def __init__(self,
hidden_size: int = 512,
emb_size: int = 512,
kernel_size: int = 5,
num_layers: int = 15,
dropout: float = 0.1,
emb_dropout: float = 0.1,
freeze: bool = False,
**kwargs):
"""
Initializes the ConvSeq2Seq Encoder.
:param hidden_size: hidden size and size of embeddings
:param ff_size: position-wise feed-forward layer size.
(Typically this is 2*hidden_size.)
:param num_layers: number of layers
:param dropout: dropout probability for Transformer layers
:param emb_dropout: Is applied to the input (word embeddings).
:param freeze: freeze the parameters of the encoder during training
:param kwargs:
"""
super(ConvSeq2SeqEncoder, self).__init__()
# build all (num_layers) layers
self.layers = nn.ModuleList([
ConvSeq2SeqEncoderLayer(hidden_size=hidden_size,
kernel_size=kernel_size,
dropout=dropout) for _ in range(num_layers)
])
self.absPE = AbsolutePositionalEncoding(emb_size)
self.emb2hidden = nn.Linear(emb_size, hidden_size)
self.emb_dropout = nn.Dropout(p=emb_dropout)
if freeze:
freeze_params(self)
def __init__(self,
embedding_dim: int = 64,
scale: bool = False,
vocab_size: int = 0,
padding_idx: int = 1,
freeze: bool = False,
**kwargs):
"""
Create new embeddings for the vocabulary.
Use scaling for the Transformer.
:param embedding_dim:
:param scale:
:param vocab_size:
:param padding_idx:
:param freeze: freeze the embeddings during training
"""
super(Embeddings, self).__init__()
self.embedding_dim = embedding_dim
self.scale = scale
self.vocab_size = vocab_size
self.lut = nn.Embedding(vocab_size,
self.embedding_dim,
padding_idx=padding_idx)
if freeze:
freeze_params(self)
def __init__(self,
num_layers: int = 4,
num_heads: int = 8,
hidden_size: int = 512,
ff_size: int = 2048,
dropout: float = 0.1,
emb_dropout: float = 0.1,
vocab_size: int = 1,
freeze: bool = False,
self_attn_func: str = "softmax",
src_attn_func: str = "softmax",
self_attn_alpha: float = 1.5,
src_attn_alpha: float = 1.5,
gen_func: str = "softmax",
gen_alpha: float = 1.5,
output_bias: bool = False,
**kwargs):
"""
Initialize a Transformer decoder.
:param num_layers: number of Transformer layers
:param num_heads: number of heads for each layer
:param hidden_size: hidden size
:param ff_size: position-wise feed-forward size
:param dropout: dropout probability (1-keep)
:param emb_dropout: dropout probability for embeddings
:param vocab_size: size of the output vocabulary
:param freeze: set to True keep all decoder parameters fixed
:param kwargs:
"""
super(TransformerDecoder, self).__init__(hidden_size,
vocab_size,
emb_dropout,
gen_func=gen_func,
gen_alpha=gen_alpha,
output_bias=output_bias)
# create num_layers decoder layers and put them in a list
self.layers = nn.ModuleList([
TransformerDecoderLayer(size=hidden_size,
ff_size=ff_size,
num_heads=num_heads,
dropout=dropout,
self_attn_func=self_attn_func,
self_attn_alpha=self_attn_alpha,
src_attn_func=src_attn_func,
src_attn_alpha=src_attn_alpha)
for _ in range(num_layers)
])
self.pe = PositionalEncoding(hidden_size)
self.layer_norm = nn.LayerNorm(hidden_size, eps=1e-6)
if freeze:
freeze_params(self)
def __init__(self,
hidden_size: int = 512,
ff_size: int = 2048,
num_layers: int = 8,
num_heads: int = 4,
dropout: float = 0.1,
emb_dropout: float = 0.1,
freeze: bool = False,
dont_minus_one=True,
shared_layers=None,
**kwargs):
"""
Initializes the Transformer.
:param hidden_size: hidden size and size of embeddings
:param ff_size: position-wise feed-forward layer size.
(Typically this is 2*hidden_size.)
:param num_layers: number of layers
:param num_heads: number of heads for multi-headed attention
:param dropout: dropout probability for Transformer layers
:param emb_dropout: Is applied to the input (word embeddings).
:param freeze: freeze the parameters of the encoder during training
:param kwargs:
"""
super(TransformerEncoder, self).__init__()
# build all (num_layers) layers
if shared_layers is not None:
#self.layers = shared_layers
self.layers = nn.ModuleList([layer for layer in shared_layers])
self.layers.append(
TransformerEncoderLayer(size=hidden_size,
ff_size=ff_size,
num_heads=num_heads,
dropout=dropout))
else:
self.layers = nn.ModuleList([
TransformerEncoderLayer(size=hidden_size,
ff_size=ff_size,
num_heads=num_heads,
dropout=dropout)
for _ in range(num_layers if dont_minus_one else num_layers -
1)
])
self.top_off = False if (dont_minus_one
and 'multi_encoder' in locals()
and multi_encoder) else True
self.layer_norm = nn.LayerNorm(hidden_size, eps=1e-6)
self.pe = PositionalEncoding(hidden_size)
self.emb_dropout = nn.Dropout(p=emb_dropout)
self._output_size = hidden_size
if freeze:
freeze_params(self)
def __init__(
self,
num_layers: int = 4,
num_heads: int = 8,
hidden_size: int = 512,
ff_size: int = 2048,
dropout: float = 0.1,
freeze: bool = False,
self_attn_func: str = "softmax",
src_attn_func: str = "softmax",
self_attn_alpha: float = 1.5,
src_attn_alpha: float = 1.5,
merge: str = "serial", # for multi-encoder models
gate_func: str = "softmax",
gate_alpha: float = 1.5,
**kwargs):
"""
Initialize a Transformer decoder.
:param num_layers:
:param num_heads:
:param hidden_size: hidden size
:param ff_size: position-wise feed-forward size
:param dropout:
:param emb_dropout: dropout probability for embeddings
:param freeze: set to True keep all decoder parameters fixed
:param kwargs: passed to generic Decoder Constructor
"""
super(TransformerDecoder, self).__init__(hidden_size, **kwargs)
self.layers = nn.ModuleList([
self.layer_module(size=hidden_size,
ff_size=ff_size,
num_heads=num_heads,
dropout=dropout,
self_attn_func=self_attn_func,
self_attn_alpha=self_attn_alpha,
src_attn_func=src_attn_func,
src_attn_alpha=src_attn_alpha,
merge=merge,
gate_func=gate_func,
gate_alpha=gate_alpha) for _ in range(num_layers)
])
self.pe = PositionalEncoding(hidden_size)
self.layer_norm = nn.LayerNorm(hidden_size, eps=1e-6)
if freeze:
freeze_params(self)
def __init__(self,
embed_file: str,
vocab: Vocabulary,
embedding_dim: int = 64,
scale: bool = False,
vocab_size: int = 0,
padding_idx: int = 1,
freeze: bool = True,
**kwargs):
super(PretrainedEmbeddings, self).__init__(embedding_dim, scale, vocab_size, padding_idx, freeze, **kwargs)
# overwrite lut with embeddings from embed_file
self.load_embeddings_from_file(embed_file, vocab)
if freeze:
freeze_params(self)
def __init__(self,
num_layers: int = 4,
num_heads: int = 8,
hidden_size: int = 512,
ff_size: int = 2048,
dropout: float = 0.1,
emb_dropout: float = 0.1,
vocab_size: int = 1,
freeze: bool = False,
**kwargs):
"""
Initialize a Transformer decoder.
:param num_layers: number of Transformer layers
:param num_heads: number of heads for each layer
:param hidden_size: hidden size
:param ff_size: position-wise feed-forward size
:param dropout: dropout probability (1-keep)
:param emb_dropout: dropout probability for embeddings
:param vocab_size: size of the output vocabulary
:param freeze: set to True keep all decoder parameters fixed
:param kwargs:
"""
super().__init__()
self._hidden_size = hidden_size
self._output_size = vocab_size
# create num_layers decoder layers and put them in a list
self.layers = nn.ModuleList([
TransformerDecoderLayer(size=hidden_size,
ff_size=ff_size,
num_heads=num_heads,
dropout=dropout) for _ in range(num_layers)
])
self.pe = PositionalEncoding(hidden_size)
self.layer_norm = nn.LayerNorm(hidden_size, eps=1e-6)
self.emb_dropout = nn.Dropout(p=emb_dropout)
self.output_layer = nn.Linear(hidden_size, hidden_size, bias=False)
#self.output_layer = self.layers[-1].feed_forward.pwff_layer[-2]
if freeze:
freeze_params(self)
def __init__(self,
num_layers: int = 4,
num_heads: int = 8,
hidden_size: int = 512,
ff_size: int = 2048,
dropout: float = 0.1,
vocab_size: int = 1,
freeze: bool = False,
**kwargs):
"""
Initialize a Transformer decoder.
:param num_layers: number of Transformer layers
:param num_heads: number of heads for each layer
:param hidden_size: hidden size
:param ff_size: position-wise feed-forward size
:param dropout: dropout probability (1-keep)
:param vocab_size: size of the output vocabulary
:param freeze: set to True keep all decoder parameters fixed
:param kwargs:
"""
super(TransformerDecoder, self).__init__()
# build all (num_layers) layers
layers = []
for _ in range(num_layers):
layer = TransformerDecoderLayer(
hidden_size,
MultiHeadedAttention(num_heads, hidden_size, dropout),
MultiHeadedAttention(num_heads, hidden_size, dropout),
PositionwiseFeedForward(hidden_size, ff_size, dropout),
dropout)
layers.append(layer)
self.layers = nn.ModuleList(layers)
self.norm = nn.LayerNorm(hidden_size)
self.pe = PositionalEncoding(hidden_size, dropout=dropout)
self._hidden_size = hidden_size
self._output_size = vocab_size
self.output_layer = nn.Linear(hidden_size, vocab_size, bias=False)
if freeze:
freeze_params(self)
def __init__(self,
num_layers: int = 4,
hidden_size: int = 512,
emb_size: int = 512,
kernel_size: int = 5,
dropout: float = 0.1,
emb_dropout: float = 0.1,
vocab_size: int = 1,
freeze: bool = False,
use_multi_head: bool = False,
num_heads: int = 8,
**kwargs):
"""
Initialize a ConvSeq2Seq decoder.
:param num_layers: number of Transformer layers
:param hidden_size: hidden size
:param dropout: dropout probability (1-keep)
:param emb_dropout: dropout probability for embeddings
:param vocab_size: size of the output vocabulary
:param freeze: set to True keep all decoder parameters fixed
:param kwargs:
"""
super(ConvSeq2SeqDecoder, self).__init__()
self._hidden_size = hidden_size
self._output_size = vocab_size
# create num_layers decoder layers and put them in a list
self.layers = nn.ModuleList([ConvSeq2SeqDecoderLayer(
hidden_size=hidden_size,
embedding_size=emb_size,
kernel_size=kernel_size,
use_multi_head=use_multi_head,
num_heads=num_heads,
dropout=dropout) for _ in range(num_layers)])
self.absPE = AbsolutePositionalEncoding(emb_size)
self.emb2hidden = nn.Linear(emb_size, hidden_size)
self.emb_dropout = nn.Dropout(p=emb_dropout)
self.output_layer = nn.Linear(hidden_size, vocab_size, bias=False)
if freeze:
freeze_params(self)
def __init__(self,
rnn_type: str = "gru",
hidden_size: int = 1,
emb_size: int = 1,
num_layers: int = 1,
dropout: float = 0.,
emb_dropout: float = 0.,
bidirectional: bool = True,
freeze: bool = False,
enforce_sorted: bool = True,
**kwargs) -> None:
"""
Create a new recurrent encoder.
:param rnn_type: RNN type: `gru` or `lstm`.
:param hidden_size: Size of each RNN.
:param emb_size: Size of the word embeddings.
:param num_layers: Number of encoder RNN layers.
:param dropout: Is applied between RNN layers.
:param emb_dropout: Is applied to the RNN input (word embeddings).
:param bidirectional: Use a bi-directional RNN.
:param freeze: freeze the parameters of the encoder during training
:param kwargs:
"""
self._enforce_sorted = enforce_sorted
super(RecurrentEncoder, self).__init__()
self.emb_dropout = nn.Dropout(p=emb_dropout, inplace=False)
self.emb_size = emb_size
rnn = nn.GRU if rnn_type == "gru" else nn.LSTM
self.rnn = rnn(emb_size,
hidden_size,
num_layers,
batch_first=True,
bidirectional=bidirectional,
dropout=dropout if num_layers > 1 else 0.)
self._output_size = 2 * hidden_size if bidirectional else hidden_size
if freeze:
freeze_params(self)
def __init__(self,
hidden_size: int = 512,
ff_size: int = 2048,
num_layers: int = 8,
num_heads: int = 4,
dropout: float = 0.1,
active_layers: list = [],
emb_dropout: float = 0.1,
freeze: bool = False,
**kwargs):
"""
Initializes the Transformer.
:param hidden_size: hidden size and size of embeddings
:param ff_size: position-wise feed-forward layer size.
(Typically this is 2*hidden_size.)
:param num_layers: number of layers
:param num_heads: number of heads for multi-headed attention
:param dropout: dropout probability for Transformer layers
:param emb_dropout: Is applied to the input (word embeddings).
:param freeze: freeze the parameters of the encoder during training
:param kwargs:
"""
super().__init__()
# build all (num_layers) layers or only some
if active_layers != []:
self.layers = nn.ModuleList([
TransformerEncoderLayer(size=hidden_size, ff_size=ff_size,
num_heads=num_heads, dropout=dropout)
for x in active_layers])
else:
self.layers = nn.ModuleList([
TransformerEncoderLayer(size=hidden_size, ff_size=ff_size,
num_heads=num_heads, dropout=dropout)
for _ in range(num_layers)])
self.layer_norm = nn.LayerNorm(hidden_size, eps=1e-6)
self.pe = PositionalEncoding(hidden_size)
self.emb_dropout = nn.Dropout(p=emb_dropout)
self._output_size = hidden_size
if freeze:
freeze_params(self)
def __init__(self,
hidden_size: int = 512,
ff_size: int = 2048,
num_layers: int = 8,
num_heads: int = 4,
dropout: float = 0.1,
emb_dropout: float = 0.1,
freeze: bool = False,
attn_func: str = "softmax",
attn_alpha: float = 1.5,
pe: bool = True,
**kwargs):
"""
Initializes the Transformer.
:param hidden_size: hidden size and size of embeddings
:param ff_size: position-wise feed-forward layer size.
(Typically this is 2*hidden_size.)
:param num_layers: number of layers
:param num_heads: number of heads for multi-headed attention
:param dropout: dropout probability for Transformer layers
:param emb_dropout: Is applied to the input (word embeddings).
:param freeze: freeze the parameters of the encoder during training
:param kwargs:
"""
super(TransformerEncoder, self).__init__()
self.layers = nn.ModuleList([
TransformerEncoderLayer(size=hidden_size,
ff_size=ff_size,
num_heads=num_heads,
dropout=dropout,
attn_func=attn_func,
attn_alpha=attn_alpha)
for _ in range(num_layers)
])
self.layer_norm = nn.LayerNorm(hidden_size, eps=1e-6)
self.pe = PositionalEncoding(hidden_size) if pe else None
self.emb_dropout = nn.Dropout(p=emb_dropout)
self._output_size = hidden_size
if freeze:
freeze_params(self)
def __init__(self,
rnn_type: str = "gru",
hidden_size: int = 1,
emb_size: int = 1,
num_layers: int = 1,
dropout: float = 0.,
bidirectional: bool = True,
freeze: bool = False,
**kwargs) -> None:
"""
Create a new recurrent encoder.
:param rnn_type:
:param hidden_size:
:param emb_size:
:param num_layers:
:param dropout:
:param bidirectional:
:param freeze: freeze the parameters of the encoder during training
:param kwargs:
"""
super(RecurrentEncoder, self).__init__()
self.rnn_input_dropout = torch.nn.Dropout(p=dropout, inplace=False)
self.type = rnn_type
self.emb_size = emb_size
rnn = nn.GRU if rnn_type == "gru" else nn.LSTM
self.rnn = rnn(emb_size,
hidden_size,
num_layers,
batch_first=True,
bidirectional=bidirectional,
dropout=dropout if num_layers > 1 else 0.)
self._output_size = 2 * hidden_size if bidirectional else hidden_size
if freeze:
freeze_params(self)
def __init__(self,
embedding_dim: int = 64,
scale: bool = False,
vocab_size: int = 0,
padding_idx: int = 1,
freeze: bool = False,
from_pretrained: bool = False,
pretrained_path: str = "",
check_embedding=False,
**kwargs):
"""
Create new embeddings for the vocabulary.
Use scaling for the Transformer.
:param embedding_dim:
:param scale:
:param vocab_size:
:param padding_idx:
:param freeze: freeze the embeddings during training
"""
super(Embeddings, self).__init__()
self.embedding_dim = embedding_dim
self.scale = scale
self.vocab_size = vocab_size
self.from_pretrained = from_pretrained
self.check_embedding = check_embedding
if from_pretrained:
print("using pretrained model")
self.weight = torch.load(pretrained_path)
print(f"loaded embeddings size {self.weight.shape}")
self.lut = nn.Embedding.from_pretrained(self.weight)
else:
self.lut = nn.Embedding(vocab_size,
self.embedding_dim,
padding_idx=padding_idx)
if freeze:
freeze_params(self)
def __init__(self,
hidden_size: int = 512,
ff_size: int = 2048,
num_layers: int = 8,
num_heads: int = 4,
dropout: float = 0.1,
freeze: bool = False,
**kwargs):
"""
Initializes the Transformer.
:param hidden_size: hidden size and size of embeddings
:param ff_size: position-wise feed-forward layer size.
(Typically this is 2*hidden_size.)
:param num_layers: number of layers
:param num_heads: number of heads for multi-headed attention
:param dropout: dropout probability
:param freeze: freeze the parameters of the encoder during training
:param kwargs:
"""
super(TransformerEncoder, self).__init__()
# build all (num_layers) layers
layers = []
for _ in range(num_layers):
layer = TransformerEncoderLayer(
hidden_size,
MultiHeadedAttention(num_heads, hidden_size, dropout),
PositionwiseFeedForward(hidden_size, ff_size, dropout),
dropout)
layers.append(layer)
self.layers = nn.ModuleList(layers)
self.norm = nn.LayerNorm(hidden_size)
self.pe = PositionalEncoding(hidden_size, dropout=dropout)
self._output_size = hidden_size
if freeze:
freeze_params(self)
def __init__(self,
rnn_type: str = "gru",
emb_size: int = 0,
hidden_size: int = 0,
encoder_output_sizes: dict = None,
attention: str = "bahdanau",
attn_merge: str = "concat",
num_layers: int = 1,
dropout: float = 0.,
hidden_dropout: float = 0.,
init_hidden: str = "bridge",
input_feeding: bool = True,
freeze: bool = False,
attn_func: str = "softmax",
attn_alpha: float = 1.5,
gate_func: str = "softmax",
gate_alpha: float = 1.5,
**kwargs) -> None:
"""
Todo: document the unique challenges of making an RNN decoder that
attends over multiple encoders
"""
super(MultiHeadRecurrentDecoder, self).__init__(hidden_size, **kwargs)
self.hidden_dropout = nn.Dropout(p=hidden_dropout)
self.head_names = sorted(encoder_output_sizes)
self.emb_size = emb_size
rnn = nn.GRU if rnn_type == "gru" else nn.LSTM
self.input_feeding = input_feeding
input_size = emb_size + hidden_size if input_feeding else emb_size
# the decoder RNN
self.rnn = rnn(input_size,
hidden_size,
num_layers,
batch_first=True,
dropout=dropout if num_layers > 1 else 0.)
# combined output sizes of all encoders
# this quantity matters for concat attention merging
# it also matters if you have a bridge for init_hidden
encoder_output_size = sum(encoder_output_sizes.values())
assert attention in ["bahdanau", "luong"], \
"Unknown attention mechanism: %s. Use 'bahdanau' or 'luong'."
if attention == "bahdanau":
attn_mechanism = partial(MultiAttention, query_size=hidden_size)
else:
attn_mechanism = MultiAttention
self.attention = attn_mechanism(attn_type=attention,
head_names=self.head_names,
key_sizes=encoder_output_sizes,
hidden_size=hidden_size,
attn_func=attn_func,
attn_alpha=attn_alpha,
attn_merge=attn_merge,
gate_func=gate_func,
gate_alpha=gate_alpha)
# to initialize from the final encoder state of last layer
assert init_hidden == "bridge", \
"only use bridge with multi-encoder models"
self.bridge_layer = nn.Sequential(
nn.Linear(encoder_output_size, hidden_size, bias=True), nn.Tanh())
if freeze:
freeze_params(self)
def __init__(self,
rnn_type: str = "gru",
hidden_size: int = 1,
emb_size: int = 1,
num_layers: int = 1,
dropout: float = 0.,
bidirectional: bool = True,
freeze: bool = False,
activation: str = "relu",
last_activation: str = "None",
layer_norm: bool = False,
emb_norm: bool = False,
same_weights: bool = False,
**kwargs) -> None:
"""
Create a new recurrent encoder.
:param rnn_type:
:param hidden_size:
:param emb_size:
:param num_layers:
:param dropout:
:param bidirectional:
:param freeze: freeze the parameters of the encoder during training
:param kwargs:
"""
super(SpeechRecurrentEncoder, self).__init__()
self.rnn_input_dropout = torch.nn.Dropout(p=dropout, inplace=False)
self.type = rnn_type
self.emb_size = emb_size
self.lila1 = nn.Linear(emb_size, hidden_size)
self.lila2 = nn.Linear(hidden_size, hidden_size)
self.same_weights = same_weights
if not self.same_weights:
self.lila3 = nn.Linear(hidden_size, hidden_size)
self.lila4 = nn.Linear(hidden_size, hidden_size)
self.activation = activation
self.last_activation = last_activation
self.conv1 = nn.Sequential(
nn.Conv1d(hidden_size, hidden_size, kernel_size=3, stride=1, padding=1),
nn.ReLU(),
nn.MaxPool1d(kernel_size=2, stride=2, padding=0))
self.conv2 = nn.Sequential(
nn.Conv1d(hidden_size, hidden_size, kernel_size=3, stride=1, padding=1),
nn.ReLU(),
nn.MaxPool1d(kernel_size=2, stride=2, padding=0))
self.layer_norm = layer_norm
self.emb_norm = emb_norm
if self.layer_norm:
self.norm1 = nn.LayerNorm(hidden_size)
self.norm2 = nn.LayerNorm(hidden_size)
self.norm_out = nn.LayerNorm(2 * hidden_size if bidirectional else hidden_size)
if self.emb_norm:
self.norm_emb = nn.LayerNorm(emb_size)
rnn = nn.GRU if rnn_type == "gru" else nn.LSTM
self.rnn = rnn(
hidden_size, hidden_size, num_layers, batch_first=True,
bidirectional=bidirectional,
dropout=dropout if num_layers > 1 else 0.)
self._output_size = 2 * hidden_size if bidirectional else hidden_size
if freeze:
freeze_params(self)
def __init__(self,
rnn_type: str = "gru",
emb_size: int = 0,
hidden_size: int = 0,
encoder: Encoder = None,
attention: str = "bahdanau",
num_layers: int = 1,
vocab_size: int = 0,
dropout: float = 0.,
emb_dropout: float = 0.,
hidden_dropout: float = 0.,
init_hidden: str = "bridge",
input_feeding: bool = True,
freeze: bool = False,
**kwargs) -> None:
"""
Create a recurrent decoder with attention.
:param rnn_type: rnn type, valid options: "lstm", "gru"
:param emb_size: target embedding size
:param hidden_size: size of the RNN
:param encoder: encoder connected to this decoder
:param attention: type of attention, valid options: "bahdanau", "luong"
:param num_layers: number of recurrent layers
:param vocab_size: target vocabulary size
:param hidden_dropout: Is applied to the input to the attentional layer.
:param dropout: Is applied between RNN layers.
:param emb_dropout: Is applied to the RNN input (word embeddings).
:param init_hidden: If "bridge" (default), the decoder hidden states are
initialized from a projection of the last encoder state,
if "zeros" they are initialized with zeros,
if "last" they are identical to the last encoder state
(only if they have the same size)
:param input_feeding: Use Luong's input feeding.
:param freeze: Freeze the parameters of the decoder during training.
:param kwargs:
"""
super().__init__()
self.emb_dropout = torch.nn.Dropout(p=emb_dropout, inplace=False)
self.type = rnn_type
self.hidden_dropout = torch.nn.Dropout(p=hidden_dropout, inplace=False)
self.hidden_size = hidden_size
self.emb_size = emb_size
rnn = nn.GRU if rnn_type == "gru" else nn.LSTM
self.input_feeding = input_feeding
if self.input_feeding: # Luong-style
# combine embedded prev word +attention vector before feeding to rnn
self.rnn_input_size = emb_size + hidden_size
else:
# just feed prev word embedding
self.rnn_input_size = emb_size
# the decoder RNN
self.rnn = rnn(self.rnn_input_size,
hidden_size,
num_layers,
batch_first=True,
dropout=dropout if num_layers > 1 else 0.)
# combine output with context vector before output layer (Luong-style)
self.att_vector_layer = nn.Linear(hidden_size + encoder.output_size,
hidden_size,
bias=True)
self.output_layer = nn.Linear(hidden_size, vocab_size, bias=False)
self._output_size = vocab_size
if attention == "bahdanau":
self.attention = BahdanauAttention(hidden_size=hidden_size,
key_size=encoder.output_size,
query_size=hidden_size)
elif attention == "luong":
self.attention = LuongAttention(hidden_size=hidden_size,
key_size=encoder.output_size)
else:
raise ConfigurationError("Unknown attention mechanism: %s. "
"Valid options: 'bahdanau', 'luong'." %
attention)
self.num_layers = num_layers
self.hidden_size = hidden_size
# to initialize from the final encoder state of last layer
self.init_hidden_option = init_hidden
if self.init_hidden_option == "bridge":
self.bridge_layer = nn.Linear(encoder.output_size,
hidden_size,
bias=True)
elif self.init_hidden_option == "last":
if encoder.output_size != self.hidden_size:
if encoder.output_size != 2 * self.hidden_size: # bidirectional
raise ConfigurationError(
"For initializing the decoder state with the "
"last encoder state, their sizes have to match "
"(encoder: {} vs. decoder: {})".format(
encoder.output_size, self.hidden_size))
if freeze:
freeze_params(self)
def __init__(self,
type: str = "gru",
emb_size: int = 0,
hidden_size: int = 0,
encoder: Encoder = None,
attention: str = "bahdanau",
num_layers: int = 0,
vocab_size: int = 0,
dropout: float = 0.,
hidden_dropout: float = 0.,
bridge: bool = False,
input_feeding: bool = True,
freeze: bool = False,
**kwargs):
"""
Create a recurrent decoder.
If `bridge` is True, the decoder hidden states are initialized from a
projection of the encoder states, else they are initialized with zeros.
:param type:
:param emb_size:
:param hidden_size:
:param encoder:
:param attention:
:param num_layers:
:param vocab_size:
:param dropout:
:param hidden_dropout:
:param bridge:
:param input_feeding:
:param freeze: freeze the parameters of the decoder during training
:param kwargs:
"""
super(RecurrentDecoder, self).__init__()
self.rnn_input_dropout = torch.nn.Dropout(p=dropout, inplace=False)
self.type = type
self.hidden_dropout = torch.nn.Dropout(p=hidden_dropout, inplace=False)
self.hidden_size = hidden_size
rnn = nn.GRU if type == "gru" else nn.LSTM
self.input_feeding = input_feeding
if self.input_feeding: # Luong-style
# combine embedded prev word +attention vector before feeding to rnn
self.rnn_input_size = emb_size + hidden_size
else:
# just feed prev word embedding
self.rnn_input_size = emb_size
# the decoder RNN
self.rnn = rnn(self.rnn_input_size, hidden_size, num_layers,
batch_first=True,
dropout=dropout if num_layers > 1 else 0.)
# combine output with context vector before output layer (Luong-style)
self.att_vector_layer = nn.Linear(
hidden_size + encoder.output_size, hidden_size, bias=True)
self.output_layer = nn.Linear(hidden_size, vocab_size, bias=False)
self.output_size = vocab_size
if attention == "bahdanau":
self.attention = BahdanauAttention(hidden_size=hidden_size,
key_size=encoder.output_size,
query_size=hidden_size)
elif attention == "luong":
self.attention = LuongAttention(hidden_size=hidden_size,
key_size=encoder.output_size)
else:
raise ValueError("Unknown attention mechanism: %s" % attention)
self.num_layers = num_layers
self.hidden_size = hidden_size
# to initialize from the final encoder state of last layer
self.bridge = bridge
if self.bridge:
self.bridge_layer = nn.Linear(
encoder.output_size, hidden_size, bias=True)
if freeze:
freeze_params(self)
def __init__(
self,
src_vocab: Vocabulary,
trg_vocab: Vocabulary,
embedding_dim: int = 300, # or 30
scale: bool = False,
vocab_size: int = 0,
padding_idx: int = 1,
freeze: bool = False,
**kwargs
):
"""
Create new embeddings for the vocabulary.
Use scaling for the Transformer.
:param embedding_dim:
:param scale:
:param vocab_size:
:param padding_idx:
:param freeze: freeze the embeddings during training
"""
super().__init__()
self.scale = scale
# TODO add support for other languages
# fasttext.util.download_model('de', if_exists='ignore')
# 30 or 300
np_embedding = f"ft_deen_{embedding_dim}.np"
if not os.path.isfile(np_embedding):
# if else call to avoid importing fasttext if possible
# because fasttext has unmet dependency on gpu nodes
import fasttext.util
src_ft = fasttext.load_model(f'cc.de.{embedding_dim}.bin')
trg_ft = fasttext.load_model(f'cc.en.{embedding_dim}.bin')
# Create smaller embeddings, to test on reverse
# fasttext.util.reduce_model(src_ft, 30)
# src_ft.save_model('cc.en.30.bin')
self.embedding_dim = src_ft.get_dimension()
vectors = []
for i, word in tqdm(enumerate(src_vocab.itos), desc="adding src vecs"):
vectors.append(src_ft.get_word_vector(word))
for i, word in tqdm(enumerate(trg_vocab.itos), desc="adding trg vecs"):
vectors.append(trg_ft.get_word_vector(word))
embedding_matrix = np.vstack(vectors)
with open(np_embedding, "wb") as np_file:
pickle.dump(embedding_matrix, np_file)
print(f"Saved joint fasttext embedding matrix as np matrix at {np_embedding}")
else:
print("Loading saved embedding ...")
with open(np_embedding, "rb") as np_file:
embedding_matrix = pickle.load(np_file)
print("Loaded saved embedding.")
self.embedding_dim = embedding_matrix.shape[-1]
self.lut = nn.Embedding(
len(src_vocab)+len(trg_vocab),
self.embedding_dim,
padding_idx=trg_vocab.stoi[PAD_TOKEN]
)
self.lut.weight = nn.Parameter(data=torch.from_numpy(embedding_matrix).float())
assert self.lut.weight is not None
# assert False, self.lut.weight.shape
# always freeze pretrained embeddings
freeze_params(self)
def __init__(self,
rnn_type: str = "gru",
emb_size: int = 0,
hidden_size: int = 0,
encoder_output_size: int = 0,
attention: str = "bahdanau",
num_layers: int = 1,
vocab_size: int = 0,
dropout: float = 0.,
emb_dropout: float = 0.,
hidden_dropout: float = 0.,
init_hidden: str = "bridge",
input_feeding: bool = True,
freeze: bool = False,
attn_func: str = "softmax",
attn_alpha: float = 1.5,
gen_func: str = "softmax",
gen_alpha: float = 1.5,
output_bias: bool = False,
multi_source: bool = False,
head_names: list = None,
attn_merge: str = "gate",
gate_func: str = "softmax",
gate_alpha: float = 1.5,
**kwargs) -> None:
"""
Create a recurrent decoder with attention.
:param rnn_type: rnn type, valid options: "lstm", "gru"
:param emb_size: target embedding size
:param hidden_size: size of the RNN
:param encoder_output_size:
:param attention: type of attention, valid options: "bahdanau", "luong"
:param num_layers: number of recurrent layers
:param vocab_size: target vocabulary size
:param hidden_dropout: applied to the input to the attentional layer.
:param dropout: Is applied between RNN layers.
:param emb_dropout: Is applied to the RNN input (word embeddings).
:param init_hidden: If "bridge" (default), the decoder hidden states
are initialized from a projection of the last encoder state,
if "zeros" they are initialized with zeros,
if "last" they are identical to the last encoder state
(only if they have the same size)
:param input_feeding: Use Luong's input feeding.
:param freeze: Freeze the parameters of the decoder during training.
:param kwargs:
"""
super(RecurrentDecoder, self).__init__(hidden_size,
vocab_size,
emb_dropout,
gen_func=gen_func,
gen_alpha=gen_alpha,
output_bias=output_bias)
self.multi_source = multi_source
self.hidden_dropout = nn.Dropout(p=hidden_dropout)
self.emb_size = emb_size
rnn = nn.GRU if rnn_type == "gru" else nn.LSTM
self.input_feeding = input_feeding
input_size = emb_size + hidden_size if input_feeding else emb_size
# the decoder RNN
self.rnn = rnn(input_size,
hidden_size,
num_layers,
batch_first=True,
dropout=dropout if num_layers > 1 else 0.)
# combine output with context vector before output layer (Luong-style)
self.att_vector_layer = nn.Linear(hidden_size + encoder_output_size,
hidden_size,
bias=True)
assert attention in ["bahdanau", "luong"], \
"Unknown attention mechanism: %s. Use 'bahdanau' or 'luong'."
if multi_source:
attn_mechanism = partial(MultiAttention,
attn_type=attention,
head_names=head_names,
attn_merge=attn_merge,
gate_func=gate_func,
gate_alpha=gate_alpha)
elif attention == "luong":
attn_mechanism = LuongAttention
else:
attn_mechanism = BahdanauAttention
if attention == "bahdanau":
attn_mechanism = partial(attn_mechanism, query_size=hidden_size)
self.attention = attn_mechanism(hidden_size=hidden_size,
key_size=encoder_output_size,
attn_func=attn_func,
attn_alpha=attn_alpha)
# init_hidden: "bridge", "zero", "last", or a dictionary describing
# an arbitrary-layered MLP
assert isinstance(init_hidden, dict) or isinstance(init_hidden, str), \
'''
Specify either a shortcut name ("bridge", "zero", "last") or a
dictionary containing a configuration for the bridge layer.
'''
if init_hidden == "zero":
self.bridge_layer = None # easy-peasy
else:
if init_hidden == "last":
# not actually clear to me if this is necessary
assert encoder_output_size in {hidden_size, 2 * hidden_size}, \
"Mismatched hidden sizes (enc: {}, dec: {})".format(
encoder_output_size, hidden_size
)
if isinstance(init_hidden, str):
bridge = init_hidden == "bridge"
# 'bridge' and 'last' are shortcuts to specific special cases
init_hidden = {
"num_layers": 1 if bridge else 0,
"activation": "tanh" if bridge else "none",
"merge": "cat"
}
if init_hidden["merge"] == "cat":
n_heads = len(head_names) if head_names is not None else 1
bridge_in_size = encoder_output_size * n_heads # for cat
else:
bridge_in_size = encoder_output_size
self.bridge_layer = Bridge(bridge_in_size,
hidden_size,
lstm=isinstance(self.rnn, nn.LSTM),
decoder_layers=self.num_layers,
**init_hidden)
if freeze:
freeze_params(self)
def __init__(self,
rnn_type: str = "gru",
emb_size: int = 0,
hidden_size: int = 0,
encoder_output_size: int = 0,
attention: str = "bahdanau",
num_layers: int = 1,
dropout: float = 0.,
hidden_dropout: float = 0.,
init_hidden: str = "bridge",
input_feeding: bool = True,
freeze: bool = False,
attn_func: str = "softmax",
attn_alpha: float = 1.5,
**kwargs) -> None:
"""
Create a recurrent decoder with attention.
:param rnn_type: rnn type, valid options: "lstm", "gru"
:param emb_size:
:param hidden_size:
:param attention: type of attention, valid options: "bahdanau", "luong"
:param num_layers:
:param hidden_dropout: applied to the input to the attentional layer.
:param dropout: applied between RNN layers.
:param emb_dropout: applied to the RNN input (word embeddings).
:param init_hidden: If "bridge" (default), the decoder hidden states
are initialized from a projection of the last encoder state,
if "zeros" they are initialized with zeros,
if "last" they are identical to the last encoder state
(only if they have the same size)
:param input_feeding: Use Luong's input feeding.
:param freeze: Freeze the parameters of the decoder during training.
:param kwargs: passed to generic Decoder constructor
"""
super(RecurrentDecoder, self).__init__(hidden_size, **kwargs)
self.hidden_dropout = nn.Dropout(p=hidden_dropout)
self.emb_size = emb_size
rnn = nn.GRU if rnn_type == "gru" else nn.LSTM
self.input_feeding = input_feeding
input_size = emb_size + hidden_size if input_feeding else emb_size
# the decoder RNN
self.rnn = rnn(input_size,
hidden_size,
num_layers,
batch_first=True,
dropout=dropout if num_layers > 1 else 0.)
# combine output with context vector before output layer (Luong-style)
self.att_vector_layer = nn.Linear(hidden_size + encoder_output_size,
hidden_size,
bias=True)
assert attention in ["bahdanau", "luong"], \
"Unknown attention mechanism: %s. Use 'bahdanau' or 'luong'."
if attention == "bahdanau":
attn_mechanism = partial(BahdanauAttention, query_size=hidden_size)
else:
attn_mechanism = LuongAttention
self.attention = attn_mechanism(hidden_size=hidden_size,
key_size=encoder_output_size,
attn_func=attn_func,
attn_alpha=attn_alpha)
# to initialize from the final encoder state of last layer
assert init_hidden in ["bridge", "zero", "last"]
self.init_hidden_option = init_hidden
if init_hidden == "bridge":
self.bridge_layer = nn.Sequential(
nn.Linear(encoder_output_size, hidden_size, bias=True),
nn.Tanh())
else:
self.bridge_layer = None
if init_hidden == "last":
out_size = encoder_output_size
assert out_size in (hidden_size, 2 * hidden_size), \
"Mismatched hidden sizes (encoder: {}, decoder: {})".format(
encoder_output_size, hidden_size
)
if freeze:
freeze_params(self)
