def _assert_same_size(outputs: TensorStruct,
output_size: OutputSize):
r"""Check if outputs match output_size
Args:
outputs: A ``Tensor`` or a (nested) ``tuple`` of tensors
output_size: Can be an Integer, a ``torch.Size``, or a (nested)
``tuple`` of Integers or ``torch.Size``.
"""
flat_output_size = nest.flatten(output_size)
flat_output = nest.flatten(outputs)
for (output, size) in zip(flat_output, flat_output_size):
if isinstance(size, torch.Size):
if output[0].size() != size:
raise ValueError("The output size does not match"
"the the required output_size")
elif output[0].size()[-1] != size:
raise ValueError(
"The output size does not match the the required output_size")
def _sum_output_size(output_size: OutputSize) -> int:
r"""Return sum of all dim values in :attr:`output_size`
Args:
output_size: Can be an ``Integer``, a ``torch.Size``, or a (nested)
``tuple`` of ``Integers`` or ``torch.Size``.
"""
flat_output_size = nest.flatten(output_size)
if isinstance(flat_output_size[0], torch.Size):
size_list = [0] * len(flat_output_size)
for (i, shape) in enumerate(flat_output_size):
size_list[i] = np.prod([dim for dim in shape])
else:
size_list = flat_output_size
sum_output_size = sum(size_list)
return sum_output_size
def forward(self, # type: ignore
inputs: TensorStruct
) -> Any:
r"""Transforms inputs to have the same structure as with
:attr:`output_size`. Values of the inputs are not changed.
:attr:`inputs` must either have the same structure, or have the same
number of elements with :attr:`output_size`.
Args:
inputs: The input (structure of) tensor to pass forward.
:returns:
A (structure of) tensors that re-packs :attr:`inputs` to have
the specified structure of :attr:`output_size`.
"""
output = inputs
flat_input = nest.flatten(inputs)
output = nest.pack_sequence_as(
self._output_size, flat_input)
return output
def test_constant_connector(self):
r"""Tests the logic of
:class:`~texar.modules.connectors.ConstantConnector`.
"""
state_size = namedtuple('LSTMStateTuple', ['c', 'h'])(256, 256)
connector_0 = ConstantConnector(state_size)
decoder_initial_state_0 = connector_0(self._batch_size)
connector_1 = ConstantConnector(state_size, hparams={"value": 1.})
decoder_initial_state_1 = connector_1(self._batch_size)
s_0 = decoder_initial_state_0
s_1 = decoder_initial_state_1
self.assertEqual(nest.flatten(s_0)[0][0, 0], 0.)
self.assertEqual(nest.flatten(s_1)[0][0, 0], 1.)
size = torch.Size([1, 2, 3])
connector_size_0 = ConstantConnector(size, hparams={"value": 2.})
size_tensor = connector_size_0(self._batch_size)
self.assertEqual(
torch.Size([self._batch_size]) + size, size_tensor.size())
self.assertEqual(size_tensor[0][0, 0, 0], 2.)
tuple_size_1 = (torch.Size([1, 2, 3]), torch.Size([4, 5, 6]))
connector_size_1 = ConstantConnector(tuple_size_1,
hparams={"value": 3.})
tuple_size_tensor = connector_size_1(self._batch_size)
tuple_size_tensor_0 = tuple_size_tensor[0]
tuple_size_tensor_1 = tuple_size_tensor[1]
self.assertEqual(
torch.Size([self._batch_size]) + torch.Size([1, 2, 3]),
tuple_size_tensor_0.size())
self.assertEqual(tuple_size_tensor_0[0][0, 0, 0], 3.)
self.assertEqual(
torch.Size([self._batch_size]) + torch.Size([4, 5, 6]),
tuple_size_tensor_1.size())
self.assertEqual(tuple_size_tensor_1[0][0, 0, 0], 3.)
tuple_size_2 = (5, 10)
connector_size_2 = ConstantConnector(tuple_size_2,
hparams={"value": 4.})
tuple_size_tensor = connector_size_2(self._batch_size)
tuple_size_tensor_0 = tuple_size_tensor[0]
tuple_size_tensor_1 = tuple_size_tensor[1]
self.assertEqual(
torch.Size([self._batch_size]) + torch.Size([5]),
tuple_size_tensor_0.size())
self.assertEqual(tuple_size_tensor_0[0][0], 4.)
self.assertEqual(
torch.Size([self._batch_size]) + torch.Size([10]),
tuple_size_tensor_1.size())
self.assertEqual(tuple_size_tensor_1[0][0], 4.)
tuple_size_3 = (torch.Size([1, 2, 3]), 10)
connector_size_3 = ConstantConnector(tuple_size_3,
hparams={"value": 4.})
tuple_size_tensor = connector_size_3(self._batch_size)
tuple_size_tensor_0 = tuple_size_tensor[0]
tuple_size_tensor_1 = tuple_size_tensor[1]
self.assertEqual(
torch.Size([self._batch_size]) + torch.Size([1, 2, 3]),
tuple_size_tensor_0.size())
self.assertEqual(tuple_size_tensor_0[0][0, 0, 0], 4.)
self.assertEqual(
torch.Size([self._batch_size]) + torch.Size([10]),
tuple_size_tensor_1.size())
self.assertEqual(tuple_size_tensor_1[0][0], 4.)
def _mlp_transform(inputs: TensorStruct,
output_size: OutputSize,
linear_layer: Optional[LinearLayer] = None,
activation_fn: Optional[ActivationFn] = None,
) -> Any:
r"""Transforms inputs through a fully-connected layer that creates
the output with specified size.
Args:
inputs: A ``Tensor`` of shape ``[batch_size, ..., finale_state]``
(i.e., batch-major), or a (nested) tuple of such elements.
A Tensor or a (nested) tuple of Tensors with shape
``[max_time, batch_size, ...]`` (i.e., time-major) can
be transposed to batch-major using
:func:`~texar.utils.transpose_batch_time` prior to this function.
output_size: Can be an ``Integer``, a ``torch.Size``, or a (nested)
``tuple`` of ``Integers`` or ``torch.Size``.
activation_fn: Activation function applied to the output.
:returns:
If :attr:`output_size` is an ``Integer`` or a ``torch.Size``,
returns a ``Tensor`` of shape ``[batch_size, *, output_size]``.
If :attr:`output_size` is a ``tuple`` of Integers or torch.Size,
returns a ``tuple`` having the same structure as:attr:`output_size`,
where each element ``Tensor`` has the same size as
defined in :attr:`output_size`.
"""
# Flatten inputs
flat_input = nest.flatten(inputs)
if len(flat_input[0].size()) == 1:
batch_size = 1
else:
batch_size = flat_input[0].size(0)
flat_input = [x.view(-1, x.size(-1)) for x in flat_input]
concat_input = torch.cat(flat_input, 0)
# Get output dimension
flat_output_size = nest.flatten(output_size)
if isinstance(flat_output_size[0], torch.Size):
size_list = [0] * len(flat_output_size)
for (i, shape) in enumerate(flat_output_size):
size_list[i] = np.prod([dim for dim in shape])
else:
size_list = flat_output_size
if linear_layer is not None:
fc_output = linear_layer(concat_input)
if activation_fn is not None:
fc_output = activation_fn(fc_output)
elif linear_layer is None and activation_fn is None:
fc_output = concat_input
flat_output = split(fc_output, size_list, dim=1) # type: ignore
flat_output = list(flat_output)
for i, _ in enumerate(flat_output):
final_state = flat_output[i].size(-1)
flat_output[i] = flat_output[i].view(batch_size, -1, final_state)
flat_output[i] = torch.squeeze(flat_output[i], 1)
if isinstance(flat_output_size[0], torch.Size):
for (i, shape) in enumerate(flat_output_size):
flat_output[i] = torch.reshape(
flat_output[i], (-1, ) + shape)
output = nest.pack_sequence_as(structure=output_size,
flat_sequence=flat_output)
return output
