def forward(
self, # type: ignore
input: torch.Tensor,
sequence_length: Union[torch.LongTensor, List[int]] = None,
dtype: Optional[torch.dtype] = None) -> torch.Tensor:
r"""Feeds forward inputs through the network layers and returns outputs.
Args:
input: The inputs to the network, which is a 3D tensor.
sequence_length (optional): An :tensor:`LongTensor` of shape
``[batch_size]`` or a python array containing the length of
each element in :attr:`inputs`. If given, time steps beyond
the length will first be masked out before feeding to the
layers.
dtype (optional): Type of the inputs. If not provided,
infers from inputs automatically.
Returns:
The output of the final layer.
"""
if sequence_length is not None:
input = mask_sequences(input,
sequence_length,
dtype=dtype,
time_major=False)
return super().forward(input)
def _discount_reward_tensor_1d(reward,
sequence_length,
discount=1.,
dtype=None):
if sequence_length is None:
raise ValueError('sequence_length must not be `None` for 1D reward.')
batch_size = tf.shape(reward)[0]
max_seq_length = tf.reduce_max(sequence_length)
dtype = dtype or reward.dtype
if discount == 1.:
dmat = tf.ones(tf.concat([[batch_size], [max_seq_length]], 0),
dtype=dtype)
else:
mask = tf.sequence_mask(sequence_length, dtype=dtype)
mask = tf.concat([mask[:, 1:], tf.zeros_like(mask[:, -1:])], axis=1)
# Make each row = [discount, ..., discount, 1, ..., 1]
dmat = mask * discount + (1 - mask)
dmat = tf.cumprod(dmat, axis=1, reverse=True)
disc_reward = dmat * tf.expand_dims(reward, -1)
disc_reward = mask_sequences(disc_reward,
sequence_length,
dtype=dtype,
tensor_rank=2)
return disc_reward
def _build(
self, # pylint: disable=arguments-differ
inputs,
sequence_length=None,
dtype=None,
mode=None):
"""Feeds forward inputs through the network layers and returns outputs.
Args:
inputs: The inputs to the network, which is a 3D tensor.
sequence_length (optional): An int tensor of shape `[batch_size]`
containing the length of each element in :attr:`inputs`.
If given, time steps beyond the length will first be masked out
before feeding to the layers.
dtype (optional): Type of the inputs. If not provided, infers
from inputs automatically.
mode (optional): A tensor taking value in
:tf_main:`tf.estimator.ModeKeys <estimator/ModeKeys>`, including
`TRAIN`, `EVAL`, and `PREDICT`. If `None`,
:func:`texar.global_mode` is used.
Returns:
The output of the final layer.
"""
if sequence_length is not None:
inputs = mask_sequences(inputs,
sequence_length,
dtype=dtype,
time_major=False,
tensor_rank=3)
return super(Conv1DNetwork, self)._build(inputs, mode=mode)
def _discount_reward_py_1d(reward, sequence_length, discount=1., dtype=None):
if sequence_length is None:
raise ValueError('sequence_length must not be `None` for 1D reward.')
reward = np.array(reward)
sequence_length = np.array(sequence_length)
batch_size = reward.shape[0]
max_seq_length = np.max(sequence_length)
dtype = dtype or reward.dtype
if discount == 1.:
dmat = np.ones([batch_size, max_seq_length], dtype=dtype)
else:
steps = np.tile(np.arange(max_seq_length), [batch_size, 1])
mask = np.asarray(steps < (sequence_length - 1)[:, None], dtype=dtype)
# Make each row = [discount, ..., discount, 1, ..., 1]
dmat = mask * discount + (1 - mask)
dmat = np.cumprod(dmat[:, ::-1], axis=1)[:, ::-1]
disc_reward = dmat * reward[:, None]
disc_reward = mask_sequences(disc_reward, sequence_length, dtype=dtype)
#mask = np.asarray(steps < sequence_length[:, None], dtype=dtype)
#disc_reward = mask * disc_reward
return disc_reward
def _forward_output_layers(inputs,
input_size,
output_layer,
time_major,
hparams,
mode,
sequence_length=None):
"""Forwards inputs through the output layers.
Args:
inputs: A Tensor of shape `[batch_size, max_time] + input_size` if
:attr:`time_major=False`, or shape
`[max_time, batch_size] + input_size` if :attr:`time_major=True`.
Returns:
A pair :attr:`(outputs, outputs_size), where
- :attr:`outputs`: A Tensor of shape \
`[batch_size, max_time] + outputs_size`.
- :attr:`outputs_size`: An `int` or 1D `int` array representing the \
output size.
"""
if output_layer is None:
return inputs, input_size
if hparams is None:
# output_layer was passed in from the constructor
if isinstance(output_layer, (list, tuple)):
raise ValueError('output_layer must not be a list or tuple.')
output, output_size = _forward_single_output_layer(
inputs, input_size, output_layer)
else:
# output_layer was built based on hparams
output_layer = _to_list(output_layer)
dropout_layer_ids = _to_list(hparams.dropout_layer_ids)
if len(dropout_layer_ids) > 0:
training = is_train_mode(mode)
output = inputs
output_size = input_size
for i, layer in enumerate(output_layer):
if i in dropout_layer_ids:
output = _apply_dropout(output, time_major, hparams, training)
output, output_size = _forward_single_output_layer(
output, output_size, layer)
if len(output_layer) in dropout_layer_ids:
output = _apply_dropout(output, time_major, hparams, training)
if sequence_length is not None:
output = mask_sequences(output,
sequence_length,
time_major=time_major,
tensor_rank=3)
return output, output_size
def test_mask_sequences(self):
"""Tests :func:`texar.utils.shapes.mask_sequences`.
"""
seq = np.ones([3, 4, 3], dtype=np.int32)
seq_length = np.array([3, 2, 1], dtype=np.int32)
masked_seq = shapes.mask_sequences(seq, seq_length)
self.assertEqual(masked_seq.shape, seq.shape)
seq_sum = np.sum(masked_seq, axis=(1, 2))
np.testing.assert_array_equal(seq_sum, seq_length * 3)
def test_mask_sequences(self):
r"""Tests :func:`texar.utils.shapes.mask_sequences`.
"""
seq = torch.ones(3, 4, 3, dtype=torch.int32)
seq_length = torch.tensor([3, 2, 1], dtype=torch.int32)
masked_seq = shapes.mask_sequences(seq, seq_length)
np.testing.assert_array_equal(masked_seq.shape, seq.shape)
seq_sum = torch.sum(masked_seq, dim=(1, 2))
np.testing.assert_array_equal(seq_sum, seq_length * 3)
def _build(self, # pylint: disable=arguments-differ
inputs,
sequence_length=None,
dtype=None,
time_major=False,
mode=None):
if sequence_length is not None:
inputs = mask_sequences(
inputs, sequence_length, dtype=dtype, time_major=time_major,
tensor_rank=3)
return super(Conv1DNetwork, self)._build(inputs, mode=mode)
def _discount_reward_py_2d(reward, sequence_length=None,
discount=1., dtype=None):
if sequence_length is not None:
reward = mask_sequences(reward, sequence_length, dtype=dtype)
dtype = dtype or reward.dtype
if discount == 1.:
disc_reward = np.cumsum(
reward[:, ::-1], axis=1, dtype=dtype)[:, ::-1]
else:
disc_reward = np.copy(reward)
for i in range(reward.shape[1]-2, -1, -1):
disc_reward[:, i] += disc_reward[:, i+1] * discount
return disc_reward
def forward(self, # type: ignore
positions: Optional[torch.LongTensor] = None,
sequence_length: Optional[torch.LongTensor] = None, **kwargs) \
-> torch.Tensor:
r"""Embeds.
Either :attr:`positions` or :attr:`sequence_length` is required:
- If both are given, :attr:`sequence_length` is used to mask out
embeddings of those time steps beyond the respective sequence
lengths.
- If only :attr:`sequence_length` is given, then positions
from `0` to `sequence_length - 1` are embedded.
Args:
positions (optional): An :tensor:`LongTensor` containing the
position IDs to embed.
sequence_length (optional): An :tensor:`LongTensor` of shape
``[batch_size]``. Time steps beyond
the respective sequence lengths will have zero-valued
embeddings.
Returns:
A Tensor of shape ``[batch_size, position_size, dim]``.
"""
if positions is None:
if sequence_length is None:
raise ValueError(
'Either `positions` or `sequence_length` is required.')
max_length = sequence_length.max()
batch_size = sequence_length.size(0)
inputs = torch.arange(max_length).to(device=sequence_length.device)
inputs = inputs.expand(batch_size, max_length)
else:
inputs = positions
if self._cache_embeddings:
outputs = F.embedding(inputs, self.signal, **kwargs)
else:
outputs = self._compute_embeddings(inputs, self.inv_timescales)
if sequence_length is not None:
outputs = mask_sequences(outputs, sequence_length)
return outputs
def _forward_output_layers(
inputs: torch.Tensor,
output_layer: Optional[nn.Module],
time_major: bool,
sequence_length: Optional[Union[torch.LongTensor, List[int]]] = None) \
-> Tuple[torch.Tensor, int]:
r"""Forwards inputs through the output layers.
Args:
inputs: A Tensor of shape ``[batch_size, max_time] + input_size`` if
:attr:`time_major` is `False`, or shape
``[max_time, batch_size] + input_size`` if :attr:`time_major` is
`True`.
output_layer (optional): :torch_nn:`Sequential` or :torch_nn:`Module`
of output layers.
time_major (bool): The shape format of the :attr:`inputs` and
:attr:`outputs` Tensors. If `True`, these tensors are of shape
`[max_time, batch_size, input_size]`. If `False` (default),
these tensors are of shape `[batch_size, max_time, input_size]`.
sequence_length (optional): A 1D :tensor:`LongTensor` of shape
``[batch_size]``. Sequence lengths of the batch inputs. Used to
copy-through state and zero-out outputs when past a batch element's
sequence length.
Returns:
A pair :attr:`(outputs, outputs_size), where
- :attr:`outputs`: A Tensor of shape
`[batch_size, max_time] + outputs_size`.
- :attr:`outputs_size`: An `int` representing the output size.
"""
if output_layer is None:
return inputs, inputs.shape[-1]
output = output_layer(inputs)
if sequence_length is not None:
output = mask_sequences(output, sequence_length, time_major=time_major)
output_size = output.shape[-1]
return output, output_size
def _discount_reward_tensor_1d(reward: torch.Tensor,
sequence_length: torch.LongTensor,
discount: float = 1.) -> torch.Tensor:
r"""Computes discounted reward.
Args:
reward: 1D Tensor with shape `[batch_size]`.
sequence_length: A Tensor of shape `[batch_size]`.
Time steps beyond the respective sequence lengths will be masked.
discount (float): A scalar. The discount factor.
Returns:
A 2D Tensor of the discounted reward.
"""
if sequence_length is None:
raise ValueError('sequence_length must not be `None` for 1D reward.')
if not isinstance(sequence_length, torch.Tensor):
sequence_length = torch.tensor(sequence_length,
dtype=torch.int64,
device=reward.device)
batch_size = reward.shape[0]
max_seq_length = torch.max(sequence_length)
dtype: torch.dtype = reward.dtype
if discount == 1.:
disc_reward = torch.unsqueeze(reward,
-1).expand(batch_size, max_seq_length)
else:
mask = sequence_mask(sequence_length, dtype=dtype)
mask = torch.cat((mask[:, 1:], torch.zeros_like(mask[:, -1:])), dim=1)
# Make each row = [discount, ..., discount, 1, ..., 1]
dmat = mask * discount + (1 - mask)
dmat = torch.flip(dmat, (1, ))
dmat = torch.cumprod(dmat, dim=1)
dmat = torch.flip(dmat, (1, ))
disc_reward = dmat * torch.unsqueeze(reward, -1)
disc_reward = mask_sequences(disc_reward, sequence_length, dtype=dtype)
return disc_reward
def _discount_reward_tensor_2d(reward, sequence_length=None,
discount=1., dtype=None):
if sequence_length is not None:
reward = mask_sequences(
reward, sequence_length, dtype=dtype, tensor_rank=2)
if discount == 1.:
disc_reward = tf.cumsum(reward, axis=1, reverse=True)
else:
# [max_time, batch_size]
rev_reward_T = tf.transpose(tf.reverse(reward, [1]), [1, 0])
rev_reward_T_cum = tf.scan(
fn=lambda acc, cur: cur + discount * acc,
elems=rev_reward_T,
initializer=tf.zeros_like(reward[:, 1]),
back_prop=False)
disc_reward = tf.reverse(
tf.transpose(rev_reward_T_cum, [1, 0]), [1])
return disc_reward
def _discount_reward_tensor_2d(reward: torch.Tensor,
sequence_length: Optional[
torch.LongTensor] = None,
discount: float = 1.) -> torch.Tensor:
r"""Computes discounted reward.
Args:
reward: 2D Tensor with shape `[batch_size, max_time]`.
sequence_length (optional): A Tensor of shape `[batch_size]`.
Time steps beyond the respective sequence lengths will be masked.
discount (float): A scalar. The discount factor.
Returns:
A 2D Tensor of the discounted reward.
"""
dtype: torch.dtype = reward.dtype
if sequence_length is not None:
reward = mask_sequences(reward, sequence_length, dtype=dtype)
if discount == 1.:
reward = torch.flip(reward, (1, ))
disc_reward = torch.cumsum(reward, dim=1)
disc_reward = torch.flip(disc_reward, (1, ))
else:
# [max_time, batch_size]
rev_reward_T = torch.flip(reward, (1, )).permute(1, 0)
res = []
acc = torch.zeros_like(reward[:, 1])
for i in range(rev_reward_T.shape[0]):
cur = rev_reward_T[i]
acc = cur + discount * acc
res.append(acc)
rev_reward_T_cum = torch.stack(res, dim=0)
disc_reward = torch.flip(rev_reward_T_cum.permute(1, 0), (1, ))
return disc_reward
def _build(self, # pylint: disable=arguments-differ
memory,
memory_sequence_length=None,
memory_attention_bias=None,
inputs=None,
sequence_length=None,
decoding_strategy='train_greedy',
beam_width=1,
alpha=0,
start_tokens=None,
end_token=None,
max_decoding_length=None,
mode=None):
"""Performs decoding.
The decoder supports 4 decoding strategies. For the first 3 strategies,
set :attr:`decoding_strategy` to the respective string.
- **"train_greedy"**: decoding in teacher-forcing fashion \
(i.e., feeding \
ground truth to decode the next step), and for each step sample \
is obtained by taking the `argmax` of logits. \
Argument :attr:`inputs` is required for this strategy. \
:attr:`sequence_length` is optional.
- **"infer_greedy"**: decoding in inference fashion (i.e., feeding \
`generated` sample to decode the next step), and for each
step sample is obtained by taking the `argmax` of logits.\
Arguments :attr:`(start_tokens, end_token)` are \
required for this strategy, and argument \
:attr:`max_decoding_length` is optional.
- **"infer_sample"**: decoding in inference fashion, and for each step\
sample is obtained by `random sampling` from the logits.
Arguments :attr:`(start_tokens, end_token)` are \
required for this strategy, and argument \
:attr:`max_decoding_length` is optional.
- **Beam Search**: set :attr:`beam_width` to > 1 to use beam search \
decoding.\
Arguments :attr:`(start_tokens, end_token)` are \
required, and argument \
:attr:`max_decoding_length` is optional.
Args:
memory: The memory to attend, e.g., the output of an RNN encoder.
A Tensor of shape `[batch_size, memory_max_time, dim]`.
memory_sequence_length (optional): A Tensor of shape `[batch_size]`
containing the sequence lengths for the batch entries in
memory. Used to create attention bias of
:attr:`memory_attention_bias` is not given. Ignored if
`memory_attention_bias` is provided.
memory_attention_bias (optional): A Tensor of shape
`[batch_size, num_heads, memory_max_time, dim]`.
An attention bias typically sets the value of a padding
position to a large negative value for masking. If not given,
:attr:`memory_sequence_length` is used to automatically
create an attention bias.
inputs (optional): Input tensor for teacher forcing decoding, of
shape `[batch_size, target_max_time, emb_dim]` containing the
target sequence word embeddings.
Used when :attr:`decoding_strategy` is set to "train_greedy".
sequence_length (optional): A Tensor of shape `[batch_size]`,
containing the sequence length of :attr:`inputs`.
Tokens beyond the respective sequence length are masked out.
Used when :attr:`decoding_strategy` is set to
"train_greedy".
decoding_strategy (str): A string specifying the decoding
strategy, including "train_greedy", "infer_greedy",
"infer_sample".
Different arguments are required based on the
strategy. See above for details. Ignored if
:attr:`beam_width` > 1.
beam_width (int): Set to > 1 to use beam search.
alpha (float): Length penalty coefficient.
Refer to https://arxiv.org/abs/1609.08144
for more details.
tart_tokens (optional): An int Tensor of shape `[batch_size]`,
containing the start tokens.
Used when `decoding_strategy` = "infer_greedy" or
"infer_sample", or `beam_width` > 1.
end_token (optional): An int 0D Tensor, the token that marks end
of decoding.
Used when `decoding_strategy` = "infer_greedy" or
"infer_sample", or `beam_width` > 1.
max_decoding_length (optional): An int scalar Tensor indicating
the maximum allowed number of decoding steps.
If `None` (default), use "max_decoding_length" defined in
:attr:`hparams`. Ignored in "train_greedy" decoding.
mode (optional): A tensor taking value in
:tf_main:`tf.estimator.ModeKeys <estimator/ModeKeys>`, including
`TRAIN`, `EVAL`, and `PREDICT`. Controls dropout mode.
If `None` (default), :func:`texar.global_mode`
is used.
Returns:
- For **"train_greedy"** decoding, returns an instance of \
:class:`~texar.modules.TransformerDecoderOutput` which contains\
`sample_id` and `logits`.
- For **"infer_greedy"** and **"infer_sample"** decoding, returns\
a tuple `(outputs, sequence_lengths)`, where `outputs` is an \
instance of :class:`~texar.modules.TransformerDecoderOutput` as\
in "train_greedy", and `sequence_lengths` is a Tensor of shape\
`[batch_size]` containing the length of each sample.
- For **beam_search** decoding, returns a `dict` containing keys\
"sample_id" and "log_prob".
- **"sample_id"** is an int Tensor of shape \
`[batch_size, max_time, beam_width]` containing generated\
token indexes. `sample_id[:,:,0]` is the highest-probable \
sample.
- **"log_porb"** is a float Tensor of shape \
`[batch_size, beam_width]` containing the log probability \
of each sequence sample.
"""
if memory_attention_bias is None:
if memory_sequence_length is None:
raise ValueError(
"`memory_sequence_length` is required if "
"`memory_attention_bias` is not given.")
#enc_padding = 1 - mask_sequences(tf.ones_like(memory),
# memory_sequence_length,
# tensor_rank=3)[:, :, 0]
enc_padding = 1 - tf.sequence_mask(
memory_sequence_length, tf.shape(memory)[1], dtype=tf.float32)
memory_attention_bias = attn.attention_bias_ignore_padding(
enc_padding)
if beam_width <= 1 and decoding_strategy == 'train_greedy':
if sequence_length is not None:
inputs = mask_sequences(inputs, sequence_length, tensor_rank=3)
decoder_self_attention_bias = (
attn.attention_bias_lower_triangle(
shape_list(inputs)[1]))
target_inputs = inputs * self._hparams.dim**0.5
_, lengths, channels = shape_list(target_inputs)
pos_embeds = self.position_embedder(lengths, channels)
inputs = target_inputs + pos_embeds
decoder_output = self._self_attention_stack(
inputs,
memory,
decoder_self_attention_bias=decoder_self_attention_bias,
memory_attention_bias=memory_attention_bias,
cache=None,
mode=mode)
logits = self.output_layer(decoder_output)
preds = tf.to_int32(tf.argmax(logits, axis=-1))
output = TransformerDecoderOutput(
logits=logits,
sample_id=preds
)
rets = output
else: # Inference decoding
if max_decoding_length is None:
max_decoding_length = self._hparams.max_decoding_length
if beam_width <= 1:
logits, preds, sequence_length = self._infer_decoding(
self._prepare_tokens_to_embeds,
start_tokens,
end_token,
decode_length=max_decoding_length,
memory=memory,
memory_attention_bias=memory_attention_bias,
decoding_strategy=decoding_strategy,
)
output = TransformerDecoderOutput(
logits=logits,
sample_id=preds)
rets = output, sequence_length
else:
# The output format is different when running beam search
sample_id, log_prob = self._beam_decode(
self._prepare_tokens_to_embeds,
start_tokens,
end_token,
beam_width=beam_width,
alpha=alpha,
decode_length=max_decoding_length,
memory=memory,
memory_attention_bias=memory_attention_bias,
)
predictions = {
'sample_id':sample_id,
'log_prob': log_prob
}
rets = predictions
if not self._built:
self._add_internal_trainable_variables()
self._built = True
return rets
def _build(self,
positions=None,
sequence_length=None,
mode=None,
**kwargs):
"""Embeds the positions.
Either :attr:`position` or :attr:`sequence_length` is required:
- If both are given, :attr:`sequence_length` is used to mask out \
embeddings of those time steps beyond the respective sequence \
lengths.
- If only :attr:`sequence_length` is given, then positions \
from `0` to `sequence_length-1` are embedded.
Args:
positions (optional): An integer tensor containing the position
ids to embed.
sequence_length (optional): An integer tensor of shape
`[batch_size]`. Time steps beyond
the respective sequence lengths will have zero-valued
embeddings.
mode (optional): A tensor taking value in
:tf_main:`tf.estimator.ModeKeys <estimator/ModeKeys>`, including
`TRAIN`, `EVAL`, and `PREDICT`. If `None`, dropout will be
controlled by :func:`texar.global_mode`.
kwargs: Additional keyword arguments for
:tf_main:`tf.nn.embedding_lookup <nn/embedding_lookup>` besides
:attr:`params` and :attr:`ids`.
Returns:
A `Tensor` of shape `shape(inputs) + embedding dimension`.
"""
# Gets embedder inputs
inputs = positions
if positions is None:
if sequence_length is None:
raise ValueError(
'Either `positions` or `sequence_length` is required.')
max_length = tf.reduce_max(sequence_length)
single_inputs = tf.range(start=0, limit=max_length, dtype=tf.int32)
# Expands `single_inputs` to have shape [batch_size, max_length]
expander = tf.expand_dims(tf.ones_like(sequence_length), -1)
inputs = expander * tf.expand_dims(single_inputs, 0)
ids_rank = len(inputs.shape.dims)
embedding = self._embedding
is_training = is_train_mode(mode)
# Gets dropout strategy
st = self._hparams.dropout_strategy
if positions is None and st == 'item':
# If `inputs` is based on `sequence_length`, then dropout
# strategies 'item' and 'item_type' have the same effect, we
# use 'item_type' to avoid unknown noise_shape in the 'item'
# strategy
st = 'item_type'
# Dropouts as 'item_type' before embedding
if st == 'item_type':
dropout_layer = self._get_dropout_layer(self._hparams,
dropout_strategy=st)
if dropout_layer:
embedding = dropout_layer.apply(inputs=embedding,
training=is_training)
# Embeds
outputs = tf.nn.embedding_lookup(embedding, inputs, **kwargs)
# Dropouts as 'item' or 'elements' after embedding
if st != 'item_type':
dropout_layer = self._get_dropout_layer(self._hparams,
ids_rank=ids_rank,
dropout_input=outputs,
dropout_strategy=st)
if dropout_layer:
outputs = dropout_layer.apply(inputs=outputs,
training=is_training)
# Optionally masks
if sequence_length is not None:
outputs = mask_sequences(outputs,
sequence_length,
tensor_rank=len(inputs.shape.dims) +
self._dim_rank)
return outputs
def _build(self, inputs, sequence_length, mode=None):
"""Encodes the inputs.
Args:
inputs: A 3D Tensor of shape `[batch_size, max_time, dim]`,
containing the word embeddings of input sequences. Note that
the embedding dimension `dim` must equal "dim" in
:attr:`hparams`.
sequence_length: A 1D Tensor of shape `[batch_size]`. Input tokens
beyond respective sequence lengths are masked out
automatically.
mode (optional): A tensor taking value in
:tf_main:`tf.estimator.ModeKeys <estimator/ModeKeys>`,
including `TRAIN`, `EVAL`, and `PREDICT`. Used to toggle
dropout.
If `None` (default), :func:`texar.global_mode` is used.
Returns:
A Tensor of shape `[batch_size, max_time, dim]` containing the
encoded vectors.
"""
# Multiply input embedding with the sqrt of its dimension for
# normalization
if not self._hparams.use_bert_config:
inputs = inputs * self._hparams.dim**0.5
inputs = mask_sequences(inputs, sequence_length, tensor_rank=3)
_, lengths, _ = shape_list(inputs)
inputs_padding = 1 - tf.sequence_mask(
sequence_length, tf.shape(inputs)[1], dtype=tf.float32)
if self._hparams.use_bert_config:
ignore_padding = attn.attention_bias_ignore_padding(
inputs_padding, bias_value=-1e4)
else:
ignore_padding = attn.attention_bias_ignore_padding(
inputs_padding)
encoder_self_attention_bias = ignore_padding
positions = tf.expand_dims(tf.range(lengths, dtype=tf.int32), 0)
pos_embeds = self.position_embedder(positions)
input_embedding = inputs + pos_embeds
if self._hparams.use_bert_config:
x = layers.layer_normalize(input_embedding)
x = tf.layers.dropout(x,
rate=self._hparams.embedding_dropout,
training=is_train_mode(mode))
else:
x = tf.layers.dropout(input_embedding,
rate=self._hparams.embedding_dropout,
training=is_train_mode(mode))
# Just to keep consistent with BERT, actually makes no difference
if self._hparams.use_bert_config:
pad_remover = None
else:
pad_remover = utils.transformer_utils.PadRemover(inputs_padding)
for i in range(self._hparams.num_blocks):
with tf.variable_scope("layer_{}".format(i)):
multihead_attention = self.multihead_attention_list[i]
# trivial difference between BERT and original Transformer
if self._hparams.use_bert_config:
_queries_input = x
else:
_queries_input = layers.layer_normalize(x)
attention_output = multihead_attention(
queries=_queries_input,
memory=_queries_input,
memory_attention_bias=encoder_self_attention_bias,
mode=mode,
)
attention_output = tf.layers.dropout(
attention_output,
rate=self._hparams.residual_dropout,
training=is_train_mode(mode),
)
x = x + attention_output
with tf.variable_scope('output'):
if self._hparams.use_bert_config:
x = layers.layer_normalize(x)
y = x
else:
y = layers.layer_normalize(x)
poswise_network = self.poswise_networks[i]
with tf.variable_scope(poswise_network.variable_scope):
original_shape = shape_list(y)
y = tf.reshape(y, [-1, self._hparams.dim])
if pad_remover:
y = tf.expand_dims(pad_remover.remove(y), axis=0)
# [1, batch_size*seq_length, hidden_dim]
layer_output = poswise_network(y, mode=mode)
sub_output = tf.layers.dropout(
layer_output,
rate=self._hparams.residual_dropout,
training=is_train_mode(mode)
)
if pad_remover:
sub_output = tf.reshape(pad_remover.restore(tf.squeeze(\
sub_output, axis=0)), original_shape \
)
else:
sub_output = tf.reshape(sub_output, original_shape)
x = x + sub_output
if self._hparams.use_bert_config:
x = layers.layer_normalize(x)
if not self._hparams.use_bert_config:
x = layers.layer_normalize(x)
if not self._built:
self._add_internal_trainable_variables()
self._built = True
return x
def forward(self, # type: ignore
inputs: Optional[torch.Tensor] = None,
sequence_length: Optional[torch.LongTensor] = None,
memory: Optional[torch.Tensor] = None,
memory_sequence_length: Optional[torch.LongTensor] = None,
memory_attention_bias: Optional[torch.Tensor] = None,
context: Optional[torch.Tensor] = None,
context_sequence_length: Optional[torch.LongTensor] = None,
helper: Optional[Helper] = None,
decoding_strategy: str = 'train_greedy',
max_decoding_length: Optional[int] = None,
impute_finished: bool = False,
infer_mode: Optional[bool] = None,
beam_width: Optional[int] = None,
length_penalty: float = 0.,
**kwargs) \
-> Union[
TransformerDecoderOutput,
Tuple[TransformerDecoderOutput, torch.LongTensor],
Dict[str, torch.Tensor]]:
r"""Performs decoding.
The interface is very similar to that of RNN decoders
(:class:`texar.modules.RNNDecoderBase`). In particular,
the function provides **3 ways** to specify the decoding method, with
varying flexibility:
1. The :attr:`decoding_strategy` argument.
- **"train_greedy"**: decoding in teacher-forcing fashion (i.e.,
feeding ground truth to decode the next step), and for each step
sample is obtained by taking the `argmax` of logits.
Argument :attr:`inputs` is required for this strategy.
:attr:`sequence_length` is optional.
- **"infer_greedy"**: decoding in inference fashion (i.e., feeding
`generated` sample to decode the next step), and for each step
sample is obtained by taking the `argmax` of logits.
Arguments :attr:`(start_tokens, end_token)` are
required for this strategy, and argument
:attr:`max_decoding_length` is optional.
- **"infer_sample"**: decoding in inference fashion, and for each
step sample is obtained by `random sampling` from the logits.
Arguments :attr:`(start_tokens, end_token)` are required for this
strategy, and argument :attr:`max_decoding_length` is optional.
This argument is used only when arguments :attr:`helper` and
:attr:`beam_width` are both `None`.
2. The :attr:`helper` argument: An instance of subclass of
:class:`texar.modules.decoders.Helper`.
This provides a superset of decoding strategies than above.
The interface is the same as in RNN decoders.
Please refer to :meth:`texar.modules.RNNDecoderBase.forward` for
detailed usage and examples.
Note that, here, though using a
:class:`~texar.decoder.TrainingHelper` corresponding to the
``"train_greedy"`` strategy above, the implementation is *slower*
than directly setting ``decoding_strategy="train_greedy"`` (though
output results are the same).
Argument :attr:`max_decoding_length` is optional.
3. **Beam search**: set :attr:`beam_width` to use beam search decoding.
Arguments :attr:`(start_tokens, end_token)` are required,
and argument :attr:`max_decoding_length` is optional.
.. warning::
Beam search is not yet implemented. Setting :attr:`beam_width`
to any value greater than 1 would raise a
:exc:`NotImplementedError`
Args:
memory (optional): The memory to attend, e.g., the output of an RNN
encoder. A :tensor:`Tensor` of shape
``[batch_size, memory_max_time, dim]``.
memory_sequence_length (optional): A :tensor:`Tensor` of shape
``[batch_size]`` containing the sequence lengths for the batch
entries in memory. Used to create attention bias of
:attr:`memory_attention_bias` is not given. Ignored if
:attr:`memory_attention_bias` is provided.
memory_attention_bias (optional): A :tensor:`Tensor` of shape
``[batch_size, num_heads, memory_max_time, dim]``.
An attention bias typically sets the value of a padding
position to a large negative value for masking. If not given,
:attr:`memory_sequence_length` is used to automatically
create an attention bias.
inputs (optional): Input tensor for teacher forcing decoding, of
shape ``[batch_size, target_max_time, emb_dim]`` containing the
target sequence word embeddings. Used when
:attr:`decoding_strategy` is set to ``"train_greedy"``.
sequence_length (optional): A :tensor:`LongTensor` of shape
``[batch_size]``, containing the sequence length of
:attr:`inputs`. Tokens beyond the respective sequence length are
masked out.
Used when :attr:`decoding_strategy` is set to
``"train_greedy"``.
decoding_strategy (str): A string specifying the decoding
strategy, including ``"train_greedy"``, ``"infer_greedy"``,
``"infer_sample"``.
Different arguments are required based on the
strategy. See above for details. Ignored if
:attr:`beam_width` or :attr:`helper` is set.
beam_width (int): Set to use beam search. If given,
:attr:`decoding_strategy` is ignored.
length_penalty (float): Length penalty coefficient used in beam
search decoding. Refer to https://arxiv.org/abs/1609.08144
for more details.
It should be larger if longer sentences are desired.
context (optional): An :tensor:`LongTensor` of shape
``[batch_size, length]``, containing the starting tokens for
decoding. If context is set, ``start_tokens`` of the
:class:`~texar.modules.Helper` will be ignored.
context_sequence_length (optional): Specify the length of context.
max_decoding_length (int, optional): The maximum allowed number of
decoding steps.
If `None` (default), use ``"max_decoding_length"`` defined in
:attr:`hparams`. Ignored in ``"train_greedy"`` decoding.
impute_finished (bool): If `True`, then states for batch
entries which are marked as finished get copied through and
the corresponding outputs get zeroed out. This causes some
slowdown at each time step, but ensures that the final state
and outputs have the correct values and that backprop ignores
time steps that were marked as finished. Ignored in
``"train_greedy"`` decoding.
helper (optional): An instance of
:class:`texar.modules.decoders.Helper`
that defines the decoding strategy. If given,
``decoding_strategy`` and helper configurations in
:attr:`hparams` are ignored.
infer_mode (optional): If not `None`, overrides mode given by
:attr:`self.training`.
Returns:
- For **"train_greedy"** decoding, returns an instance of
:class:`~texar.modules.TransformerDecoderOutput` which contains
`sample_id` and `logits`.
- For **"infer_greedy"** and **"infer_sample"** decoding or
decoding with :attr:`helper`, returns
a tuple ``(outputs, sequence_lengths)``, where ``outputs`` is an
instance of :class:`~texar.modules.TransformerDecoderOutput` as
in `"train_greedy"`, and ``sequence_lengths`` is a
:tensor:`LongTensor` of shape ``[batch_size]`` containing the
length of each sample.
- For **beam search** decoding, returns a ``dict`` containing keys
``"sample_id"`` and ``"log_prob"``.
- ``"sample_id"`` is a :tensor:`LongTensor` of shape
``[batch_size, max_time, beam_width]`` containing generated
token indexes. ``sample_id[:,:,0]`` is the highest-probable
sample.
- ``"log_prob"`` is a :tensor:`Tensor` of shape
``[batch_size, beam_width]`` containing the log probability
of each sequence sample.
"""
if memory is not None:
if memory_attention_bias is None:
if memory_sequence_length is None:
raise ValueError("`memory_sequence_length` is required if "
"`memory_attention_bias` is not given.")
enc_padding = 1 - sequence_mask(memory_sequence_length,
memory.size(1),
dtype=torch.float32)
memory_attention_bias = attn.attention_bias_ignore_padding(
enc_padding)
# record the context, which will be used in step function
# for dynamic_decode
if context is not None:
if context_sequence_length is None:
raise ValueError("'context_sequence_length' must not be None"
"when 'context' is specified.")
self._state_context = context[:, 1:]
self._state_context_sequence_length = context_sequence_length - 1
else:
self._state_context = None
self._state_context_sequence_length = None
# Faster code path for teacher-forcing training
if (helper is None and beam_width is None
and decoding_strategy == 'train_greedy'):
if inputs is None:
raise ValueError(
"'input' must not be none "
"when using 'train_greedy' decoding strategy.")
if sequence_length is not None:
inputs = mask_sequences(inputs, sequence_length)
decoder_self_attention_bias = (attn.attention_bias_lower_triangle(
inputs.size(1)))
decoder_output = self._self_attention_stack(
inputs,
memory,
decoder_self_attention_bias,
memory_attention_bias,
cache=None)
logits = self._output_layer(decoder_output)
sample_id = torch.argmax(logits, dim=-1)
return TransformerDecoderOutput(logits, sample_id)
# Inference code path.
if max_decoding_length is None:
max_decoding_length = self._hparams.max_decoding_length
self._state_max_decoding_length = max_decoding_length
if beam_width is None or beam_width == 1: # Inference-like decoding
# Prepare helper
if helper is None:
kwargs.update(decoding_strategy=decoding_strategy)
if context is not None:
kwargs.update(start_tokens=context[:, 0])
helper = self._create_or_get_helper(infer_mode, **kwargs)
assert isinstance(helper, EmbeddingHelper)
self._state_cache = self._init_cache(memory,
memory_attention_bias,
beam_search_decoding=False,
batch_size=helper.batch_size)
if context is not None:
assert self._state_context is not None
pad_length = max_decoding_length - self._state_context.size(1)
if pad_length > 0:
self._state_context = torch.cat(
(self._state_context,
self._state_context.new_zeros(
self._state_context.size(0), pad_length)),
dim=1)
outputs, cache, sequence_lengths = self.dynamic_decode(
helper,
inputs=None,
sequence_length=None,
initial_state=None,
max_decoding_length=max_decoding_length,
impute_finished=impute_finished)
del cache # not used
if context is not None:
# Here the length of sample_id will be larger than that
# of logit by 1, because there will be a additional
# start_token in the returned sample_id.
# the start_id should be the first token of the
# given context
start_tokens = context[:, 0]
outputs = TransformerDecoderOutput(
logits=outputs.logits,
sample_id=torch.cat(
[start_tokens.unsqueeze(1), outputs.sample_id], dim=1))
sequence_lengths = sequence_lengths + 1
return outputs, sequence_lengths
else: # Beam-search decoding
# Ignore `decoding_strategy` and # assume `helper` is not set.
if helper is not None:
raise ValueError("Must not set 'beam_width' and 'helper' "
"simultaneously.")
if context is not None:
start_tokens = context[:, 0]
else:
if 'start_tokens' not in kwargs:
raise ValueError(
"'start_tokens' must be specified when using"
"beam search decoding.")
start_tokens = kwargs['start_tokens']
_batch_size = start_tokens.size(0)
self._state_cache = self._init_cache(memory,
memory_attention_bias,
beam_search_decoding=True,
batch_size=_batch_size)
end_token: int = kwargs.get('end_token') # type: ignore
# The output format is different when running beam search.
sample_id, log_prob = self._beam_decode(
start_tokens,
end_token,
embedding_fn=kwargs['embedding'],
beam_width=beam_width,
length_penalty=length_penalty,
decode_length=max_decoding_length)
return {'sample_id': sample_id, 'log_prob': log_prob}
def mask_and_reduce(sequence,
sequence_length,
rank=2,
average_across_batch=True,
average_across_timesteps=False,
average_across_remaining=False,
sum_over_batch=False,
sum_over_timesteps=True,
sum_over_remaining=True,
dtype=None,
time_major=False):
"""Masks out sequence entries that are beyond the respective sequence
lengths, and reduces (average or sum) away dimensions.
This is a combined function of :func:`~texar.utils.shapes.mask_sequences`
and :func:`~texar.losses.losses_utils.reduce_batch_time`.
Args:
sequence: A Tensor of sequence values.
If `time_major=False` (default), this must be a Tensor of shape:
`[batch_size, max_time, d_2, ..., d_rank]`, where the rank of
the Tensor is specified with :attr:`rank`.
If `time_major=True`, this must be a Tensor of shape:
`[max_time, batch_size, d_2, ..., d_rank].`
sequence_length: A Tensor of shape `[batch_size]`. Time steps beyond
the respective sequence lengths will be made zero. If `None`,
not masking is performed.
rank (int): The rank of :attr:`sequence`. Must be >= 2. Default is 2,
i.e., :attr:`sequence` is a 2D Tensor consisting of batch and time
dimensions.
average_across_timesteps (bool): If set, average the sequence across
the time dimension. Must not set :attr:`average_across_timesteps`
and :attr:`sum_over_timesteps` at the same time.
average_across_batch (bool): If set, average the sequence across the
batch dimension. Must not set :attr:`average_across_batch`'
and :attr:`sum_over_batch` at the same time.
average_across_remaining (bool): If set, average the sequence across the
remaining dimensions. Must not set :attr:`average_across_remaining`'
and :attr:`sum_over_remaining` at the same time.
sum_over_timesteps (bool): If set, sum the loss across the
time dimension. Must not set :attr:`average_across_timesteps`
and :attr:`sum_over_timesteps` at the same time.
sum_over_batch (bool): If set, sum the loss across the
batch dimension. Must not set :attr:`average_across_batch`
and :attr:`sum_over_batch` at the same time.
sum_over_remaining (bool): If set, sum the loss across the
remaining dimension. Must not set :attr:`average_across_remaining`
and :attr:`sum_over_remaining` at the same time.
time_major (bool): The shape format of the inputs. If `True`,
:attr:`sequence` must have shape `[max_time, batch_size, ...]`.
If `False` (default), :attr:`sequence` must have
shape `[batch_size, max_time, ...]`.
dtype (dtype): Type of :attr:`sequence`. If `None`, infer from
:attr:`sequence` automatically.
"""
if rank < 2:
raise ValueError('`rank` must be >= 2.')
if time_major:
sequence = rnn._transpose_batch_time(sequence)
if sequence_length is None:
sequence = mask_sequences(sequence,
sequence_length,
dtype=dtype,
time_major=False,
tensor_rank=rank)
if rank > 2:
if average_across_remaining and sum_over_remaining:
raise ValueError("Only one of `average_across_remaining` and "
"`sum_over_remaining` can be set.")
if average_across_remaining:
sequence = tf.reduce_mean(sequence, axis=range(2, rank))
elif sum_over_remaining:
sequence = tf.reduce_sum(sequence, axis=range(2, rank))
sequence = reduce_batch_time(sequence, sequence_length,
average_across_batch,
average_across_timesteps, sum_over_batch,
sum_over_timesteps)
reduce_time = average_across_timesteps or sum_over_timesteps
reduce_batch = average_across_batch or sum_over_batch
if not reduce_time and not reduce_batch and time_major:
sequence = rnn._transpose_batch_time(sequence)
return sequence
def forward(
self, # type: ignore
positions: Optional[torch.LongTensor] = None,
sequence_length: Optional[torch.LongTensor] = None,
**kwargs):
r"""Embeds the positions.
Either :attr:`positions` or :attr:`sequence_length` is required:
- If both are given, :attr:`sequence_length` is used to mask out
embeddings of those time steps beyond the respective sequence
lengths.
- If only :attr:`sequence_length` is given, then positions
from 0 to ``sequence_length - 1`` are embedded.
Args:
positions (optional): A :tensor:`LongTensor` containing the position
IDs to embed.
sequence_length (optional): An :tensor:`LongTensor` of shape
``[batch_size]``. Time steps beyond the respective sequence
lengths will have zero-valued embeddings.
kwargs: Additional keyword arguments for
:torch_nn:`functional.embedding` besides
:attr:`params` and :attr:`ids`.
Returns:
A `Tensor` of shape `shape(inputs) + embedding dimension`.
"""
# Gets embedder inputs
if positions is None:
if sequence_length is None:
raise ValueError(
'Either `positions` or `sequence_length` is required.')
max_length = torch.max(sequence_length)
single_inputs = torch.arange(start=0, end=max_length)
# Expands `single_inputs` to have shape [batch_size, max_length]
inputs = single_inputs.unsqueeze(0)
inputs = inputs.expand(len(sequence_length), -1).contiguous()
else:
inputs = positions
ids_rank = inputs.dim()
embedding = self._embedding
inputs = inputs.to(device=embedding.device)
# Gets dropout strategy
st = self._hparams.dropout_strategy
# Dropouts as 'item_type' before embedding
if st == 'item_type':
noise_shape = self._get_noise_shape(dropout_strategy=st,
dropout_input=embedding)
embedding = self._dropout_layer(embedding, noise_shape)
# Embeds
outputs = torch.nn.functional.embedding(inputs.type(torch.long),
embedding, **kwargs)
# Dropouts as 'item' or 'elements' after embedding
if st != 'item_type':
noise_shape = self._get_noise_shape(dropout_strategy=st,
dropout_input=outputs,
ids_rank=ids_rank)
outputs = self._dropout_layer(outputs, noise_shape)
# Optionally masks
if sequence_length is not None:
outputs = mask_sequences(outputs, sequence_length)
return outputs
def _build(self, inputs, sequence_length, mode=None):
"""Encodes the inputs.
Args:
inputs: A 3D Tensor of shape `[batch_size, max_time, dim]`,
containing the word embeddings of input sequences. Note that
the embedding dimension `dim` must equal "dim" in
:attr:`hparams`.
sequence_length: A 1D Tensor of shape `[batch_size]`. Input tokens
beyond respective sequence lengths are masked out
automatically.
mode (optional): A tensor taking value in
:tf_main:`tf.estimator.ModeKeys <estimator/ModeKeys>`,
including `TRAIN`, `EVAL`, and `PREDICT`. Used to toggle
dropout.
If `None` (default), :func:`texar.global_mode` is used.
Returns:
A Tensor of shape `[batch_size, max_time, dim]` containing the
encoded vectors.
"""
# Multiply input embedding with the sqrt of its dimension for
# normalization
inputs = inputs * self._hparams.dim**0.5
inputs = mask_sequences(inputs, sequence_length, tensor_rank=3)
_, lengths, _ = shape_list(inputs)
inputs_padding = 1 - tf.sequence_mask(
sequence_length, tf.shape(inputs)[1], dtype=tf.float32)
ignore_padding = attn.attention_bias_ignore_padding(inputs_padding)
encoder_self_attention_bias = ignore_padding
pos_embeds = self.position_embedder(lengths, self._hparams.dim)
input_embedding = inputs + pos_embeds
x = tf.layers.dropout(input_embedding,
rate=self._hparams.embedding_dropout,
training=is_train_mode(mode))
pad_remover = utils.transformer_utils.PadRemover(inputs_padding)
for i in range(self._hparams.num_blocks):
with tf.variable_scope("layer_{}".format(i)):
with tf.variable_scope('self_attention'):
selfatt_output = attn.multihead_attention(
queries=layers.layer_normalize(x),
memory=None,
memory_attention_bias=encoder_self_attention_bias,
num_heads=self._hparams.num_heads,
dropout_rate=self._hparams.attention_dropout,
num_units=self._hparams.dim,
scope='multihead_attention')
x = x + tf.layers.dropout(
selfatt_output,
rate=self._hparams.residual_dropout,
training=is_train_mode(mode),
)
poswise_network = FeedForwardNetwork(
hparams=self._hparams['poswise_feedforward'])
with tf.variable_scope(poswise_network.variable_scope):
y = layers.layer_normalize(x)
original_shape = shape_list(y)
y = tf.reshape(y, [-1, self._hparams.dim])
y = tf.expand_dims(pad_remover.remove(y), axis=0)
# [1, batch_size*seq_length, hidden_dim]
sub_output = tf.layers.dropout(
poswise_network(y),
rate=self._hparams.residual_dropout,
training=is_train_mode(mode))
sub_output = tf.reshape(pad_remover.restore(tf.squeeze(\
sub_output, axis=0)), original_shape \
)
x = x + sub_output
encoder_output = layers.layer_normalize(x)
if not self._built:
self._add_internal_trainable_variables()
self._built = True
return encoder_output
def _dynamic_rnn_loop(cell: RNNCellBase[State],
inputs: torch.Tensor,
initial_state: State,
sequence_length: torch.LongTensor) \
-> Tuple[torch.Tensor, State]:
r"""Internal implementation of Dynamic RNN.
Args:
cell: An instance of RNNCell.
inputs: A ``Tensor`` of shape ``[time, batch_size, input_size]``,
or a nested tuple of such elements.
initial_state: A ``Tensor`` of shape ``[batch_size, state_size]``,
or if ``cell.state_size`` is a tuple, then this should be a tuple
of tensors having shapes ``[batch_size, s]`` for ``s`` in
``cell.state_size``.
sequence_length: (optional) An ``int32`` ``Tensor``
of shape ``[batch_size]``.
Returns:
Tuple ``(final_outputs, final_state)``.
final_outputs:
A ``Tensor`` of shape ``[time, batch_size, cell.output_size]``. If
``cell.output_size`` is a (possibly nested) tuple of ints or
``TensorShape`` objects, then this returns a
(possibly nested) tuple of Tensors matching the corresponding
shapes.
final_state:
A ``Tensor``, or possibly nested tuple of Tensors, matching
in length and shapes to ``initial_state``.
"""
state = initial_state
time_steps = inputs.shape[0]
all_outputs = []
all_state: MaybeTuple[List[torch.Tensor]]
if isinstance(state, tuple):
all_state = ([], [])
else:
all_state = []
for i in range(time_steps):
output, state = cell(inputs[i], state)
all_outputs.append(output)
if isinstance(state, tuple):
all_state[0].append(state[0])
all_state[1].append(state[1])
else:
all_state.append(state) # type: ignore
# pylint: disable=fixme
# TODO: Do not compute everything regardless of sequence_length
final_outputs = torch.stack(all_outputs, dim=0)
final_outputs = mask_sequences(final_outputs,
sequence_length=sequence_length,
time_major=True)
final_state: MaybeTuple[List[torch.Tensor]]
if isinstance(state, tuple):
final_state = ([], [])
else:
final_state = []
for batch_idx, time_idx in enumerate(sequence_length.tolist()):
if time_idx > 0:
if isinstance(state, tuple):
final_state[0].append(all_state[0][time_idx - 1][batch_idx])
final_state[1].append(all_state[1][time_idx - 1][batch_idx])
else:
final_state.append( # type: ignore
all_state[time_idx - 1][batch_idx])
else:
if isinstance(initial_state, tuple):
final_state[0].append(initial_state[0][batch_idx])
final_state[1].append(initial_state[1][batch_idx])
else:
final_state.append(initial_state[batch_idx]) # type: ignore
if isinstance(state, tuple):
final_state = (torch.stack(final_state[0],
dim=0), torch.stack(final_state[1], dim=0))
else:
final_state = torch.stack(final_state, dim=0) # type: ignore
return final_outputs, final_state
def mask_and_reduce(sequence: torch.Tensor,
sequence_length: Optional[torch.LongTensor],
rank: int = 2,
average_across_batch: bool = True,
average_across_timesteps: bool = False,
average_across_remaining: bool = False,
sum_over_batch: bool = False,
sum_over_timesteps: bool = True,
sum_over_remaining: bool = True,
dtype: Optional[torch.dtype] = None,
time_major: bool = False) -> torch.Tensor:
r"""Masks out sequence entries that are beyond the respective sequence
lengths, and reduces (average or sum) away dimensions.
This is a combination of :func:`~texar.utils.shapes.mask_sequences`
and :func:`~texar.losses.losses_utils.reduce_batch_time`.
Args:
sequence: A tensor of sequence values.
If `time_major=False` (default), this must be a tensor of shape
`[batch_size, max_time, d_2, ..., d_rank]`, where the rank of
the tensor is specified with :attr:`rank`.
The batch and time dimensions are exchanged if `time_major` is True.
sequence_length: A tensor of shape `[batch_size]`. Time steps beyond
the respective sequence lengths will be made zero. If `None`,
no masking is performed.
rank (int): The rank of :attr:`sequence`. Must be >= 2. Default is 2,
i.e., `sequence` is a 2D Tensor consisting of batch and time
dimensions.
average_across_timesteps (bool): If set, average the sequence across
the time dimension. Must not set `average_across_timesteps`
and `sum_over_timesteps` at the same time.
average_across_batch (bool): If set, average the sequence across the
batch dimension. Must not set `average_across_batch`'
and `sum_over_batch` at the same time.
average_across_remaining (bool): If set, average the sequence across the
remaining dimensions. Must not set `average_across_remaining`'
and `sum_over_remaining` at the same time.
sum_over_timesteps (bool): If set, sum the sequence across the time
dimension. Must not set `average_across_timesteps` and
`sum_over_timesteps` at the same time.
sum_over_batch (bool): If set, sum the sequence across the batch
dimension. Must not set `average_across_batch` and `sum_over_batch`
at the same time.
sum_over_remaining (bool): If set, sum the sequence across the remaining
dimension. Must not set `average_across_remaining` and
`sum_over_remaining` at the same time.
dtype (torch.dtype): The dtype of the returned mask.
time_major (bool): The shape format of the inputs. If `True`,
:attr:`sequence` must have shape `[max_time, batch_size, ...]`.
If `False` (default), `sequence` must have
shape `[batch_size, max_time, ...]`.
Returns:
A tensor containing the masked and reduced sequence.
"""
if rank < 2:
raise ValueError('`rank` must be >= 2.')
if time_major:
sequence = transpose_batch_time(sequence)
if sequence_length is not None:
sequence = mask_sequences(sequence,
sequence_length,
dtype=dtype,
time_major=False)
if rank > 2:
if average_across_remaining and sum_over_remaining:
raise ValueError("Only one of `average_across_remaining` and "
"`sum_over_remaining` can be set.")
if average_across_remaining:
for axis in sorted(list(range(2, rank)), reverse=True):
sequence = torch.mean(sequence, dim=axis)
elif sum_over_remaining:
for axis in sorted(list(range(2, rank)), reverse=True):
sequence = torch.sum(sequence, dim=axis)
sequence = reduce_batch_time(sequence,
sequence_length,
average_across_batch,
average_across_timesteps,
sum_over_batch,
sum_over_timesteps)
reduce_time = average_across_timesteps or sum_over_timesteps
reduce_batch = average_across_batch or sum_over_batch
if not reduce_time and not reduce_batch and time_major:
sequence = transpose_batch_time(sequence)
return sequence
def _dynamic_rnn_loop(cell: RNNCellBase[State],
inputs: torch.Tensor,
initial_state: State,
sequence_length: torch.LongTensor) \
-> Tuple[torch.Tensor, State]:
r"""Internal implementation of Dynamic RNN.
Args:
cell: An instance of RNNCell.
inputs: A ``Tensor`` of shape ``[time, batch_size, input_size]``,
or a nested tuple of such elements.
initial_state: A ``Tensor`` of shape ``[batch_size, state_size]``,
or if ``cell.state_size`` is a tuple, then this should be a tuple
of tensors having shapes ``[batch_size, s]`` for ``s`` in
``cell.state_size``.
sequence_length: (optional) An ``int32`` ``Tensor``
of shape ``[batch_size]``.
Returns:
Tuple ``(final_outputs, final_state)``.
final_outputs:
A ``Tensor`` of shape ``[time, batch_size, cell.output_size]``. If
``cell.output_size`` is a (possibly nested) tuple of ints or
``torch.Size`` objects, then this returns a
(possibly nested) tuple of Tensors matching the corresponding
shapes.
final_state:
A ``Tensor``, or possibly nested tuple of Tensors, matching
in length and shapes to ``initial_state``.
"""
state = initial_state
time_steps = inputs.shape[0]
all_outputs = []
all_state = map_structure(lambda _: no_map(list), state)
for i in range(time_steps):
output, state = cell(inputs[i], state)
all_outputs.append(output)
map_structure_zip(lambda xs, x: xs.append(x), (all_state, state))
# TODO: Do not compute everything regardless of sequence_length
final_outputs = torch.stack(all_outputs, dim=0)
final_outputs = mask_sequences(final_outputs,
sequence_length=sequence_length,
time_major=True)
final_state = map_structure(lambda _: no_map(list), state)
# pylint: disable=cell-var-from-loop
# Our use case is fine because the function is called immediately and
# exclusively in the current iteration of the loop.
for batch_idx, time_idx in enumerate(sequence_length.tolist()):
if time_idx > 0:
map_structure_zip(
lambda xs, x: xs.append(x[time_idx - 1][batch_idx]),
(final_state, all_state))
else:
map_structure_zip(lambda xs, x: xs.append(x[batch_idx]),
(final_state, initial_state))
# pylint: enable=cell-var-from-loop
final_state = map_structure(lambda x: torch.stack(x, dim=0), final_state)
return final_outputs, final_state
def _build(
self, # pylint: disable=arguments-differ
memory=None,
memory_sequence_length=None,
memory_attention_bias=None,
inputs=None,
sequence_length=None,
decoding_strategy='train_greedy',
beam_width=None,
length_penalty=0.,
start_tokens=None,
end_token=None,
context=None,
context_sequence_length=None,
softmax_temperature=None,
max_decoding_length=None,
impute_finished=False,
helper=None,
mode=None):
"""Performs decoding.
The interface is very similar to that of RNN decoders
(:meth:`texar.modules.RNNDecoderBase._build`). In particular,
the function provides **3 ways** to specify the decoding method, with
varying flexibility:
1. The :attr:`decoding_strategy` argument.
- **"train_greedy"**: decoding in teacher-forcing fashion (i.e.,
feeding ground truth to decode the next step), and for each step
sample is obtained by taking the `argmax` of logits.
Argument :attr:`inputs` is required for this strategy.
:attr:`sequence_length` is optional.
- **"infer_greedy"**: decoding in inference fashion (i.e., feeding
`generated` sample to decode the next step), and for each step
sample is obtained by taking the `argmax` of logits.
Arguments :attr:`(start_tokens, end_token)` are
required for this strategy, and argument
:attr:`max_decoding_length` is optional.
- **"infer_sample"**: decoding in inference fashion, and for each
step sample is obtained by `random sampling` from the logits.
Arguments :attr:`(start_tokens, end_token)` are required for this
strategy, and argument :attr:`max_decoding_length` is optional.
This argument is used only when arguments :attr:`helper` and
:attr:`beam_width` are both `None`.
2. The :attr:`helper` argument: An instance of subclass of
:tf_main:`tf.contrib.seq2seq.Helper <contrib/seq2seq/Helper>`.
This provides a superset of decoding strategies than above.
The interface is the same as in RNN decoders.
Please refer to :meth:`texar.modules.RNNDecoderBase._build` for
detailed usage and examples.
Note that, here, though using a :tf_main:`TrainingHelper
<contrib/seq2seq/TrainingHelper>` corresponding to the
"train_greedy" strategy above, the implementation is *slower* than
directly setting `decoding_strategy="train_greedy"` (though the
output results are the same).
Argument :attr:`max_decoding_length` is optional.
3. **Beam search**: set :attr:`beam_width` to use beam search decoding.
Arguments :attr:`(start_tokens, end_token)` are required,
and argument :attr:`max_decoding_length` is optional.
Args:
memory (optional): The memory to attend, e.g., the output of an RNN encoder.
A Tensor of shape `[batch_size, memory_max_time, dim]`.
memory_sequence_length (optional): A Tensor of shape `[batch_size]`
containing the sequence lengths for the batch entries in
memory. Used to create attention bias of
:attr:`memory_attention_bias` is not given. Ignored if
`memory_attention_bias` is provided.
memory_attention_bias (optional): A Tensor of shape
`[batch_size, num_heads, memory_max_time, dim]`.
An attention bias typically sets the value of a padding
position to a large negative value for masking. If not given,
:attr:`memory_sequence_length` is used to automatically
create an attention bias.
inputs (optional): Input tensor for teacher forcing decoding, of
shape `[batch_size, target_max_time, emb_dim]` containing the
target sequence word embeddings.
Used when :attr:`decoding_strategy` is set to "train_greedy".
sequence_length (optional): A Tensor of shape `[batch_size]`,
containing the sequence length of :attr:`inputs`.
Tokens beyond the respective sequence length are masked out.
Used when :attr:`decoding_strategy` is set to
"train_greedy".
decoding_strategy (str): A string specifying the decoding
strategy, including "train_greedy", "infer_greedy",
"infer_sample".
Different arguments are required based on the
strategy. See above for details. Ignored if
:attr:`beam_width` or :attr:`helper` is set.
beam_width (int): Set to use beam search. If given,
:attr:`decoding_strategy` is ignored.
length_penalty (float): Length penalty coefficient used in beam search
decoding. Refer to https://arxiv.org/abs/1609.08144
for more details.
It Should be larger if longer sentences are wanted.
start_tokens (optional): An int Tensor of shape `[batch_size]`,
containing the start tokens.
Used when :attr:`decoding_strategy` = "infer_greedy" or
"infer_sample", or :attr:`beam_width` is set.
Ignored when context is set.
end_token (optional): An int 0D Tensor, the token that marks end
of decoding.
Used when :attr:`decoding_strategy` = "infer_greedy" or
"infer_sample", or :attr:`beam_width` is set.
context (optional): An int Tensor of shape `[batch_size, length]`,
containing the starting tokens for decoding.
If context is set, the start_tokens will be ignored.
context_sequence_length (optional): specify the length of context.
softmax_temperature (optional): A float 0D Tensor, value to divide
the logits by before computing the softmax. Larger values
(above 1.0) result in more random samples. Must > 0. If `None`,
1.0 is used.
Used when :attr:`decoding_strategy` = "infer_sample"`.
max_decoding_length (optional): An int scalar Tensor indicating
the maximum allowed number of decoding steps.
If `None` (default), use "max_decoding_length" defined in
:attr:`hparams`. Ignored in "train_greedy" decoding.
impute_finished (bool): If `True`, then states for batch
entries which are marked as finished get copied through and
the corresponding outputs get zeroed out. This causes some
slowdown at each time step, but ensures that the final state
and outputs have the correct values and that backprop ignores
time steps that were marked as finished. Ignored in
"train_greedy" decoding.
helper (optional): An instance of
:tf_main:`Helper <contrib/seq2seq/Helper>` that defines the
decoding strategy. If given, :attr:`decoding_strategy` is
ignored.
mode (optional): A tensor taking value in
:tf_main:`tf.estimator.ModeKeys <estimator/ModeKeys>`, including
`TRAIN`, `EVAL`, and `PREDICT`. Controls dropout mode.
If `None` (default), :func:`texar.global_mode`
is used.
Returns:
- For **"train_greedy"** decoding, returns an instance of \
:class:`~texar.modules.TransformerDecoderOutput` which contains\
`sample_id` and `logits`.
- For **"infer_greedy"** and **"infer_sample"** decoding or\
decoding with :attr:`helper`, returns\
a tuple `(outputs, sequence_lengths)`, where `outputs` is an \
instance of :class:`~texar.modules.TransformerDecoderOutput` as\
in "train_greedy", and `sequence_lengths` is a Tensor of shape\
`[batch_size]` containing the length of each sample.
- For **beam search** decoding, returns a `dict` containing keys\
"sample_id" and "log_prob".
- **"sample_id"** is an int Tensor of shape \
`[batch_size, max_time, beam_width]` containing generated\
token indexes. `sample_id[:,:,0]` is the highest-probable \
sample.
- **"log_prob"** is a float Tensor of shape \
`[batch_size, beam_width]` containing the log probability \
of each sequence sample.
"""
if memory is not None:
if memory_attention_bias is None:
if memory_sequence_length is None:
raise ValueError("`memory_sequence_length` is required if "
"`memory_attention_bias` is not given.")
enc_padding = 1 - tf.sequence_mask(memory_sequence_length,
tf.shape(memory)[1],
dtype=tf.float32)
memory_attention_bias = attn.attention_bias_ignore_padding(
enc_padding)
# record the context, which will be used in step function
# for dynamic_decode
if context is not None:
start_tokens = context[:, 0]
self.context = context[:, 1:]
self.context_sequence_length = context_sequence_length - 1
else:
self.context = None
if helper is None and beam_width is None and \
decoding_strategy == 'train_greedy': # Teacher-forcing
if sequence_length is not None:
inputs = mask_sequences(inputs, sequence_length, tensor_rank=3)
decoder_self_attention_bias = (attn.attention_bias_lower_triangle(
shape_list(inputs)[1]))
if self._hparams.scale_embeds:
target_inputs = inputs * self._hparams.dim**0.5
else:
target_inputs = inputs
_, lengths, _ = shape_list(target_inputs)
positions = tf.expand_dims(tf.range(lengths, dtype=tf.int32), 0)
pos_embeds = self.position_embedder(positions)
inputs = target_inputs + pos_embeds
decoder_output = self._self_attention_stack(
inputs,
memory,
decoder_self_attention_bias=decoder_self_attention_bias,
memory_attention_bias=memory_attention_bias,
cache=None,
mode=mode)
logits = self.output_layer(decoder_output)
preds = tf.to_int32(tf.argmax(logits, axis=-1))
rets = TransformerDecoderOutput(logits=logits, sample_id=preds)
else:
if max_decoding_length is None:
max_decoding_length = self._hparams.max_decoding_length
self._inputs_to_outputs = self._inputs_to_outputs_fn(
max_decoding_length + 1)
if beam_width is None: #Inference-like decoding
# Prepare helper
if helper is not None:
# ignore `decoding_strategy`
pass
else:
if decoding_strategy == "infer_greedy":
helper = tf.contrib.seq2seq.GreedyEmbeddingHelper(
self._embedding, start_tokens, end_token)
elif decoding_strategy == "infer_sample":
helper = tf.contrib.seq2seq.SampleEmbeddingHelper(
self._embedding, start_tokens, end_token,
softmax_temperature)
else:
raise ValueError(
"Unknown decoding strategy: {}".format(
decoding_strategy))
self._helper = helper
self._cache = self._init_cache(memory,
memory_attention_bias,
beam_search_decoding=False)
if context is not None:
self.context = tf.pad(
self.context,
[[0, 0],
[0, max_decoding_length - tf.shape(self.context)[1]]])
outputs, cache, sequence_lengths = dynamic_decode(
decoder=self,
impute_finished=impute_finished,
maximum_iterations=max_decoding_length,
output_time_major=False,
scope=self.variable_scope)
if context is not None:
# Here the length of sample_id will be larger than that
# of logit by 1, because there will be a additional
# start_token in the returned sample_id.
# the start_id should be the first token of the
# given context
outputs = TransformerDecoderOutput(
logits=outputs.logits,
sample_id=tf.concat([
tf.expand_dims(start_tokens, 1), outputs.sample_id
],
axis=1))
sequence_lengths = sequence_lengths + 1
rets = outputs, sequence_lengths
else: #Beam-search decoding
# ignore `decoding_strategy`
# assume `helper` is not set
if helper is not None:
raise ValueError("Must not set 'beam_width' and 'helper' "
"simultaneously.")
_batch_size = tf.shape(start_tokens)[0]
self._cache = self._init_cache(memory,
memory_attention_bias,
beam_search_decoding=True,
batch_size=_batch_size)
# The output format is different when running beam search
sample_id, log_prob = self._beam_decode(
start_tokens,
end_token,
beam_width=beam_width,
length_penalty=length_penalty,
decode_length=max_decoding_length,
)
rets = {'sample_id': sample_id, 'log_prob': log_prob}
if not self._built:
self._add_internal_trainable_variables()
self._built = True
return rets