def build(self, input_shape):
# Create a trainable weight variable for this layer.
self._inferred_dimension_permutation = infer_dimension_permutation(input_shape) \
if self.dimension_permutation == DimensionPermutation.AUTO \
else self.dimension_permutation
if self._inferred_dimension_permutation == DimensionPermutation.BATCH_FIRST:
self._num_nodes_in = num_nodes_in = np.prod(input_shape[1:])
else:
num_scopes_in, num_decomps_in, _, num_nodes_in = input_shape
self._num_nodes_in = num_nodes_in
if num_scopes_in != 1 or num_decomps_in != 1:
raise ValueError("Number of scopes and decomps must both be 1")
initializer = self.accumulator_initializer
accumulator_constraint = self.linear_accumulator_constraint
if self.logspace_accumulators:
initializer = logspace_wrapper_initializer(initializer)
accumulator_constraint = None
self.accumulators = self.add_weight(
name='weights',
shape=(num_nodes_in, ),
initializer=initializer,
regularizer=self.accumulator_regularizer,
constraint=accumulator_constraint)
def build(self, input_shape: Tuple[Optional[int], ...]) -> None:
"""
Build the internal components for this leaf layer.
Args:
input_shape: Shape of the input Tensor.
"""
# Create a trainable weight variable for this layer.
_, num_scopes_vertical, num_scopes_horizontal, num_channels_in = input_shape
weights_shape = (1, 1, num_channels_in, self.num_sums)
initializer = self.accumulator_initializer
accumulator_contraint = self.linear_accumulator_constraint
if self.logspace_accumulators:
initializer = logspace_wrapper_initializer(initializer)
accumulator_contraint = self.logspace_accumulator_constraint
self._accumulators = self.add_weight(
name="sum_weights",
shape=weights_shape,
initializer=initializer,
regularizer=self.accumulator_regularizer,
constraint=accumulator_contraint,
)
super(DenseSum, self).build(input_shape)
def build(self, input_shape):
# Create a trainable weight variable for this layer.
self._num_scopes, self._num_decomps, _, num_nodes_in = input_shape
weights_shape = (self._num_scopes, self._num_decomps, num_nodes_in, self.num_sums)
initializer = self.accumulator_initializer
accumulator_constraint = self.linear_accumulator_constraint
if self.logspace_accumulators:
initializer = logspace_wrapper_initializer(self.accumulator_initializer)
accumulator_constraint = None
self._accumulators = self.add_weight(
name='sum_weights', shape=weights_shape, initializer=initializer,
regularizer=self.accumulator_regularizer, constraint=accumulator_constraint
)
super(DenseSum, self).build(input_shape)
def build(self, input_shape):
# Create a trainable weight variable for this layer.
_, num_scopes_vertical, num_scopes_horizontal, num_channels_in = input_shape
weights_shape = (1, 1, num_channels_in, self.num_sums)
initializer = self.accumulator_initializer
accumulator_contraint = self.linear_accumulator_constraint
if self.logspace_accumulators:
initializer = logspace_wrapper_initializer(initializer)
accumulator_contraint = None
self.accumulators = self.add_weight(
name='sum_weights', shape=weights_shape, initializer=initializer,
regularizer=self.accumulator_regularizer, constraint=accumulator_contraint
)
super(Conv2DSum, self).build(input_shape)
def build(self, input_shape):
_, num_scopes_in, num_decomps_in, self._num_nodes_in = input_shape
if num_scopes_in != 1 or num_decomps_in != 1:
raise ValueError("Number of scopes and decomps must both be 1")
initializer = self.accumulator_initializer
accumulator_constraint = self.linear_accumulator_constraint
if self.logspace_accumulators:
initializer = logspace_wrapper_initializer(initializer)
accumulator_constraint = None
self.accumulators = self.add_weight(
name='weights',
shape=(self._num_nodes_in, ),
initializer=initializer,
regularizer=self.accumulator_regularizer,
constraint=accumulator_constraint)
def build(self, input_shape: Tuple[Optional[int], ...]) -> None:
"""
Build the internal components for this layer.
Args:
input_shape: Shape of the input Tensor.
"""
# Create a trainable weight variable for this layer.
_, self._num_scopes, self._num_decomps, self._num_nodes_in = input_shape
weights_shape = (
self._num_scopes,
self._num_decomps,
self._num_nodes_in,
self.num_sums,
)
initializer = self.accumulator_initializer
accumulator_constraint = self.linear_accumulator_constraint
if self.logspace_accumulators:
initializer = logspace_wrapper_initializer(
self.accumulator_initializer)
accumulator_constraint = self.logspace_accumulator_constraint
self._accumulators = self.add_weight(
name="sum_weights",
shape=weights_shape,
initializer=initializer,
regularizer=self.accumulator_regularizer,
constraint=accumulator_constraint,
)
if accumulator_constraint is not None:
self._accumulators.assign(
accumulator_constraint(self._accumulators))
self._forward_normalize = not isinstance(
accumulator_constraint,
(GreaterEqualEpsilonNormalized, LogNormalized))
super(DenseSum, self).build(input_shape)