def __init__(self,
source_inputter,
target_inputter,
num_layers,
num_units,
num_heads,
ffn_inner_dim,
dropout=0.1,
attention_dropout=0.1,
relu_dropout=0.1,
position_encoder=SinusoidalPositionEncoder(),
decoder_self_attention_type="scaled_dot",
name="transformer"):
"""Initializes a Transformer model.
Args:
source_inputter: A :class:`opennmt.inputters.inputter.Inputter` to process
the source data.
target_inputter: A :class:`opennmt.inputters.inputter.Inputter` to process
the target data. Currently, only the
:class:`opennmt.inputters.text_inputter.WordEmbedder` is supported.
num_layers: The shared number of layers.
num_units: The number of hidden units.
num_heads: The number of heads in each self-attention layers.
ffn_inner_dim: The inner dimension of the feed forward layers.
dropout: The probability to drop units in each layer output.
attention_dropout: The probability to drop units from the attention.
relu_dropout: The probability to drop units from the ReLU activation in
the feed forward layer.
position_encoder: A :class:`opennmt.layers.position.PositionEncoder` to
apply on the inputs.
decoder_self_attention_type: Type of self attention in the decoder,
"scaled_dot" or "average" (case insensitive).
name: The name of this model.
"""
encoder = SelfAttentionEncoder(num_layers,
num_units=num_units,
num_heads=num_heads,
ffn_inner_dim=ffn_inner_dim,
dropout=dropout,
attention_dropout=attention_dropout,
relu_dropout=relu_dropout,
position_encoder=position_encoder)
decoder = SelfAttentionDecoder(
num_layers,
num_units=num_units,
num_heads=num_heads,
ffn_inner_dim=ffn_inner_dim,
dropout=dropout,
attention_dropout=attention_dropout,
relu_dropout=relu_dropout,
position_encoder=position_encoder,
self_attention_type=decoder_self_attention_type)
super(Transformer, self).__init__(source_inputter,
target_inputter,
encoder,
decoder,
daisy_chain_variables=True,
name=name)
def __init__(self,
source_inputter,
target_inputter,
num_layers,
num_heads,
ffn_inner_dim,
dropout=0.1,
attention_dropout=0.0,
position_encoder=PositionEmbedder(),
name="transformer"):
"""Initializes a Transformer model.
Args:
source_inputter: A :class:`opennmt.inputters.inputter.Inputter` to process
the source data.
target_inputter: A :class:`opennmt.inputters.inputter.Inputter` to process
the target data. Currently, only the
:class:`opennmt.inputters.text_inputter.WordEmbedder` is supported.
num_layers: The shared number of layers.
num_heads: The number of heads in each self-attention layers.
ffn_inner_dim: The inner dimension of the feed forward layers.
dropout: The probability to drop units in each layer output.
attention_dropout: The probability to drop units from the attention.
position_encoder: A :class:`opennmt.utils.position.PositionEncoder` to
apply on the inputs.
name: The name of this model.
"""
encoder = SelfAttentionEncoder(num_layers,
num_heads=num_heads,
ffn_inner_dim=ffn_inner_dim,
dropout=dropout,
attention_dropout=attention_dropout,
keep_layers_output=True,
position_encoder=position_encoder)
decoder = SelfAttentionDecoder(num_layers,
num_heads=num_heads,
ffn_inner_dim=ffn_inner_dim,
dropout=dropout,
attention_dropout=attention_dropout,
position_encoder=position_encoder)
super(Transformer, self).__init__(source_inputter,
target_inputter,
encoder,
decoder,
name=name)
def __init__(self,
source_inputter,
target_inputter,
num_layers,
num_units,
num_heads,
ffn_inner_dim,
dropout=0.1,
attention_dropout=0.1,
relu_dropout=0.1,
position_encoder=SinusoidalPositionEncoder(),
decoder_self_attention_type="scaled_dot",
share_embeddings=EmbeddingsSharingLevel.NONE,
share_encoders=False,
alignment_file_key="train_alignments",
name="transformer"):
"""Initializes a Transformer model.
Args:
source_inputter: A :class:`opennmt.inputters.inputter.Inputter` to process
the source data. If this inputter returns parallel inputs, a multi
source Transformer architecture will be constructed.
target_inputter: A :class:`opennmt.inputters.inputter.Inputter` to process
the target data. Currently, only the
:class:`opennmt.inputters.text_inputter.WordEmbedder` is supported.
num_layers: The shared number of layers.
num_units: The number of hidden units.
num_heads: The number of heads in each self-attention layers.
ffn_inner_dim: The inner dimension of the feed forward layers.
dropout: The probability to drop units in each layer output.
attention_dropout: The probability to drop units from the attention.
relu_dropout: The probability to drop units from the ReLU activation in
the feed forward layer.
position_encoder: A :class:`opennmt.layers.position.PositionEncoder` to
apply on the inputs.
decoder_self_attention_type: Type of self attention in the decoder,
"scaled_dot" or "average" (case insensitive).
share_embeddings: Level of embeddings sharing, see
:class:`opennmt.models.sequence_to_sequence.EmbeddingsSharingLevel`
for possible values.
share_encoders: In case of multi source architecture, whether to share the
separate encoders parameters or not.
alignment_file_key: The data configuration key of the training alignment
file to support guided alignment.
name: The name of this model.
"""
encoders = [
SelfAttentionEncoder(num_layers,
num_units=num_units,
num_heads=num_heads,
ffn_inner_dim=ffn_inner_dim,
dropout=dropout,
attention_dropout=attention_dropout,
relu_dropout=relu_dropout,
position_encoder=position_encoder)
for _ in range(source_inputter.num_outputs)
]
if len(encoders) > 1:
encoder = ParallelEncoder(encoders,
outputs_reducer=None,
states_reducer=None,
share_parameters=share_encoders)
else:
encoder = encoders[0]
decoder = SelfAttentionDecoder(
num_layers,
num_units=num_units,
num_heads=num_heads,
ffn_inner_dim=ffn_inner_dim,
dropout=dropout,
attention_dropout=attention_dropout,
relu_dropout=relu_dropout,
position_encoder=position_encoder,
self_attention_type=decoder_self_attention_type)
self._num_units = num_units
super(Transformer,
self).__init__(source_inputter,
target_inputter,
encoder,
decoder,
share_embeddings=share_embeddings,
alignment_file_key=alignment_file_key,
daisy_chain_variables=True,
name=name)
def __init__(self,
source_inputter,
target_inputter,
num_layers,
num_units,
num_heads,
ffn_inner_dim,
dropout=0.1,
attention_dropout=0.1,
ffn_dropout=0.1,
ffn_activation=tf.nn.relu,
position_encoder_class=SinusoidalPositionEncoder,
share_embeddings=EmbeddingsSharingLevel.NONE,
share_encoders=False,
maximum_relative_position=None,
attention_span=None,
num_attended_heads=1):
"""Initializes a Transformer model.
Args:
source_inputter: A :class:`opennmt.inputters.Inputter` to process
the source data. If this inputter returns parallel inputs, a multi
source Transformer architecture will be constructed.
target_inputter: A :class:`opennmt.inputters.Inputter` to process
the target data. Currently, only the
:class:`opennmt.inputters.WordEmbedder` is supported.
num_layers: The shared number of layers.
num_units: The number of hidden units.
num_heads: The number of heads in each self-attention layers.
ffn_inner_dim: The inner dimension of the feed forward layers.
dropout: The probability to drop units in each layer output.
attention_dropout: The probability to drop units from the attention.
ffn_dropout: The probability to drop units from the ReLU activation in
the feed forward layer.
ffn_activation: The activation function to apply between the two linear
transformations of the feed forward layer.
position_encoder_class: The :class:`opennmt.layers.PositionEncoder`
class to use for position encoding (or a callable that returns an
instance).
share_embeddings: Level of embeddings sharing, see
:class:`opennmt.models.EmbeddingsSharingLevel` for possible values.
share_encoders: In case of multi source architecture, whether to share the
separate encoders parameters or not.
maximum_relative_position: Maximum relative position representation
(from https://arxiv.org/abs/1803.02155).
attention_span: Maximum relative position to attend to
(from https://arxiv.org/abs/1904.03107).
num_attended_heads: How many heads should be attended. Defaults to 1
as each head only attends to itself in vanilla Transformer. Increase to
an odd number < `num_heads` to also model head interaction.
(from ttps://arxiv.org/abs/1904.03107).
"""
encoders = [
SelfAttentionEncoder(
num_layers,
num_units=num_units,
num_heads=num_heads,
ffn_inner_dim=ffn_inner_dim,
dropout=dropout,
attention_dropout=attention_dropout,
ffn_dropout=ffn_dropout,
ffn_activation=ffn_activation,
position_encoder_class=position_encoder_class,
maximum_relative_position=maximum_relative_position,
attention_span=attention_span,
num_attended_heads=num_attended_heads)
for _ in range(source_inputter.num_outputs)
]
if len(encoders) > 1:
encoder = ParallelEncoder(
encoders if not share_encoders else encoders[0],
outputs_reducer=None,
states_reducer=None)
else:
encoder = encoders[0]
decoder = SelfAttentionDecoder(
num_layers,
num_units=num_units,
num_heads=num_heads,
ffn_inner_dim=ffn_inner_dim,
dropout=dropout,
attention_dropout=attention_dropout,
ffn_dropout=ffn_dropout,
ffn_activation=ffn_activation,
position_encoder_class=position_encoder_class,
num_sources=source_inputter.num_outputs,
maximum_relative_position=maximum_relative_position)
self._num_units = num_units
super(Transformer, self).__init__(source_inputter,
target_inputter,
encoder,
decoder,
share_embeddings=share_embeddings)
memory, _, _ = encoder.encode(source_embedding,
memory_sequence_length,
mode=mode)
tf_encoder_result = memory
tf_encoder_result = tf.reshape(tf_encoder_result,
[batch_size, -1, encoder_hidden_dim])
with tf.variable_scope("transformer/decoder", reuse=tf.AUTO_REUSE):
target_inputter.build()
with tf.variable_scope("transformer/decoder", reuse=tf.AUTO_REUSE):
decoder = SelfAttentionDecoder(num_layers=6,
num_units=512,
num_heads=8,
ffn_inner_dim=2048,
dropout=0.0,
attention_dropout=0.0,
relu_dropout=0.0)
start_tokens = tf.fill([batch_size], start_of_sentence_id)
end_token = end_of_sentence_id
target_ids, _, target_length, _ = decoder.dynamic_decode_and_search(
target_inputter.embedding,
start_tokens,
end_token,
vocab_size=vocab_size,
beam_width=beam_width,
memory=memory,
memory_sequence_length=memory_sequence_length)
def __init__(self,
source_inputter,
target_inputter,
num_layers,
num_units,
num_heads,
ffn_inner_dim,
dropout=0.1,
attention_dropout=0.1,
ffn_dropout=0.1,
ffn_activation=tf.nn.relu,
position_encoder_class=SinusoidalPositionEncoder,
share_embeddings=EmbeddingsSharingLevel.NONE,
share_encoders=False,
maximum_relative_position=None):
"""Initializes a Transformer model.
Args:
source_inputter: A :class:`opennmt.inputters.Inputter` to process
the source data. If this inputter returns parallel inputs, a multi
source Transformer architecture will be constructed.
target_inputter: A :class:`opennmt.inputters.Inputter` to process
the target data. Currently, only the
:class:`opennmt.inputters.WordEmbedder` is supported.
num_layers: The number of layers or a 2-tuple with the number of encoder
layers and decoder layers.
num_units: The number of hidden units.
num_heads: The number of heads in each self-attention layers.
ffn_inner_dim: The inner dimension of the feed forward layers.
dropout: The probability to drop units in each layer output.
attention_dropout: The probability to drop units from the attention.
ffn_dropout: The probability to drop units from the ReLU activation in
the feed forward layer.
ffn_activation: The activation function to apply between the two linear
transformations of the feed forward layer.
position_encoder_class: The :class:`opennmt.layers.PositionEncoder`
class to use for position encoding (or a callable that returns an
instance).
share_embeddings: Level of embeddings sharing, see
:class:`opennmt.models.EmbeddingsSharingLevel` for possible values.
share_encoders: In case of multi source architecture, whether to share the
separate encoders parameters or not.
maximum_relative_position: Maximum relative position representation
(from https://arxiv.org/abs/1803.02155).
"""
if isinstance(num_layers, (list, tuple)):
num_encoder_layers, num_decoder_layers = num_layers
else:
num_encoder_layers, num_decoder_layers = num_layers, num_layers
encoders = [
SelfAttentionEncoder(
num_encoder_layers,
num_units=num_units,
num_heads=num_heads,
ffn_inner_dim=ffn_inner_dim,
dropout=dropout,
attention_dropout=attention_dropout,
ffn_dropout=ffn_dropout,
ffn_activation=ffn_activation,
position_encoder_class=position_encoder_class,
maximum_relative_position=maximum_relative_position)
for _ in range(source_inputter.num_outputs)
]
if len(encoders) > 1:
encoder = ParallelEncoder(
encoders if not share_encoders else encoders[0],
outputs_reducer=None,
states_reducer=None)
else:
encoder = encoders[0]
decoder = SelfAttentionDecoder(
num_decoder_layers,
num_units=num_units,
num_heads=num_heads,
ffn_inner_dim=ffn_inner_dim,
dropout=dropout,
attention_dropout=attention_dropout,
ffn_dropout=ffn_dropout,
ffn_activation=ffn_activation,
position_encoder_class=position_encoder_class,
num_sources=source_inputter.num_outputs,
maximum_relative_position=maximum_relative_position)
self._num_units = num_units
self._num_encoder_layers = num_encoder_layers
self._num_decoder_layers = num_decoder_layers
self._num_heads = num_heads
self._with_relative_position = maximum_relative_position is not None
self._is_ct2_compatible = (
isinstance(encoder, SelfAttentionEncoder)
and num_encoder_layers == num_decoder_layers
and ffn_activation == tf.nn.relu and
((self._with_relative_position and position_encoder_class is None)
or (not self._with_relative_position
and position_encoder_class == SinusoidalPositionEncoder)))
super(Transformer, self).__init__(source_inputter,
target_inputter,
encoder,
decoder,
share_embeddings=share_embeddings)
def __init__(
self,
source_inputter,
target_inputter,
num_layers,
num_units,
num_heads,
ffn_inner_dim,
dropout=0.1,
attention_dropout=0.1,
ffn_dropout=0.1,
ffn_activation=tf.nn.relu,
position_encoder_class=SinusoidalPositionEncoder,
share_embeddings=EmbeddingsSharingLevel.NONE,
share_encoders=False,
maximum_relative_position=None,
attention_reduction=MultiHeadAttentionReduction.FIRST_HEAD_LAST_LAYER,
pre_norm=True,
):
"""Initializes a Transformer model.
Args:
source_inputter: A :class:`opennmt.inputters.Inputter` to process
the source data. If this inputter returns parallel inputs, a multi
source Transformer architecture will be constructed.
target_inputter: A :class:`opennmt.inputters.Inputter` to process
the target data. Currently, only the
:class:`opennmt.inputters.WordEmbedder` is supported.
num_layers: The number of layers or a 2-tuple with the number of encoder
layers and decoder layers.
num_units: The number of hidden units.
num_heads: The number of heads in each self-attention layers.
ffn_inner_dim: The inner dimension of the feed forward layers.
dropout: The probability to drop units in each layer output.
attention_dropout: The probability to drop units from the attention.
ffn_dropout: The probability to drop units from the ReLU activation in
the feed forward layer.
ffn_activation: The activation function to apply between the two linear
transformations of the feed forward layer.
position_encoder_class: The :class:`opennmt.layers.PositionEncoder`
class to use for position encoding (or a callable that returns an
instance).
share_embeddings: Level of embeddings sharing, see
:class:`opennmt.models.EmbeddingsSharingLevel` for possible values.
share_encoders: In case of multi source architecture, whether to share the
separate encoders parameters or not.
maximum_relative_position: Maximum relative position representation
(from https://arxiv.org/abs/1803.02155).
attention_reduction: A :class:`opennmt.layers.MultiHeadAttentionReduction`
value to specify how to reduce target-source multi-head attention
matrices.
pre_norm: If ``True``, layer normalization is applied before each
sub-layer. Otherwise it is applied after. The original paper uses
``pre_norm=False``, but the authors later suggested that ``pre_norm=True``
"seems better for harder-to-learn models, so it should probably be the
default."
"""
if isinstance(num_layers, (list, tuple)):
num_encoder_layers, num_decoder_layers = num_layers
else:
num_encoder_layers, num_decoder_layers = num_layers, num_layers
encoders = [
SelfAttentionEncoder(
num_encoder_layers,
num_units=num_units,
num_heads=num_heads,
ffn_inner_dim=ffn_inner_dim,
dropout=dropout,
attention_dropout=attention_dropout,
ffn_dropout=ffn_dropout,
ffn_activation=ffn_activation,
position_encoder_class=position_encoder_class,
maximum_relative_position=maximum_relative_position,
pre_norm=pre_norm,
) for _ in range(source_inputter.num_outputs)
]
if len(encoders) > 1:
encoder = ParallelEncoder(
encoders if not share_encoders else encoders[0],
outputs_reducer=None,
states_reducer=None,
)
else:
encoder = encoders[0]
decoder = SelfAttentionDecoder(
num_decoder_layers,
num_units=num_units,
num_heads=num_heads,
ffn_inner_dim=ffn_inner_dim,
dropout=dropout,
attention_dropout=attention_dropout,
ffn_dropout=ffn_dropout,
ffn_activation=ffn_activation,
position_encoder_class=position_encoder_class,
num_sources=source_inputter.num_outputs,
maximum_relative_position=maximum_relative_position,
attention_reduction=attention_reduction,
pre_norm=pre_norm,
)
self._num_units = num_units
self._num_encoder_layers = num_encoder_layers
self._num_decoder_layers = num_decoder_layers
self._num_heads = num_heads
self._with_relative_position = maximum_relative_position is not None
self._is_ct2_compatible = (
isinstance(encoder, SelfAttentionEncoder) and pre_norm
and ffn_activation is tf.nn.relu and
((self._with_relative_position and position_encoder_class is None)
or (not self._with_relative_position
and position_encoder_class == SinusoidalPositionEncoder)))
super().__init__(
source_inputter,
target_inputter,
encoder,
decoder,
share_embeddings=share_embeddings,
)
