def __call__(self, shape, dtype, **kwargs):
value = linalg_ops_impl.eye(*shape, dtype=dtype)
if "partition_shape" in kwargs and "partition_offset" in kwargs:
return array_ops.slice(value, kwargs["partition_offset"],
kwargs["partition_shape"])
raise AssertionError("PartitionAwareIdentity do not support "
"non-partitioned initialization")
def _block_orth(self, p1, p2, p3):
"""Construct a 3 x 3 kernel. Used to construct orthgonal kernel.
Args:
p1: A symmetric projection matrix.
p2: A symmetric projection matrix.
p3: A symmetric projection matrix.
Returns:
A 2 x 2 x 2 kernel.
Raises:
ValueError: If the dimensions of p1, p2 and p3 are different.
"""
p1_shape = p1.shape.as_list()
if p1_shape != p2.shape.as_list() or p1_shape != p3.shape.as_list():
raise ValueError("The dimension of the matrices must be the same.")
n = p1_shape[0]
eye = linalg_ops_impl.eye(n, dtype=self.dtype)
kernel2x2x2 = {}
def matmul(p1, p2, p3):
return math_ops.matmul(math_ops.matmul(p1, p2), p3)
def cast(i, p):
"""Return p or (1-p)."""
return i * p + (1-i) * (eye - p)
for i in [0, 1]:
for j in [0, 1]:
for k in [0, 1]:
kernel2x2x2[i, j, k] = matmul(cast(i, p1), cast(j, p2), cast(k, p3))
return kernel2x2x2
def _block_orth(self, p1, p2, p3):
"""Construct a 3 x 3 kernel. Used to construct orthgonal kernel.
Args:
p1: A symmetric projection matrix.
p2: A symmetric projection matrix.
p3: A symmetric projection matrix.
Returns:
A 2 x 2 x 2 kernel.
Raises:
ValueError: If the dimensions of p1, p2 and p3 are different.
"""
p1_shape = p1.shape.as_list()
if p1_shape != p2.shape.as_list() or p1_shape != p3.shape.as_list():
raise ValueError("The dimension of the matrices must be the same.")
n = p1_shape[0]
eye = linalg_ops_impl.eye(n, dtype=self.dtype)
kernel2x2x2 = {}
def matmul(p1, p2, p3):
return math_ops.matmul(math_ops.matmul(p1, p2), p3)
def cast(i, p):
"""Return p or (1-p)."""
return i * p + (1 - i) * (eye - p)
for i in [0, 1]:
for j in [0, 1]:
for k in [0, 1]:
kernel2x2x2[i, j, k] = matmul(cast(i, p1), cast(j, p2),
cast(k, p3))
return kernel2x2x2
def __call__(self, shape, dtype=None, partition_info=None):
full_shape = shape if partition_info is None else partition_info.full_shape
if len(full_shape) != 2:
raise ValueError(
"Identity matrix initializer can only be used for 2D matrices.")
if dtype is None:
dtype = self.dtype
initializer = linalg_ops_impl.eye(*full_shape, dtype=dtype)
if partition_info is not None:
initializer = array_ops.slice(initializer, partition_info.var_offset,
shape)
return self.gain * initializer
def _block_orth(self, projection_matrix):
"""Construct a kernel. Used to construct orthgonal kernel.
Args:
projection_matrix: A symmetric projection matrix of size n x n.
Returns:
[projection_matrix, (1 - projection_matrix)].
"""
n = projection_matrix.shape.as_list()[0]
kernel = {}
eye = linalg_ops_impl.eye(n, dtype=self.dtype)
kernel[0] = projection_matrix
kernel[1] = eye - projection_matrix
return kernel
def __call__(self, shape, dtype=None, partition_info=None):
full_shape = shape if partition_info is None else partition_info.full_shape
if len(full_shape) != 2:
raise ValueError(
"Identity matrix initializer can only be used for 2D matrices."
)
if dtype is None:
dtype = self.dtype
gain_identity = self.diagonal_gain * linalg_ops_impl.eye(*full_shape,
dtype=dtype)
gain_ones = self.off_diagonal_gain * array_ops.ones(full_shape,
dtype=dtype)
gdn_init = math_ops.sqrt(gain_identity + math_ops.square(gain_ones))
return gdn_init
def _block_orth(self, projection_matrix):
"""Construct a kernel. Used to construct orthgonal kernel.
Args:
projection_matrix: A symmetric projection matrix of size n x n.
Returns:
[projection_matrix, (1 - projection_matrix)].
"""
n = projection_matrix.shape.as_list()[0]
kernel = {}
eye = linalg_ops_impl.eye(n, dtype=self.dtype)
kernel[0] = projection_matrix
kernel[1] = eye - projection_matrix
return kernel
def eye(num_rows,
num_columns=None,
batch_shape=None,
dtype=dtypes.float32,
name=None):
"""Construct an identity matrix, or a batch of matrices.
See also `tf.ones`, `tf.zeros`, `tf.fill`, `tf.one_hot`.
```python
# Construct one identity matrix.
tf.eye(2)
==> [[1., 0.],
[0., 1.]]
# Construct a batch of 3 identity matrices, each 2 x 2.
# batch_identity[i, :, :] is a 2 x 2 identity matrix, i = 0, 1, 2.
batch_identity = tf.eye(2, batch_shape=[3])
# Construct one 2 x 3 "identity" matrix
tf.eye(2, num_columns=3)
==> [[ 1., 0., 0.],
[ 0., 1., 0.]]
```
Args:
num_rows: Non-negative `int32` scalar `Tensor` giving the number of rows
in each batch matrix.
num_columns: Optional non-negative `int32` scalar `Tensor` giving the number
of columns in each batch matrix. Defaults to `num_rows`.
batch_shape: A list or tuple of Python integers or a 1-D `int32` `Tensor`.
If provided, the returned `Tensor` will have leading batch dimensions of
this shape.
dtype: The type of an element in the resulting `Tensor`
name: A name for this `Op`. Defaults to "eye".
Returns:
A `Tensor` of shape `batch_shape + [num_rows, num_columns]`
"""
return linalg_ops_impl.eye(num_rows,
num_columns=num_columns,
batch_shape=batch_shape,
dtype=dtype,
name=name)
def eye(num_rows,
num_columns=None,
batch_shape=None,
dtype=dtypes.float32,
name=None):
"""Construct an identity matrix, or a batch of matrices.
```python
# Construct one identity matrix.
tf.eye(2)
==> [[1., 0.],
[0., 1.]]
# Construct a batch of 3 identity matricies, each 2 x 2.
# batch_identity[i, :, :] is a 2 x 2 identity matrix, i = 0, 1, 2.
batch_identity = tf.eye(2, batch_shape=[3])
# Construct one 2 x 3 "identity" matrix
tf.eye(2, num_columns=3)
==> [[ 1., 0., 0.],
[ 0., 1., 0.]]
```
Args:
num_rows: Non-negative `int32` scalar `Tensor` giving the number of rows
in each batch matrix.
num_columns: Optional non-negative `int32` scalar `Tensor` giving the number
of columns in each batch matrix. Defaults to `num_rows`.
batch_shape: A list or tuple of Python integers or a 1-D `int32` `Tensor`.
If provided, the returned `Tensor` will have leading batch dimensions of
this shape.
dtype: The type of an element in the resulting `Tensor`
name: A name for this `Op`. Defaults to "eye".
Returns:
A `Tensor` of shape `batch_shape + [num_rows, num_columns]`
"""
return linalg_ops_impl.eye(num_rows,
num_columns=num_columns,
batch_shape=batch_shape,
dtype=dtype,
name=name)
def __call__(self, shape, dtype=dtypes.float32, **kwargs):
"""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.
**kwargs: Additional keyword arguments.
Raises:
ValueError: If the dtype is not floating point
ValueError: If the requested shape does not have exactly two axes.
"""
self._validate_kwargs(kwargs, support_partition=False)
dtype = _assert_float_dtype(dtype)
if len(shape) != 2:
raise ValueError(
"Identity matrix initializer can only be used for 2D matrices.")
initializer = linalg_ops_impl.eye(*shape, dtype=dtype)
return self.gain * initializer
def __call__(self, shape, dtype=dtypes.float32, **kwargs):
"""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.
**kwargs: Additional keyword arguments.
Raises:
ValueError: If the dtype is not floating point
ValueError: If the requested shape does not have exactly two axes.
"""
self._validate_kwargs(kwargs, support_partition=False)
dtype = _assert_float_dtype(dtype)
if len(shape) != 2:
raise ValueError("The tensor to initialize, specified by argument `shape`"
" must be at least two-dimensional. Received shape="
f"{shape}")
initializer = linalg_ops_impl.eye(*shape, dtype=dtype)
return self.gain * initializer
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
"""
partition_info = None # Keeps logic so can be readded later if necessary
dtype = _assert_float_dtype(dtype)
full_shape = shape if partition_info is None else partition_info.full_shape
if len(full_shape) != 2:
raise ValueError(
"Identity matrix initializer can only be used for 2D matrices.")
initializer = linalg_ops_impl.eye(*full_shape, dtype=dtype)
if partition_info is not None:
initializer = array_ops.slice(initializer, partition_info.var_offset,
shape)
return self.gain * initializer
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
"""
partition_info = None # Keeps logic so can be readded later if necessary
dtype = _assert_float_dtype(dtype)
full_shape = shape if partition_info is None else partition_info.full_shape
if len(full_shape) != 2:
raise ValueError(
"Identity matrix initializer can only be used for 2D matrices.")
initializer = linalg_ops_impl.eye(*full_shape, dtype=dtype)
if partition_info is not None:
initializer = array_ops.slice(initializer, partition_info.var_offset,
shape)
return self.gain * initializer
def _block_orth(self, p1, p2):
"""Construct a 2 x 2 kernel. Used to construct orthgonal kernel.
Args:
p1: A symmetric projection matrix.
p2: A symmetric projection matrix.
Returns:
A 2 x 2 kernel [[p1p2, p1(1-p2)],
[(1-p1)p2, (1-p1)(1-p2)]].
Raises:
ValueError: If the dimensions of p1 and p2 are different.
"""
if p1.shape.as_list() != p2.shape.as_list():
raise ValueError("The dimension of the matrices must be the same.")
n = p1.shape.as_list()[0]
kernel2x2 = {}
eye = linalg_ops_impl.eye(n, dtype=self.dtype)
kernel2x2[0, 0] = math_ops.matmul(p1, p2)
kernel2x2[0, 1] = math_ops.matmul(p1, (eye - p2))
kernel2x2[1, 0] = math_ops.matmul((eye - p1), p2)
kernel2x2[1, 1] = math_ops.matmul((eye - p1), (eye - p2))
return kernel2x2
def _block_orth(self, p1, p2):
"""Construct a 2 x 2 kernel. Used to construct orthgonal kernel.
Args:
p1: A symmetric projection matrix.
p2: A symmetric projection matrix.
Returns:
A 2 x 2 kernel [[p1p2, p1(1-p2)],
[(1-p1)p2, (1-p1)(1-p2)]].
Raises:
ValueError: If the dimensions of p1 and p2 are different.
"""
if p1.shape.as_list() != p2.shape.as_list():
raise ValueError("The dimension of the matrices must be the same.")
n = p1.shape.as_list()[0]
kernel2x2 = {}
eye = linalg_ops_impl.eye(n, dtype=self.dtype)
kernel2x2[0, 0] = math_ops.matmul(p1, p2)
kernel2x2[0, 1] = math_ops.matmul(p1, (eye - p2))
kernel2x2[1, 0] = math_ops.matmul((eye - p1), p2)
kernel2x2[1, 1] = math_ops.matmul((eye - p1), (eye - p2))
return kernel2x2
def __call__(self, shape, dtype, partition):
value = linalg_ops_impl.eye(*shape, dtype=dtype)
if partition is not None:
value = array_ops.slice(value, partition.offsets, partition.shape)
return value
def __call__(self, shape, dtype, shard_info):
value = linalg_ops_impl.eye(*shape, dtype=dtype)
if shard_info is not None:
value = array_ops.slice(value, shard_info.offset, shard_info.shape)
return value