def test_transformer_decoder(num_beams, decoder_inputs_transformer):
batch_size, emb_size, nlayers, sl, vin, ven = decoder_inputs_transformer
ntokens, nhid, max_tokens = 10, 2, 20
embedding = TransformerEmbeddings(ntokens=ntokens,
emb_size=emb_size,
dropout=0.0,
pad_token=1)
encoder = TransformerDecoderLayers(nlayers=nlayers,
input_size=emb_size,
num_heads=2,
nhid=emb_size)
projection_layer = Projection(output_size=ntokens,
input_size=emb_size,
tie_encoder=None,
dropout=0.0)
decoder = TransformerDecoder(decoder_layer=encoder,
projection_layer=projection_layer,
pad_token=1,
eos_token=2,
max_tokens=max_tokens,
embedding_layer=embedding)
decoder = to_gpu(decoder)
outputs = decoder(vin, ven, num_beams=num_beams)
if num_beams > 0:
assert_dims(outputs,
[None, num_beams * batch_size, (emb_size, ntokens)])
# actual beam outputs can be found in beam_outputs
assert decoder.beam_outputs is not None
assert_dims(decoder.beam_outputs, [None, batch_size, num_beams])
# the sl can go up to max_tokens + 1(for the extra 0 token at the end)
assert 0 < decoder.beam_outputs.shape[0] <= max_tokens + 1
else:
assert_dims(outputs, [None, batch_size, (emb_size, ntokens)])
assert decoder.beam_outputs is None
def __init__(self,
ntoken,
emb_size=512,
nlayers=6,
pad_token=None,
eos_token=None,
max_tokens=200,
share_embedding_layer=False,
tie_decoder=True,
**kwargs):
super().__init__()
ntoken = get_list(ntoken, 2)
self.nlayers = nlayers
dropout = get_kwarg(kwargs, name="dropout", default_value=0.1)
num_heads = get_kwarg(kwargs, name="num_heads", default_value=8)
nhid = get_kwarg(kwargs, name="nhid", default_value=2048)
encoder_embedding_layer = TransformerEmbeddings(ntokens=ntoken[0],
emb_size=emb_size,
dropout=dropout,
pad_token=pad_token)
encoder_layer = TransformerEncoderLayers(num_layers=nlayers,
input_size=emb_size,
num_heads=num_heads,
nhid=nhid)
self.encoder = Encoder(embedding_layer=encoder_embedding_layer,
encoder_layer=encoder_layer)
if share_embedding_layer:
decoder_embedding_layer = encoder_embedding_layer
else:
decoder_embedding_layer = TransformerEmbeddings(
ntokens=ntoken[-1],
emb_size=emb_size,
dropout=dropout,
pad_token=pad_token)
decoder_layer = TransformerDecoderLayers(nlayers=nlayers,
input_size=emb_size,
num_heads=num_heads,
nhid=nhid)
projection_layer = Projection(
output_size=ntoken[-1],
input_size=emb_size,
dropout=dropout,
tie_encoder=decoder_embedding_layer if tie_decoder else None)
self.decoder = TransformerDecoder(
decoder_layer=decoder_layer,
projection_layer=projection_layer,
embedding_layer=decoder_embedding_layer,
pad_token=pad_token,
eos_token=eos_token,
max_tokens=max_tokens,
)
self.nt = ntoken[-1]
# xavier uniform initialization
for p in self.parameters():
if p.dim() > 1:
nn.init.xavier_uniform_(p)
def rnn_decoder(decoder_params):
decoder_embedding_layer = DropoutEmbeddings(
ntokens=decoder_params.ntokens,
emb_size=decoder_params.emb_size,
)
if decoder_params.attention:
# attention decoder must have double the input_size to accommodate for the attention concat
decoder_rnn = RNNLayers(input_size=decoder_params.emb_size * 2,
output_size=decoder_params.emb_size,
nhid=decoder_params.nhid,
bidir=False,
nlayers=decoder_params.nlayers,
cell_type="gru")
projection_layer = AttentionProjection(
output_size=decoder_params.ntokens,
input_size=decoder_params.emb_size,
att_nhid=decoder_params.att_hid,
tie_encoder=None,
dropout=0.0)
decoder = AttentionDecoder(decoder_layer=decoder_rnn,
embedding_layer=decoder_embedding_layer,
projection_layer=projection_layer,
pad_token=1,
eos_token=2,
max_tokens=decoder_params.max_tokens)
else:
decoder_rnn = RNNLayers(input_size=decoder_params.emb_size,
output_size=decoder_params.emb_size,
nhid=decoder_params.nhid,
bidir=False,
nlayers=decoder_params.nlayers,
cell_type="gru")
projection_layer = Projection(output_size=decoder_params.ntokens,
input_size=decoder_params.emb_size,
dropout=0.0,
tie_encoder=None)
decoder = Decoder(
decoder_layer=decoder_rnn,
projection_layer=projection_layer,
embedding_layer=decoder_embedding_layer,
pad_token=0,
eos_token=1,
max_tokens=decoder_params.max_tokens,
)
decoder = to_gpu(decoder)
decoder.reset(decoder_params.batch_size)
return decoder, decoder_params
def rnn_decoder(decoder_params):
if decoder_params.attention:
decoder = RNNAttentionDecoder(cell_type="gru", ntoken=decoder_params.ntokens,
emb_sz=decoder_params.emb_size, nhid=decoder_params.nhid,
nlayers=decoder_params.nlayers,
pad_token=1, eos_token=2,
max_tokens=decoder_params.max_tokens)
decoder.projection_layer = AttentionProjection(n_out=decoder_params.ntokens,
n_in=decoder_params.emb_size,
att_nhid=decoder_params.att_hid,
tie_encoder=None,
dropout=0.0)
else:
decoder = EmbeddingRNNDecoder(cell_type="gru", ntoken=decoder_params.ntokens,
emb_sz=decoder_params.emb_size, nhid=decoder_params.nhid,
nlayers=decoder_params.nlayers,
pad_token=1, eos_token=2,
max_tokens=decoder_params.max_tokens)
decoder.projection_layer = Projection(n_out=decoder_params.ntokens,
n_in=decoder_params.emb_size, tie_encoder=None, dropout=0.0)
decoder = to_gpu(decoder)
decoder.reset(decoder_params.batch_size)
return decoder, decoder_params
def __init__(self,
ntoken: int,
emb_sz: HParam,
nhid: HParam,
nlayers: HParam,
pad_token: int,
eos_token: int,
max_tokens: int = 50,
share_embedding_layer: bool = False,
tie_decoder: bool = True,
bidir: bool = False,
session_constraint: bool = False,
**kwargs):
"""
Args:
ntoken (int): Number of tokens for the encoder and the decoder
emb_sz (Union[List[int],int]): Embedding size for the encoder and decoder embeddings
nhid (Union[List[int],int]): Number of hidden dims for the encoder (first two values) and the decoder
nlayers (Union[List[int],int]): Number of layers for the encoder and the decoder
pad_token (int): The index of the token used for padding
eos_token (int): The index of the token used for eos
max_tokens (int): The maximum number of steps the decoder iterates before stopping
share_embedding_layer (bool): if True the decoder shares its input and output embeddings
tie_decoder (bool): if True the encoder and the decoder share their embeddings
bidir (bool): if True use a bidirectional encoder
session_constraint (bool) If true the session will be concated as a constraint to the decoder input
**kwargs: Extra embeddings that will be passed to the encoder and the decoder
"""
super().__init__()
# allow for the same or different parameters between encoder and decoder
ntoken, emb_sz, nhid, nlayers = get_list(ntoken), get_list(
emb_sz, 2), get_list(nhid, 3), get_list(nlayers, 3)
dropoutd = get_kwarg(kwargs, name="dropout_d",
default_value=0.5) # output dropout
dropoute = get_kwarg(kwargs, name="dropout_e",
default_value=0.1) # encoder embedding dropout
dropoute = get_list(dropoute, 2)
dropouti = get_kwarg(kwargs, name="dropout_i",
default_value=0.65) # input dropout
dropouti = get_list(dropouti, 2)
dropouth = get_kwarg(kwargs, name="dropout_h",
default_value=0.3) # RNN output layers dropout
dropouth = get_list(dropouth, 3)
wdrop = get_kwarg(kwargs, name="wdrop",
default_value=0.5) # RNN weights dropout
wdrop = get_list(wdrop, 3)
train_init = kwargs.pop(
"train_init", False) # Have trainable initial states to the RNNs
dropoutinit = get_kwarg(
kwargs, name="dropout_init",
default_value=0.1) # RNN initial states dropout
dropoutinit = get_list(dropoutinit, 3)
self.cell_type = "gru"
self.nt = ntoken[-1]
self.pr_force = 1.0
encoder_embedding_layer = DropoutEmbeddings(ntokens=ntoken[0],
emb_size=emb_sz[0],
dropoute=dropoute[0],
dropouti=dropouti[0])
encoder_rnn = RNNLayers(input_size=emb_sz[0],
output_size=kwargs.get("output_size_encoder",
emb_sz[0]),
nhid=nhid[0],
bidir=bidir,
dropouth=dropouth[0],
wdrop=wdrop[0],
nlayers=nlayers[0],
cell_type=self.cell_type,
train_init=train_init,
dropoutinit=dropoutinit[0])
self.query_encoder = Encoder(embedding_layer=encoder_embedding_layer,
encoder_layer=encoder_rnn)
self.se_enc = RNNLayers(cell_type=self.cell_type,
input_size=encoder_rnn.output_size,
output_size=nhid[1],
nhid=nhid[1],
nlayers=1,
dropouth=dropouth[1],
wdrop=wdrop[1],
train_init=train_init,
dropoutinit=dropoutinit[1])
if share_embedding_layer:
decoder_embedding_layer = encoder_embedding_layer
else:
decoder_embedding_layer = DropoutEmbeddings(ntokens=ntoken[0],
emb_size=emb_sz[1],
dropoute=dropoute[1],
dropouti=dropouti[1])
input_size_decoder = kwargs.get("input_size_decoder", emb_sz[1])
input_size_decoder = input_size_decoder + self.se_enc.output_size if session_constraint else input_size_decoder
decoder_rnn = RNNLayers(input_size=input_size_decoder,
output_size=kwargs.get("output_size_decoder",
emb_sz[1]),
nhid=nhid[2],
bidir=False,
dropouth=dropouth[2],
wdrop=wdrop[2],
nlayers=nlayers[2],
cell_type=self.cell_type,
train_init=train_init,
dropoutinit=dropoutinit[2])
self.session_constraint = session_constraint
# allow for changing sizes of decoder output
input_size = decoder_rnn.output_size
nhid = emb_sz[1] if input_size != emb_sz[1] else None
projection_layer = Projection(
output_size=ntoken[0],
input_size=input_size,
nhid=nhid,
dropout=dropoutd,
tie_encoder=decoder_embedding_layer if tie_decoder else None)
self.decoder = Decoder(
decoder_layer=decoder_rnn,
projection_layer=projection_layer,
embedding_layer=decoder_embedding_layer,
pad_token=pad_token,
eos_token=eos_token,
max_tokens=max_tokens,
)
self.decoder_state_linear = nn.Linear(
in_features=self.se_enc.output_size,
out_features=self.decoder.layers[0].output_size)
def __init__(self,
ntoken: int,
emb_sz: HParam,
nhid: HParam,
nlayers: HParam,
pad_token: int,
eos_token: int,
max_tokens: int = 50,
share_embedding_layer: bool = False,
tie_decoder: bool = True,
bidir: bool = False,
**kwargs):
"""
Args:
ntoken (int): Number of tokens for the encoder and the decoder
emb_sz (Union[List[int],int]): Embedding size for the encoder and decoder embeddings
nhid (Union[List[int],int]): Number of hidden dims for the encoder (first two values) and the decoder
nlayers (Union[List[int],int]): Number of layers for the encoder and the decoder
pad_token (int): The index of the token used for padding
eos_token (int): The index of the token used for eos
max_tokens (int): The maximum number of steps the decoder iterates before stopping
share_embedding_layer (bool): if True the decoder shares its input and output embeddings
tie_decoder (bool): if True the encoder and the decoder share their embeddings
bidir (bool): if True use a bidirectional encoder
**kwargs: Extra embeddings that will be passed to the encoder and the decoder
"""
super().__init__()
# allow for the same or different parameters between encoder and decoder
ntoken, emb_sz, nhid, nlayers = get_list(ntoken), get_list(
emb_sz, 2), get_list(nhid, 3), get_list(nlayers, 3)
dropoutd, kwargs = get_kwarg(kwargs,
name="dropoutd",
default_value=0.5)
self.cell_type = "gru"
self.query_encoder = EmbeddingRNNEncoder(ntoken=ntoken[0],
emb_sz=emb_sz[0],
nhid=nhid[0],
nlayers=nlayers[0],
pad_token=pad_token,
bidir=bidir,
out_dim=nhid[0],
cell_type=self.cell_type,
**kwargs)
self.session_encoder = RNNEncoder(in_dim=nhid[0],
nhid=nhid[1],
out_dim=nhid[2],
nlayers=1,
bidir=False,
cell_type=self.cell_type,
**kwargs)
self.decoder = EmbeddingRNNDecoder(
ntoken=ntoken[-1],
emb_sz=emb_sz[-1],
nhid=nhid[-1],
nlayers=nlayers[-1],
pad_token=pad_token,
eos_token=eos_token,
max_tokens=max_tokens,
# Share the embedding layer between encoder and decoder
embedding_layer=self.query_encoder.encoder_with_dropout.embed
if share_embedding_layer else None,
# potentially tie the output projection with the decoder embedding
cell_type=self.cell_type,
out_dim=nhid[-1],
**kwargs)
enc = self.decoder.encoder if tie_decoder else None
self.decoder.projection_layer = Projection(
n_out=ntoken[-1],
n_in=nhid[-1],
nhid=emb_sz[-1],
dropout=dropoutd,
tie_encoder=enc if tie_decoder else None)
self.decoder_state_linear = nn.Linear(
in_features=nhid[-1],
out_features=self.decoder.rnns[0].output_size)
self.nt = ntoken[-1]
def __init__(self,
ntoken: HParam,
emb_sz: HParam,
nhid: HParam,
nlayers: HParam,
pad_token: int,
eos_token: int,
max_tokens: int = 50,
share_embedding_layer: bool = False,
tie_decoder: bool = True,
bidir: bool = False,
**kwargs):
"""
Args:
ntoken (Union[List[int],int]): Number of tokens for the encoder and the decoder
emb_sz (Union[List[int],int]): Embedding size for the encoder and decoder embeddings
nhid (Union[List[int],int]): Number of hidden dims for the encoder and the decoder
nlayers (Union[List[int],int]): Number of layers for the encoder and the decoder
pad_token (int): The index of the token used for padding
eos_token (int): The index of the token used for eos
max_tokens (int): The maximum number of steps the decoder iterates before stopping
share_embedding_layer (bool): if True the decoder shares its input and output embeddings
tie_decoder (bool): if True the encoder and the decoder share their embeddings
bidir (bool): if True use a bidirectional encoder
**kwargs: Extra embeddings that will be passed to the encoder and the decoder
"""
super().__init__()
# allow for the same or different parameters between encoder and decoder
ntoken, emb_sz, nhid, nlayers = get_list(ntoken, 2), get_list(
emb_sz, 2), get_list(nhid, 2), get_list(nlayers, 2)
dropoutd = get_kwarg(kwargs, name="dropout_d",
default_value=0.5) # output dropout
dropoute = get_kwarg(kwargs, name="dropout_e",
default_value=0.1) # encoder embedding dropout
dropoute = get_list(dropoute, 2)
dropouti = get_kwarg(kwargs, name="dropout_i",
default_value=0.65) # input dropout
dropouti = get_list(dropouti, 2)
dropouth = get_kwarg(kwargs, name="dropout_h",
default_value=0.3) # RNN output layers dropout
dropouth = get_list(dropouth, 2)
wdrop = get_kwarg(kwargs, name="wdrop",
default_value=0.5) # RNN weights dropout
wdrop = get_list(wdrop, 2)
self.cell_type = get_kwarg(kwargs,
name="cell_type",
default_value="lstm")
encoder_embedding_layer = DropoutEmbeddings(ntokens=ntoken[0],
emb_size=emb_sz[0],
dropoute=dropoute[0],
dropouti=dropouti[0])
self.bidir = bidir
self.nlayers = nlayers[0]
self.nt = ntoken[-1] # number of possible tokens
self.pr_force = 1.0 # teacher forcing probability
encoder_rnn = RNNLayers(
input_size=emb_sz[0],
output_size=kwargs.get("out_dim", emb_sz[0]),
nhid=nhid[0],
bidir=bidir,
dropouth=dropouth[0],
wdrop=wdrop[0],
nlayers=nlayers[0],
cell_type=self.cell_type,
)
self.encoder = Encoder(embedding_layer=encoder_embedding_layer,
encoder_layer=encoder_rnn)
if share_embedding_layer:
decoder_embedding_layer = encoder_embedding_layer
else:
decoder_embedding_layer = DropoutEmbeddings(ntokens=ntoken[-1],
emb_size=emb_sz[-1],
dropoute=dropoute[1],
dropouti=dropouti[1])
decoder_rnn = RNNLayers(input_size=kwargs.get("input_size",
emb_sz[-1]),
output_size=kwargs.get("output_size",
emb_sz[-1]),
nhid=nhid[-1],
bidir=False,
dropouth=dropouth[1],
wdrop=wdrop[1],
nlayers=nlayers[-1],
cell_type=self.cell_type)
projection_layer = Projection(
output_size=ntoken[-1],
input_size=emb_sz[-1],
dropout=dropoutd,
tie_encoder=decoder_embedding_layer if tie_decoder else None)
self.decoder = Decoder(
decoder_layer=decoder_rnn,
projection_layer=projection_layer,
embedding_layer=decoder_embedding_layer,
pad_token=pad_token,
eos_token=eos_token,
max_tokens=max_tokens,
)
def __init__(self,
ntoken: HParam,
emb_sz: HParam,
nhid: HParam,
nlayers: HParam,
pad_token: int,
eos_token: int,
max_tokens: int = 50,
share_embedding_layer: bool = False,
tie_decoder: bool = True,
bidir: bool = False,
**kwargs):
"""
Args:
ntoken (Union[List[int],int]): Number of tokens for the encoder and the decoder
emb_sz (Union[List[int],int]): Embedding size for the encoder and decoder embeddings
nhid (Union[List[int],int]): Number of hidden dims for the encoder and the decoder
nlayers (Union[List[int],int]): Number of layers for the encoder and the decoder
pad_token (int): The index of the token used for padding
eos_token (int): The index of the token used for eos
max_tokens (int): The maximum number of steps the decoder iterates before stopping
share_embedding_layer (bool): if True the decoder shares its input and output embeddings
tie_decoder (bool): if True the encoder and the decoder share their embeddings
bidir (bool): if True use a bidirectional encoder
**kwargs: Extra embeddings that will be passed to the encoder and the decoder
"""
super(Seq2Seq, self).__init__()
# allow for the same or different parameters between encoder and decoder
ntoken, emb_sz, nhid, nlayers = get_list(ntoken, 2), get_list(
emb_sz, 2), get_list(nhid, 2), get_list(nlayers, 2)
if "dropoutd" in kwargs:
dropoutd = kwargs.pop("dropoutd")
else:
dropoutd = 0.5
self.encoder = EmbeddingRNNEncoder(ntoken=ntoken[0],
emb_sz=emb_sz[0],
nhid=nhid[0],
nlayers=nlayers[0],
pad_token=pad_token,
bidir=bidir,
**kwargs)
self.decoder = EmbeddingRNNDecoder(
ntoken=ntoken[-1],
emb_sz=emb_sz[-1],
nhid=nhid[-1],
nlayers=nlayers[-1],
pad_token=pad_token,
eos_token=eos_token,
max_tokens=max_tokens,
# Share the embedding layer between encoder and decoder
embedding_layer=self.encoder.encoder_with_dropout.embed
if share_embedding_layer else None,
# potentially tie the output projection with the decoder embedding
**kwargs)
enc = self.decoder.encoder if tie_decoder else None
self.decoder.projection_layer = Projection(
n_out=ntoken[-1],
n_in=emb_sz[-1],
dropout=dropoutd,
tie_encoder=enc if tie_decoder else None)
self.nt = ntoken[-1]