def benchmarkQROp(self):
for shape_ in self.shapes:
with ops.Graph().as_default(), \
session.Session(config=benchmark.benchmark_config()) as sess, \
ops.device("/cpu:0"):
matrix_value = np.random.uniform(
low=-1.0, high=1.0, size=shape_).astype(np.float32)
matrix = variables.Variable(matrix_value)
q, r = linalg_ops.qr(matrix)
variables.global_variables_initializer().run()
self.run_op_benchmark(
sess,
control_flow_ops.group(q, r),
min_iters=25,
name="QR_cpu_{shape}".format(shape=shape_))
if test.is_gpu_available(True):
with ops.Graph().as_default(), \
session.Session(config=benchmark.benchmark_config()) as sess, \
ops.device("/device:GPU:0"):
matrix_value = np.random.uniform(
low=-1.0, high=1.0, size=shape_).astype(np.float32)
matrix = variables.Variable(matrix_value)
q, r = linalg_ops.qr(matrix)
variables.global_variables_initializer().run()
self.run_op_benchmark(
sess,
control_flow_ops.group(q, r),
min_iters=25,
name="QR_gpu_{shape}".format(shape=shape_))
def benchmarkQROp(self):
for shape_ in self.shapes:
with ops.Graph().as_default(), \
session.Session(config=benchmark.benchmark_config()) as sess, \
ops.device("/cpu:0"):
matrix_value = np.random.uniform(
low=-1.0, high=1.0, size=shape_).astype(np.float32)
matrix = variables.Variable(matrix_value)
q, r = linalg_ops.qr(matrix)
variables.global_variables_initializer().run()
self.run_op_benchmark(
sess,
control_flow_ops.group(q, r),
min_iters=25,
name="QR_cpu_{shape}".format(shape=shape_))
if test.is_gpu_available(True):
with ops.Graph().as_default(), \
session.Session(config=benchmark.benchmark_config()) as sess, \
ops.device("/device:GPU:0"):
matrix_value = np.random.uniform(
low=-1.0, high=1.0, size=shape_).astype(np.float32)
matrix = variables.Variable(matrix_value)
q, r = linalg_ops.qr(matrix)
variables.global_variables_initializer().run()
self.run_op_benchmark(
sess,
control_flow_ops.group(q, r),
min_iters=25,
name="QR_gpu_{shape}".format(shape=shape_))
def testWrongDimensions(self):
# The input to qr should be a tensor of at least rank 2.
scalar = constant_op.constant(1.)
with self.assertRaisesRegexp(ValueError,
"Shape must be at least rank 2 but is rank 0"):
linalg_ops.qr(scalar)
vector = constant_op.constant([1., 2.])
with self.assertRaisesRegexp(ValueError,
"Shape must be at least rank 2 but is rank 1"):
linalg_ops.qr(vector)
def testWrongDimensions(self):
# The input to qr should be a tensor of at least rank 2.
scalar = constant_op.constant(1.)
with self.assertRaisesRegexp(
ValueError, "Shape must be at least rank 2 but is rank 0"):
linalg_ops.qr(scalar)
vector = constant_op.constant([1., 2.])
with self.assertRaisesRegexp(
ValueError, "Shape must be at least rank 2 but is rank 1"):
linalg_ops.qr(vector)
def testWrongDimensions(self):
# The input to svd should be a tensor of at least rank 2.
scalar = constant_op.constant(1.)
with self.assertRaisesRegex(
(ValueError, errors_impl.InvalidArgumentError), "rank.* 2.*0"):
linalg_ops.qr(scalar)
vector = constant_op.constant([1., 2.])
with self.assertRaisesRegex(
(ValueError, errors_impl.InvalidArgumentError), "rank.* 2.*1"):
linalg_ops.qr(vector)
def __call__(self, shape, dtype=None, partition_info=None):
if dtype is None:
dtype = self.dtype
# Check the shape
if len(shape) < 2:
raise ValueError("The tensor to initialize must be "
"at least two-dimensional")
# Flatten the input shape with the last dimension remaining
# its original shape so it works for conv2d
num_rows = 1
for dim in shape[:-1]:
num_rows *= dim
num_cols = shape[-1]
flat_shape = (num_rows, num_cols)
# Generate a random matrix
a = random_ops.random_normal(flat_shape, dtype=dtype, seed=self.seed)
# Compute the qr factorization
q, r = linalg_ops.qr(a, full_matrices=False)
# Make Q uniform
square_len = math_ops.minimum(num_rows, num_cols)
d = array_ops.diag_part(r[:square_len, :square_len])
ph = d / math_ops.abs(d)
q *= ph
# Pad zeros to Q (if rows smaller than cols)
if num_rows < num_cols:
padding = array_ops.zeros([num_rows, num_cols - num_rows], dtype=dtype)
q = array_ops.concat([q, padding], 1)
return self.gain * array_ops.reshape(q, shape)
def __call__(self, shape, dtype=None, partition_info=None):
if dtype is None:
dtype = self.dtype
# Check the shape
if len(shape) < 3 or len(shape) > 5:
raise ValueError("The tensor to initialize must be at least "
"three-dimensional and at most five-dimensional")
if shape[-2] > shape[-1]:
raise ValueError("In_filters cannot be greater than out_filters.")
# Generate a random matrix
a = random_ops.random_normal([shape[-1], shape[-1]],
dtype=dtype,
seed=self.seed)
# Compute the qr factorization
q, r = linalg_ops.qr(a, full_matrices=False)
# Make Q uniform
d = array_ops.diag_part(r)
q *= math_ops.sign(d)
q = q[:shape[-2], :]
q *= math_ops.sqrt(math_ops.cast(self.gain, dtype=dtype))
if len(shape) == 3:
weight = array_ops.scatter_nd([[(shape[0] - 1) // 2]],
array_ops.expand_dims(q, 0), shape)
elif len(shape) == 4:
weight = array_ops.scatter_nd([[(shape[0] - 1) // 2,
(shape[1] - 1) // 2]],
array_ops.expand_dims(q, 0), shape)
else:
weight = array_ops.scatter_nd([[(shape[0] - 1) // 2,
(shape[1] - 1) // 2,
(shape[2] - 1) // 2]],
array_ops.expand_dims(q, 0), shape)
return weight
def Test(self):
np.random.seed(42)
a = np.random.uniform(low=-1.0, high=1.0, size=shape_).astype(dtype_)
if dtype_ in [np.complex64, np.complex128]:
a += 1j * np.random.uniform(
low=-1.0, high=1.0, size=shape_).astype(dtype_)
# Optimal stepsize for central difference is O(epsilon^{1/3}).
epsilon = np.finfo(dtype_).eps
delta = 0.1 * epsilon**(1.0 / 3.0)
if dtype_ in [np.float32, np.complex64]:
tol = 3e-2
else:
tol = 1e-6
with self.session(use_gpu=True):
tf_a = constant_op.constant(a)
tf_b = linalg_ops.qr(tf_a, full_matrices=full_matrices_)
for b in tf_b:
x_init = np.random.uniform(
low=-1.0, high=1.0, size=shape_).astype(dtype_)
if dtype_ in [np.complex64, np.complex128]:
x_init += 1j * np.random.uniform(
low=-1.0, high=1.0, size=shape_).astype(dtype_)
theoretical, numerical = gradient_checker.compute_gradient(
tf_a,
tf_a.get_shape().as_list(),
b,
b.get_shape().as_list(),
x_init_value=x_init,
delta=delta)
self.assertAllClose(theoretical, numerical, atol=tol, rtol=tol)
def __call__(self, shape, dtype=None, partition_info=None):
if dtype is None:
dtype = self.dtype
# Check the shape
if len(shape) < 2:
raise ValueError("The tensor to initialize must be "
"at least two-dimensional")
# Flatten the input shape with the last dimension remaining
# its original shape so it works for conv2d
num_rows = 1
for dim in shape[:-1]:
num_rows *= dim
num_cols = shape[-1]
flat_shape = (num_cols,
num_rows) if num_rows < num_cols else (num_rows,
num_cols)
# Generate a random matrix
a = random_ops.random_normal(flat_shape, dtype=dtype, seed=self.seed)
# Compute the qr factorization
q, r = linalg_ops.qr(a, full_matrices=False)
# Make Q uniform
d = array_ops.diag_part(r)
ph = d / math_ops.abs(d)
q *= ph
if num_rows < num_cols:
q = array_ops.matrix_transpose(q)
return self.gain * array_ops.reshape(q, shape)
def _test(self, dtype, shape, full_matrices):
np.random.seed(1)
x_np = np.random.uniform(
low=-1.0, high=1.0, size=np.prod(shape)).reshape(shape).astype(dtype)
with self.test_session() as sess:
x_tf = array_ops.placeholder(dtype)
with self.test_scope():
q_tf, r_tf = linalg_ops.qr(x_tf, full_matrices=full_matrices)
q_tf_val, r_tf_val = sess.run([q_tf, r_tf], feed_dict={x_tf: x_np})
q_dims = q_tf_val.shape
np_q = np.ndarray(q_dims, dtype)
np_q_reshape = np.reshape(np_q, (-1, q_dims[-2], q_dims[-1]))
new_first_dim = np_q_reshape.shape[0]
x_reshape = np.reshape(x_np, (-1, x_np.shape[-2], x_np.shape[-1]))
for i in range(new_first_dim):
if full_matrices:
np_q_reshape[i, :, :], _ = np.linalg.qr(
x_reshape[i, :, :], mode="complete")
else:
np_q_reshape[i, :, :], _ = np.linalg.qr(
x_reshape[i, :, :], mode="reduced")
np_q = np.reshape(np_q_reshape, q_dims)
self.CompareOrthogonal(np_q, q_tf_val, min(shape[-2:]))
self.CheckApproximation(x_np, q_tf_val, r_tf_val)
self.CheckUnitary(q_tf_val)
def Test(self):
np.random.seed(42)
a = np.random.uniform(low=-1.0, high=1.0, size=shape_).astype(dtype_)
if dtype_ in [np.complex64, np.complex128]:
a += 1j * np.random.uniform(low=-1.0, high=1.0,
size=shape_).astype(dtype_)
# Optimal stepsize for central difference is O(epsilon^{1/3}).
epsilon = np.finfo(dtype_).eps
delta = 0.1 * epsilon**(1.0 / 3.0)
if dtype_ in [np.float32, np.complex64]:
tol = 3e-2
else:
tol = 1e-6
with self.session(use_gpu=True):
tf_a = constant_op.constant(a)
tf_b = linalg_ops.qr(tf_a, full_matrices=full_matrices_)
for b in tf_b:
x_init = np.random.uniform(low=-1.0, high=1.0,
size=shape_).astype(dtype_)
if dtype_ in [np.complex64, np.complex128]:
x_init += 1j * np.random.uniform(
low=-1.0, high=1.0, size=shape_).astype(dtype_)
theoretical, numerical = gradient_checker.compute_gradient(
tf_a,
tf_a.get_shape().as_list(),
b,
b.get_shape().as_list(),
x_init_value=x_init,
delta=delta)
self.assertAllClose(theoretical, numerical, atol=tol, rtol=tol)
def __call__(self, shape, dtype=None, partition_info=None):
if dtype is None:
dtype = self.dtype
# Check the shape
if len(shape) < 2:
raise ValueError("The tensor to initialize must be "
"at least two-dimensional")
# Flatten the input shape with the last dimension remaining
# its original shape so it works for conv2d
num_rows = 1
for dim in shape[:-1]:
num_rows *= dim
num_cols = shape[-1]
flat_shape = (num_cols, num_rows) if num_rows < num_cols else (num_rows,
num_cols)
# Generate a random matrix
a = random_ops.random_normal(flat_shape, dtype=dtype, seed=self.seed)
# Compute the qr factorization
q, r = linalg_ops.qr(a, full_matrices=False)
# Make Q uniform
d = array_ops.diag_part(r)
q *= math_ops.sign(d)
if num_rows < num_cols:
q = array_ops.matrix_transpose(q)
return self.gain * array_ops.reshape(q, shape)
def _test(self, x_np, full_matrices, full_rank=True):
dtype = x_np.dtype
shape = x_np.shape
with self.session() as sess:
x_tf = array_ops.placeholder(dtype)
with self.device_scope():
q_tf, r_tf = linalg_ops.qr(x_tf, full_matrices=full_matrices)
q_tf_val, r_tf_val = sess.run([q_tf, r_tf], feed_dict={x_tf: x_np})
q_dims = q_tf_val.shape
np_q = np.ndarray(q_dims, dtype)
np_q_reshape = np.reshape(np_q, (-1, q_dims[-2], q_dims[-1]))
new_first_dim = np_q_reshape.shape[0]
x_reshape = np.reshape(x_np, (-1, x_np.shape[-2], x_np.shape[-1]))
for i in range(new_first_dim):
if full_matrices:
np_q_reshape[i, :, :], _ = np.linalg.qr(
x_reshape[i, :, :], mode="complete")
else:
np_q_reshape[i, :, :], _ = np.linalg.qr(
x_reshape[i, :, :], mode="reduced")
np_q = np.reshape(np_q_reshape, q_dims)
if full_rank:
# Q is unique up to sign/phase if the matrix is full-rank.
self.CompareOrthogonal(np_q, q_tf_val, min(shape[-2:]))
self.CheckApproximation(x_np, q_tf_val, r_tf_val)
self.CheckUnitary(q_tf_val)
def testConcurrentExecutesWithoutError(self):
with self.session(use_gpu=True) as sess:
all_ops = []
for full_matrices_ in True, False:
for rows_ in 4, 5:
for cols_ in 4, 5:
matrix1 = random_ops.random_normal([rows_, cols_], seed=42)
matrix2 = random_ops.random_normal([rows_, cols_], seed=42)
q1, r1 = linalg_ops.qr(matrix1, full_matrices=full_matrices_)
q2, r2 = linalg_ops.qr(matrix2, full_matrices=full_matrices_)
all_ops += [q1, r1, q2, r2]
val = self.evaluate(all_ops)
for i in range(8):
q = 4 * i
self.assertAllEqual(val[q], val[q + 2]) # q1 == q2
self.assertAllEqual(val[q + 1], val[q + 3]) # r1 == r2
def testConcurrentExecutesWithoutError(self):
with self.session(use_gpu=True) as sess:
all_ops = []
for full_matrices_ in True, False:
for rows_ in 4, 5:
for cols_ in 4, 5:
matrix1 = random_ops.random_normal([rows_, cols_], seed=42)
matrix2 = random_ops.random_normal([rows_, cols_], seed=42)
q1, r1 = linalg_ops.qr(matrix1, full_matrices=full_matrices_)
q2, r2 = linalg_ops.qr(matrix2, full_matrices=full_matrices_)
all_ops += [q1, r1, q2, r2]
val = self.evaluate(all_ops)
for i in range(8):
q = 4 * i
self.assertAllClose(val[q], val[q + 2]) # q1 == q2
self.assertAllClose(val[q + 1], val[q + 3]) # r1 == r2
def __call__(self, shape, dtype=None, partition_info=None):
if dtype is None:
dtype = self.dtype
# Check the shape
if len(shape) < 3 or len(shape) > 5:
raise ValueError("The tensor to initialize must be at least "
"three-dimensional and at most five-dimensional")
if shape[-2] > shape[-1]:
raise ValueError("In_filters cannot be greater than out_filters.")
# Generate a random matrix
a = random_ops.random_normal([shape[-1], shape[-1]],
dtype=dtype, seed=self.seed)
# Compute the qr factorization
q, r = linalg_ops.qr(a, full_matrices=False)
# Make Q uniform
d = array_ops.diag_part(r)
q *= math_ops.sign(d)
q = q[:shape[-2], :]
q *= math_ops.sqrt(math_ops.cast(self.gain, dtype=dtype))
if len(shape) == 3:
weight = array_ops.scatter_nd([[(shape[0]-1)//2]],
array_ops.expand_dims(q, 0), shape)
elif len(shape) == 4:
weight = array_ops.scatter_nd([[(shape[0]-1)//2, (shape[1]-1)//2]],
array_ops.expand_dims(q, 0), shape)
else:
weight = array_ops.scatter_nd([[(shape[0]-1)//2, (shape[1]-1)//2,
(shape[2]-1)//2]],
array_ops.expand_dims(q, 0), shape)
return weight
def testConcurrentExecutesWithoutError(self):
seed = [42, 24]
all_ops = []
for full_matrices_ in True, False:
for rows_ in 4, 5:
for cols_ in 4, 5:
matrix_shape = [rows_, cols_]
matrix1 = stateless_random_ops.stateless_random_normal(
matrix_shape, seed)
matrix2 = stateless_random_ops.stateless_random_normal(
matrix_shape, seed)
self.assertAllEqual(matrix1, matrix2)
q1, r1 = linalg_ops.qr(matrix1, full_matrices=full_matrices_)
q2, r2 = linalg_ops.qr(matrix2, full_matrices=full_matrices_)
all_ops += [q1, q2, r1, r2]
val = self.evaluate(all_ops)
for i in range(0, len(val), 2):
self.assertAllClose(val[i], val[i + 1])
def Test(self):
np.random.seed(42)
# Optimal stepsize for central difference is O(epsilon^{1/3}).
epsilon = np.finfo(dtype_).eps
delta = 0.1 * epsilon**(1.0 / 3.0)
if dtype_ in [np.float32, np.complex64]:
tol = 3e-2
else:
tol = 1e-6
# TODO(b/157171666): Sadly we have to double the computation because
# gradient_checker_v2.compute_gradient expects a list of functions.
funcs = [
lambda a: linalg_ops.qr(a, full_matrices=full_matrices_)[0],
lambda a: linalg_ops.qr(a, full_matrices=full_matrices_)[1]
]
for f in funcs:
theoretical, numerical = gradient_checker_v2.compute_gradient(
f, [RandomInput()], delta=delta)
self.assertAllClose(theoretical, numerical, atol=tol, rtol=tol)
def _orthogonal_matrix(self, n):
"""Construct an n x n orthogonal matrix.
Args:
n: dimension.
Returns:
a n x n orthogonal matrix.
"""
a = random_ops.random_normal([n, n], dtype=self.dtype, seed=self.seed)
if self.seed:
self.seed += 1
q, r = linalg_ops.qr(a)
d = array_ops.diag_part(r)
# make q uniform
q *= math_ops.sign(d)
return q
def _orthogonal_matrix(self, n):
"""Construct an n x n orthogonal matrix.
Args:
n: dimension.
Returns:
a n x n orthogonal matrix.
"""
a = random_ops.random_normal([n, n], dtype=self.dtype, seed=self.seed)
if self.seed:
self.seed += 1
q, r = linalg_ops.qr(a)
d = array_ops.diag_part(r)
# make q uniform
q *= math_ops.sign(d)
return q
def Test(self):
np.random.seed(1)
x_np = np.random.uniform(
low=-1.0, high=1.0,
size=np.prod(shape_)).reshape(shape_).astype(dtype_)
if is_complex:
x_np += 1j * np.random.uniform(
low=-1.0, high=1.0,
size=np.prod(shape_)).reshape(shape_).astype(dtype_)
# TODO(rmlarsen): Debug failure due to invalid parameter to ORMQR.
rows_ = shape_[-2]
cols_ = shape_[-1]
use_gpu = False if rows_ < cols_ or (full_matrices_
and rows_ != cols_) else True
with self.test_session(use_gpu=use_gpu) as sess:
if use_static_shape_:
x_tf = constant_op.constant(x_np)
else:
x_tf = array_ops.placeholder(dtype_)
q_tf, r_tf = linalg_ops.qr(x_tf, full_matrices=full_matrices_)
if use_static_shape_:
q_tf_val, r_tf_val = sess.run([q_tf, r_tf])
else:
q_tf_val, r_tf_val = sess.run([q_tf, r_tf],
feed_dict={x_tf: x_np})
q_dims = q_tf_val.shape
np_q = np.ndarray(q_dims, dtype_)
np_q_reshape = np.reshape(np_q, (-1, q_dims[-2], q_dims[-1]))
new_first_dim = np_q_reshape.shape[0]
x_reshape = np.reshape(x_np, (-1, x_np.shape[-2], x_np.shape[-1]))
for i in range(new_first_dim):
if full_matrices_:
np_q_reshape[i,:,:], _ = \
np.linalg.qr(x_reshape[i,:,:], mode="complete")
else:
np_q_reshape[i,:,:], _ = \
np.linalg.qr(x_reshape[i,:,:], mode="reduced")
np_q = np.reshape(np_q_reshape, q_dims)
CompareOrthogonal(self, np_q, q_tf_val, min(shape_[-2:]))
CheckApproximation(self, x_np, q_tf_val, r_tf_val)
CheckUnitary(self, q_tf_val)
def Test(self):
np.random.seed(42)
a = np.random.uniform(low=-1.0, high=1.0, size=shape_).astype(dtype_)
if dtype_ in [np.complex64, np.complex128]:
a += 1j * np.random.uniform(
low=-1.0, high=1.0, size=shape_).astype(dtype_)
# Optimal stepsize for central difference is O(epsilon^{1/3}).
epsilon = np.finfo(dtype_).eps
delta = 0.1 * epsilon**(1.0 / 3.0)
if dtype_ in [np.float32, np.complex64]:
tol = 3e-2
else:
tol = 1e-6
with self.session(use_gpu=True):
tf_a = constant_op.constant(a)
tf_b = linalg_ops.qr(tf_a, full_matrices=full_matrices_)
for b in tf_b:
x_init = np.random.uniform(
low=-1.0, high=1.0, size=shape_).astype(dtype_)
if dtype_ in [np.complex64, np.complex128]:
x_init += 1j * np.random.uniform(
low=-1.0, high=1.0, size=shape_).astype(dtype_)
# The compute_gradient on qr gradident will call blas kernel on GPU,
# which might lead to deviated results from the reference with tf32.
if dtype_ == np.float32:
with ops.device("/cpu:0"):
theoretical, numerical = gradient_checker.compute_gradient(
tf_a,
tf_a.get_shape().as_list(),
b,
b.get_shape().as_list(),
x_init_value=x_init,
delta=delta)
else:
theoretical, numerical = gradient_checker.compute_gradient(
tf_a,
tf_a.get_shape().as_list(),
b,
b.get_shape().as_list(),
x_init_value=x_init,
delta=delta)
self.assertAllClose(theoretical, numerical, atol=tol, rtol=tol)
def Test(self):
np.random.seed(1)
x_np = np.random.uniform(
low=-1.0, high=1.0,
size=np.prod(shape_)).reshape(shape_).astype(dtype_)
if is_complex:
x_np += 1j * np.random.uniform(
low=-1.0, high=1.0,
size=np.prod(shape_)).reshape(shape_).astype(dtype_)
# rocBLAS on ROCm stack does not support complex<double> GEMM yet
with self.session(
use_gpu=True and not test.is_built_with_rocm()) as sess:
if use_static_shape_:
x_tf = constant_op.constant(x_np)
else:
x_tf = array_ops.placeholder(dtype_)
q_tf, r_tf = linalg_ops.qr(x_tf, full_matrices=full_matrices_)
if use_static_shape_:
q_tf_val, r_tf_val = self.evaluate([q_tf, r_tf])
else:
q_tf_val, r_tf_val = sess.run([q_tf, r_tf],
feed_dict={x_tf: x_np})
q_dims = q_tf_val.shape
np_q = np.ndarray(q_dims, dtype_)
np_q_reshape = np.reshape(np_q, (-1, q_dims[-2], q_dims[-1]))
new_first_dim = np_q_reshape.shape[0]
x_reshape = np.reshape(x_np, (-1, x_np.shape[-2], x_np.shape[-1]))
for i in range(new_first_dim):
if full_matrices_:
np_q_reshape[i, :, :], _ = np.linalg.qr(x_reshape[i, :, :],
mode="complete")
else:
np_q_reshape[i, :, :], _ = np.linalg.qr(x_reshape[i, :, :],
mode="reduced")
np_q = np.reshape(np_q_reshape, q_dims)
CompareOrthogonal(self, np_q, q_tf_val, min(shape_[-2:]))
CheckApproximation(self, x_np, q_tf_val, r_tf_val)
CheckUnitary(self, q_tf_val)
def Test(self):
if not use_static_shape_ and context.executing_eagerly():
return
np.random.seed(1)
x_np = np.random.uniform(
low=-1.0, high=1.0,
size=np.prod(shape_)).reshape(shape_).astype(dtype_)
if is_complex:
x_np += 1j * np.random.uniform(
low=-1.0, high=1.0,
size=np.prod(shape_)).reshape(shape_).astype(dtype_)
if use_static_shape_:
x_tf = constant_op.constant(x_np)
else:
x_tf = array_ops.placeholder(dtype_)
q_tf, r_tf = linalg_ops.qr(x_tf, full_matrices=full_matrices_)
if use_static_shape_:
q_tf_val, r_tf_val = self.evaluate([q_tf, r_tf])
else:
with self.session(use_gpu=True) as sess:
q_tf_val, r_tf_val = sess.run([q_tf, r_tf],
feed_dict={x_tf: x_np})
q_dims = q_tf_val.shape
np_q = np.ndarray(q_dims, dtype_)
np_q_reshape = np.reshape(np_q, (-1, q_dims[-2], q_dims[-1]))
new_first_dim = np_q_reshape.shape[0]
x_reshape = np.reshape(x_np, (-1, x_np.shape[-2], x_np.shape[-1]))
for i in range(new_first_dim):
if full_matrices_:
np_q_reshape[i, :, :], _ = np.linalg.qr(x_reshape[i, :, :],
mode="complete")
else:
np_q_reshape[i, :, :], _ = np.linalg.qr(x_reshape[i, :, :],
mode="reduced")
np_q = np.reshape(np_q_reshape, q_dims)
CompareOrthogonal(self, np_q, q_tf_val, min(shape_[-2:]))
CheckApproximation(self, x_np, q_tf_val, r_tf_val)
CheckUnitary(self, q_tf_val)
def Test(self):
np.random.seed(1)
x_np = np.random.uniform(
low=-1.0, high=1.0, size=np.prod(shape_)).reshape(shape_).astype(dtype_)
if is_complex:
x_np += 1j * np.random.uniform(
low=-1.0, high=1.0,
size=np.prod(shape_)).reshape(shape_).astype(dtype_)
for full_matrices in False, True:
with self.test_session() as sess:
if use_static_shape_:
x_tf = constant_op.constant(x_np)
else:
x_tf = array_ops.placeholder(dtype_)
q_tf, r_tf = linalg_ops.qr(x_tf, full_matrices=full_matrices)
if use_static_shape_:
q_tf_val, r_tf_val = sess.run([q_tf, r_tf])
else:
q_tf_val, r_tf_val = sess.run([q_tf, r_tf], feed_dict={x_tf: x_np})
q_dims = q_tf_val.shape
np_q = np.ndarray(q_dims, dtype_)
np_q_reshape = np.reshape(np_q, (-1, q_dims[-2], q_dims[-1]))
new_first_dim = np_q_reshape.shape[0]
x_reshape = np.reshape(x_np, (-1, x_np.shape[-2], x_np.shape[-1]))
for i in range(new_first_dim):
if full_matrices:
np_q_reshape[i,:,:], _ = \
np.linalg.qr(x_reshape[i,:,:], mode="complete")
else:
np_q_reshape[i,:,:], _ = \
np.linalg.qr(x_reshape[i,:,:], mode="reduced")
np_q = np.reshape(np_q_reshape, q_dims)
CompareOrthogonal(self, np_q, q_tf_val, min(shape_[-2:]))
CheckApproximation(self, x_np, q_tf_val, r_tf_val)
CheckUnitary(self, q_tf_val)
def _initializer(shape, dtype=None, partition_info=None):
if len(shape) < 2:
raise ValueError(
"The tensor to initialize must be at least two-dimensional")
num_rows = 1
for dim in [shape[i] for i in flatten_axis]:
num_rows *= dim
num_cols = shape[list(set(range(len(shape))) - set(flatten_axis))[0]]
flat_shape = (num_cols,
num_rows) if num_rows < num_cols else (num_rows,
num_cols)
a = random(flat_shape, type='uniform', stddev=1.0)
q, r = linalg_ops.qr(a, full_matrices=False)
q *= np.sqrt(flat_shape[0])
if num_rows < num_cols:
q = tf.matrix_transpose(q)
return stddev * tf.reshape(q, shape)
def _NoGrad(x):
with backprop.GradientTape() as tape:
tape.watch(x)
ret = linalg_ops.qr(x, full_matrices=True)
return tape.gradient(ret, x)
