def reshape_spec(args, inputs, outputs):
outputs[0]._dtype = inputs[0].dtype
try:
shape = args['dims']
out_shape = []
n_elements, n_elements_known = None, None
try:
for i, s in enumerate(shape):
if s == -1:
out_shape.append(1)
elif s == 0:
out_shape.append(inputs[0].shape[i])
else:
out_shape.append(s)
except IndexError:
out_shape = None
try:
n_elements = math_util.prod(inputs[0].shape)
n_elements_known = math_util.prod(out_shape)
except TypeError:
pass
for i, s in enumerate(shape):
if s == -1:
try:
out_shape[i] = n_elements // n_elements_known
except TypeError:
out_shape[i] = None
except (KeyError, TypeError):
out_shape = None
outputs[0]._shape = tuple(out_shape) if out_shape is not None else None
return outputs
def flatten_spec(args, inputs, outputs):
outputs[0].dtype = inputs[0].dtype
keep_axes = args['keep_axes']
axis, num_axes = args['axis'], args['num_axes']
if keep_axes is not None:
out_shape = [None] * keep_axes
else:
out_shape = None
try:
in_shape = list(inputs[0].shape[:])
if keep_axes is not None:
if len(in_shape) <= keep_axes:
out_shape[:len(in_shape)] = in_shape
else:
for i in range(keep_axes - 1):
out_shape[i] = in_shape[i]
try:
out_shape[keep_axes - 1] = \
math_util.prod(in_shape[keep_axes - 1:])
except (TypeError, IndexError):
out_shape[keep_axes - 1] = None
else:
if num_axes == -1:
num_axes = len(in_shape) - axis
num_axes = max(num_axes, 1)
try:
num_flatten = math_util.prod(in_shape[axis:axis + num_axes])
except TypeError:
num_flatten = None
out_shape = in_shape[: axis] + [num_flatten] + in_shape[axis + num_axes:]
except (TypeError, IndexError):
pass
outputs[0].shape = out_shape
return outputs
def numel(self):
"""Return the total number of elements.
Returns
-------
int
The number of elements.
"""
return math_util.prod(self)
def size(self):
"""Return the total number of elements in this tensor.
Returns
-------
int
The total count of elements.
"""
if self._shape is None:
return 0
if None in self._shape:
return numpy.inf
return math_util.prod(self._shape)
def size(self):
"""Return the total number of elements in this tensor.
Returns
-------
int
The total count of elements.
"""
if self._is_variable:
return self._impl.size
if self._shape is None:
return 0
if None in self._shape:
return float('inf')
return math_util.prod(self._shape)
def _calculate_fan_in_and_fan_out(tensor):
"""Return the fan value according to tensor size."""
dimensions = tensor.ndimension()
if dimensions < 2:
raise ValueError("Excepted 2 or higher tensor dimensions.")
if dimensions == 2:
fan_in = tensor.size(1)
fan_out = tensor.size(0)
else:
num_input = tensor.size(1)
num_output = tensor.size(0)
receptive_field_size = 1
if tensor.dim() > 2:
receptive_field_size = math_util.prod(tensor.shape[2:])
fan_in = num_input * receptive_field_size
fan_out = num_output * receptive_field_size
return fan_in, fan_out
def flatten_spec(args, inputs, outputs):
outputs[0]._dtype = inputs[0].dtype
axis = args['axis']
end_axis = args.get('end_axis', None)
end_axis = axis if end_axis is None else end_axis
try:
in_shape = list(inputs[0].shape[:])
out_shape = in_shape[:axis]
num_axes = len(in_shape[axis:end_axis]) + 1
try:
out_shape += [math_util.prod(in_shape[axis:axis + num_axes])]
except TypeError:
out_shape += [None]
out_shape += in_shape[axis + num_axes:]
outputs[0]._shape = tuple(out_shape)
except (TypeError, IndexError):
outputs[0]._shape = None
return outputs
def _create_weights(self):
"""Create a flat weights."""
gate_size = self._hidden_size * self._num_gates
# Compute the shape of weight and bias.
matrix_shapes, bias_shapes = [], []
for layer in range(self._num_layers):
for direction in range(self._num_directions):
layer_input_size = self._input_size if layer == 0 \
else self._hidden_size * self._num_directions
w_ih_shape = [gate_size, layer_input_size]
w_hh_shape = [gate_size, self._hidden_size]
b_ih_shape, b_hh_shape = [gate_size], [gate_size]
matrix_shapes.extend([w_ih_shape, w_hh_shape])
bias_shapes.extend([b_ih_shape, b_hh_shape])
# Create single float32 weights.
weights_count = 0
self._weights_shapes = matrix_shapes + bias_shapes
for shape in self._weights_shapes:
weights_count += math_util.prod(shape)
self._weights = Tensor([weights_count])
self._weights.requires_grad = True
def flatten_parameters(self):
"""Flatten parameters into a single weights."""
gate_size = self._num_gates * self.hidden_size
# Compute the shape of weight and bias.
matrix_shapes, bias_shapes = [], []
for layer in range(self.num_layers):
for direction in range(int(self.bidirectional) + 1):
layer_input_size = self.input_size if layer == 0 \
else self.hidden_size * self.num_directions
w_ih_shape = [gate_size, layer_input_size]
w_hh_shape = [gate_size, self.hidden_size]
b_ih_shape, b_hh_shape = [gate_size], [gate_size]
matrix_shapes.extend([w_ih_shape, w_hh_shape])
bias_shapes.extend([b_ih_shape, b_hh_shape])
# Compute total number of parameters.
self._weights_count = 0
self._weights_shapes = matrix_shapes + bias_shapes
for shape in self._weights_shapes:
self._weights_count += math_util.prod(shape)
# Create the flat float32 weights.
self.weights = Parameter(Tensor(self._weights_count))
def repeat_spec(args, inputs, outputs):
outputs[0]._dtype = inputs[0].dtype
if 'repeats_desc' in args:
return outputs
axis, repeats = args['axis'], args['repeats']
if axis is None:
try:
num_elements = math_util.prod(inputs[0].shape[:])
outputs[0]._shape = (num_elements * repeats, )
except TypeError:
outputs[0]._shape = (None, )
else:
try:
out_shape = list(inputs[0].shape[:])
except TypeError:
return outputs
if axis < len(out_shape):
try:
out_shape[axis] *= repeats
except TypeError:
out_shape[axis] = None
outputs[0]._shape = tuple(out_shape)
return outputs
