def __init__(self, opt, dicts):
self.layers = opt.layers
self.input_feed = opt.input_feed
input_size = opt.word_vec_size
if self.input_feed:
input_size += opt.rnn_size
super(Decoder, self).__init__()
self.word_lut = nn.Embedding(dicts.size(),
opt.word_vec_size,
padding_idx=onmt.Constants.PAD)
stackedCell = StackedLSTM if opt.rnn_type == "LSTM" else StackedGRU
self.rnn = stackedCell(opt.layers, input_size, opt.rnn_size,
opt.dropout)
self.attn = onmt.modules.GlobalAttention(opt.rnn_size)
self.context_gate = None
if opt.context_gate is not None:
self.context_gate = ContextGateFactory(opt.context_gate,
opt.word_vec_size,
opt.rnn_size, opt.rnn_size,
opt.rnn_size)
self.dropout = nn.Dropout(opt.dropout)
self.hidden_size = opt.rnn_size
def __init__(self, opt, dicts):
"""
Args:
opt: model options
dicts: Target `Dict` object
"""
self.layers = opt.dec_layers
self.decoder_type = opt.decoder_type
self._coverage = opt.coverage_attn
self.hidden_size = opt.rnn_size
self.input_feed = opt.input_feed
input_size = opt.tgt_word_vec_size
if self.input_feed:
input_size += opt.rnn_size
super(Decoder, self).__init__()
self.embeddings = Embeddings(opt.tgt_word_vec_size,
opt, dicts, None)
pad_id = dicts.stoi[onmt.IO.PAD_WORD]
if self.decoder_type == "transformer":
self.transformer = nn.ModuleList(
[onmt.modules.TransformerDecoder(self.hidden_size, opt, pad_id)
for _ in range(opt.dec_layers)])
else:
if self.input_feed:
if opt.rnn_type == "LSTM":
stackedCell = onmt.modules.StackedLSTM
else:
stackedCell = onmt.modules.StackedGRU
self.rnn = stackedCell(opt.dec_layers, input_size,
opt.rnn_size, opt.dropout)
else:
self.rnn = getattr(nn, opt.rnn_type)(
input_size, opt.rnn_size,
num_layers=opt.dec_layers,
dropout=opt.dropout
)
self.context_gate = None
if opt.context_gate is not None:
self.context_gate = ContextGateFactory(
opt.context_gate, input_size,
opt.rnn_size, opt.rnn_size,
opt.rnn_size
)
self.dropout = nn.Dropout(opt.dropout)
# Std attention layer.
self.attn = onmt.modules.GlobalAttention(
opt.rnn_size,
coverage=self._coverage,
attn_type=opt.global_attention)
# Separate Copy Attention.
self._copy = False
if opt.copy_attn:
self.copy_attn = onmt.modules.GlobalAttention(
opt.rnn_size, attn_type=opt.global_attention)
self._copy = True
def __init__(self, opt, dicts):
"""
Args:
opt: model options
dicts: Target `Dict` object
"""
self.layers = opt.layers
self.decoder_layer = opt.decoder_layer
self._coverage = opt.coverage_attn
self.exhaustion_loss = opt.exhaustion_loss
self.fertility_loss = True if opt.supervised_fertility else False
self.hidden_size = opt.rnn_size
self.input_feed = opt.input_feed
input_size = opt.word_vec_size
if self.input_feed:
input_size += opt.rnn_size
super(Decoder, self).__init__()
self.embeddings = Embeddings(opt, dicts, None)
if self.decoder_layer == "transformer":
self.transformer = nn.ModuleList([
onmt.modules.TransformerDecoder(self.hidden_size, opt)
for _ in range(opt.layers)
])
else:
if opt.rnn_type == "LSTM":
stackedCell = onmt.modules.StackedLSTM
else:
stackedCell = onmt.modules.StackedGRU
self.rnn = stackedCell(opt.layers, input_size, opt.rnn_size,
opt.dropout)
self.context_gate = None
if opt.context_gate is not None:
self.context_gate = ContextGateFactory(opt.context_gate,
opt.word_vec_size,
input_size,
opt.rnn_size,
opt.rnn_size)
self.dropout = nn.Dropout(opt.dropout)
# Std attention layer.
self.attn = onmt.modules.GlobalAttention(
opt.rnn_size,
coverage=self._coverage,
attn_type=opt.attention_type,
attn_transform=opt.attn_transform,
c_attn=opt.c_attn)
self.fertility = opt.fertility
self.predict_fertility = opt.predict_fertility
self.guided_fertility = opt.guided_fertility
self.supervised_fertility = opt.supervised_fertility
# Separate Copy Attention.
self._copy = False
if opt.copy_attn:
self.copy_attn = onmt.modules.GlobalAttention(
opt.rnn_size, attn_type=opt.attention_type)
self._copy = True
def __init__(self, opt, dicts):
"""
Args:
opt: model options
dicts: Target `Dict` object
"""
self.layers = opt.layers
self.input_feed = opt.input_feed
input_size = opt.word_vec_size
if self.input_feed:
input_size += opt.rnn_size
self.decoder_layer = opt.__dict__.get("decoder_layer", "")
self.positional_encoding = opt.__dict__.get("position_encoding", "")
self._coverage = opt.__dict__.get("coverage_attn", False)
self.hidden_size = opt.rnn_size
super(Decoder, self).__init__()
self.word_lut = nn.Embedding(dicts.size(),
opt.word_vec_size,
padding_idx=onmt.Constants.PAD)
if self.decoder_layer == "transformer":
self.transformer = nn.ModuleList([
onmt.modules.TransformerDecoder(self.hidden_size, opt)
for _ in range(opt.layers)
])
stackedCell = StackedLSTM if opt.rnn_type == "LSTM" else StackedGRU
self.rnn = stackedCell(opt.layers, input_size, opt.rnn_size,
opt.dropout)
self.context_gate = None
if opt.context_gate is not None:
self.context_gate = ContextGateFactory(opt.context_gate,
opt.word_vec_size,
opt.rnn_size, opt.rnn_size,
opt.rnn_size)
self.dropout = nn.Dropout(opt.dropout)
self.emb_dropout = nn.Dropout(opt.dropout)
# Std attention layer.
self.attn = onmt.modules.GlobalAttention(opt.rnn_size, self._coverage)
# Separate Copy Attention.
self._copy = False
if opt.__dict__.get("copy_attn", False):
self.copy_attn = onmt.modules.GlobalAttention(opt.rnn_size)
self._copy = True
# Use positional encodings
if self.positional_encoding:
self.pe = make_positional_encodings(opt.word_vec_size, 5000).cuda()
def __init__(self, opt):
"""
Args:
opt: model options
dicts: Target `Dict` object
"""
self.layers = opt.layers
self.decoder_layer = opt.decoder_layer
self._coverage = opt.coverage_attn
self.hidden_size = opt.rnn_size
self.input_feed = opt.input_feed
input_size = opt.word_vec_size
if self.input_feed:
input_size += opt.rnn_size
super(DecoderVideo, self).__init__()
#self.embeddings = Embeddings(opt, dicts, None)
if opt.rnn_type == "LSTM":
stackedCell = onmt.modules.StackedLSTM
else:
stackedCell = onmt.modules.StackedGRU
self.rnn = stackedCell(opt.layers, input_size,
opt.rnn_size, opt.dropout)
self.context_gate = None
if opt.context_gate is not None:
self.context_gate = ContextGateFactory(
opt.context_gate, input_size,
opt.rnn_size, opt.rnn_size,
opt.rnn_size
)
self.dropout = nn.Dropout(opt.dropout)
# Std attention layer.
self.attn = onmt.modules.GlobalAttention(opt.rnn_size,
coverage=self._coverage,
attn_type=opt.attention_type)
# Separate Copy Attention.
self._copy = False
if opt.copy_attn:
self.copy_attn = onmt.modules.GlobalAttention(
opt.rnn_size, attn_type=opt.attention_type)
self._copy = True
def __init__(self, rnn_type, num_layers, hidden_size,
attn_type, coverage_attn, context_gate,
copy_attn, dropout, embeddings):
"""
See make_decoder() comment for arguments description.
"""
super(RNNDecoderBase, self).__init__()
# Basic attributes.
self.decoder_type = 'rnn'
self.num_layers = num_layers
self.hidden_size = hidden_size
self.embeddings = embeddings
self.dropout = nn.Dropout(dropout)
# Build the RNN.
self.rnn = self._build_rnn(rnn_type, self._input_size, hidden_size,
num_layers, dropout)
# Set up the context gate.
self.context_gate = None
if context_gate is not None:
self.context_gate = ContextGateFactory(
context_gate, self._input_size,
hidden_size, hidden_size, hidden_size
)
# Set up the standard attention.
self._coverage = coverage_attn
self.attn = onmt.modules.GlobalAttention(
hidden_size, coverage=coverage_attn,
attn_type=attn_type
)
# Set up a separated copy attention layer, if needed.
self._copy = False
if copy_attn:
self.copy_attn = onmt.modules.GlobalAttention(
hidden_size, attn_type=attn_type
)
self._copy = True
