), np.ravel(y)))
def _covector(self, x):
return np.conj(x)
VSpace.register(np.ndarray,
lambda x: ComplexArrayVSpace(x)
if np.iscomplexobj(x)
else ArrayVSpace(x))
for type_ in [float, np.float64, np.float32, np.float16]:
ArrayVSpace.register(type_)
for type_ in [complex, np.complex64, np.complex128]:
ComplexArrayVSpace.register(type_)
return self.value.size
@property
def ndim(self):
return len(self.shape)
def randn(self):
a = scipy.sparse.random(m=self.shape[0], n=self.shape[1])
return self.t(a)
def zeros(self):
return self.t(self.shape)
def _inner_prod(self, x, y):
x_vec = x.reshape((-1, 1))
y_vec = y.reshape((-1, 1))
dot_prod = np.sum(x_vec.T.dot(y_vec))
return dot_prod
# def __eq__(self, other):
# issametype = type(self) == type(other)
# issamevals = (self.value != other.value).nnz == 0
# return issametype # and issamevals
VSpace.register(scipy.sparse.csc_matrix, lambda x: SparseArrayVSpace(x))
VSpace.register(scipy.sparse.csr_matrix, lambda x: SparseArrayVSpace(x))
VSpace.register(scipy.sparse.coo_matrix, lambda x: SparseArrayVSpace(x))
VSpace.register(scipy.sparse.dia_matrix, lambda x: SparseArrayVSpace(x))
VSpace.register(np.matrixlib.defmatrix.matrix, lambda x: VSpace(x))