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
if context.executing_eagerly():
with ops.device("cpu:0"): # TODO(b/73102536)
q, r = gen_linalg_ops.qr(a, full_matrices=False)
else:
q, r = gen_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 __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 = gen_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 __call__(self, shape, dtype=dtypes.float32):
"""Returns a tensor object initialized as specified by the initializer.
Args:
shape: Shape of the tensor.
dtype: Optional dtype of the tensor. Only floating point types are
supported.
Raises:
ValueError: If the dtype is not floating point or the input shape is not
valid.
"""
dtype = _assert_float_dtype(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 = (max(num_cols, num_rows), min(num_cols, num_rows))
# Generate a random matrix
a = self._random_generator.random_normal(flat_shape, dtype=dtype)
# Compute the qr factorization
q, r = gen_linalg_ops.qr(a, full_matrices=False)
# Make Q uniform
d = array_ops.diag_part(r)
q *= math_ops.sign(d)
if tf.linalg.det(q) < 0:
q[:, 0] *= -1
if num_rows < num_cols:
q = array_ops.matrix_transpose(q)
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 = gen_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 __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 = gen_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 __call__(self, shape, dtype=dtypes.float32):
"""Returns a tensor object initialized as specified by the initializer.
Args:
shape: Shape of the tensor.
dtype: Optional dtype of the tensor. Only floating point types are
supported.
Raises:
ValueError: If the dtype is not floating point or the input shape is not
valid.
"""
dtype = _assert_float_dtype(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 = (max(num_cols, num_rows), min(num_cols, num_rows))
# Generate a random matrix
a = random_ops.random_normal(flat_shape, dtype=dtype, seed=self.seed)
# Compute the qr factorization
q, r = gen_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 orth(self, dim=None):
'''
Orthogonalize the tensor.
Parameters
----------
dim : int, optional
The dimension of the orthogonal dim-matricization of self. The
default is None, returning a copy of the original tensor.
Returns
-------
tensor : FullTensor
A tensor whose dim-matricization is an orthogonal matrix
corresponding to the Q factor of a QR factorization of the
dim-matricization of self.
r_matrix : numpy.ndarray
The R factor.
'''
tensor = FullTensor(self) # Copy the tensor
if dim is None:
return tensor, np.array([])
if dim == -1:
dim = tensor.order - 1
dims = np.concatenate(
(np.arange(dim), np.arange(dim + 1, tensor.order), [dim]))
tensor = tensor.transpose(dims)
shape0 = np.array(tensor.shape)
tensor = tensor.reshape([np.prod(shape0[:-1]), shape0[-1]])
try:
from tensorflow.python.ops.gen_linalg_ops import qr
q_tf, r_tf = qr(tensor.data, full_matrices=False)
tensor.data, r_matrix = q_tf.numpy(), r_tf.numpy()
except ImportError:
tensor.data, r_matrix = np.linalg.qr(tensor.data)
shape0[-1] = r_matrix.shape[0]
tensor = tensor.reshape(shape0)
tensor = tensor.itranspose(dims)
tensor.is_orth = True
tensor.orth_dim = dim
return tensor, r_matrix
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 = gen_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 = gen_linalg_ops.qr(a)
d = array_ops.diag_part(r)
# make q uniform
q *= math_ops.sign(d)
return q
def __call__(self, shape, dtype=None, **kwargs):
"""Returns a tensor object initialized to an orthogonal matrix.
Args:
shape: Shape of the tensor.
dtype: Optional dtype of the tensor. Only floating point types are
supported. If not specified, `tf.keras.backend.floatx()` is used,
which default to `float32` unless you configured it otherwise
(via `tf.keras.backend.set_floatx(float_dtype)`)
**kwargs: Additional keyword arguments.
"""
_validate_kwargs(self.__class__.__name__,
kwargs,
support_partition=False)
dtype = _assert_float_dtype(_get_dtype(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 = (max(num_cols, num_rows), min(num_cols, num_rows))
# Generate a random matrix
a = self._random_generator.random_normal(flat_shape, dtype=dtype)
# Compute the qr factorization
q, r = gen_linalg_ops.qr(a, full_matrices=False)
# Make Q uniform
d = array_ops.tensor_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)
