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
class Seq2SeqAttention(nn.Module):
def __init__(self,
ntoken: HParam,
emb_sz: HParam,
nhid: HParam,
nlayers: HParam,
att_nhid: int,
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
att_nhid (int): Number of hidden dims for the attention Module
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="dropoutd",
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)
cell_type = get_kwarg(kwargs, name="cell_type", default_value="lstm")
self.nlayers = nlayers
self.nhid = nhid
self.emb_sz = emb_sz
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", emb_sz[0]),
nhid=nhid[0],
bidir=bidir,
dropouth=dropouth[0],
wdrop=wdrop[0],
nlayers=nlayers[0],
cell_type=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] * 2),
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=cell_type)
projection_layer = AttentionProjection(
output_size=ntoken[-1],
input_size=emb_sz[-1],
dropout=dropoutd,
att_nhid=att_nhid,
tie_encoder=decoder_embedding_layer if tie_decoder else None)
self.decoder = AttentionDecoder(
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 forward(self, *inputs, num_beams=0):
with torch.set_grad_enabled(self.training):
encoder_inputs, decoder_inputs = inputs
# reset the states for the new batch
bs = encoder_inputs.size(1)
self.encoder.reset(bs)
self.decoder.reset(bs)
outputs = self.encoder(encoder_inputs)
# as initial state we use the initial decoder state (zeros)
state = self.decoder.hidden
assert_dims(
outputs,
[self.nlayers[0], None, bs, (self.nhid[0], self.emb_sz[0])])
# pass the encoder outputs as keys to the attention projection_layer
self.decoder.projection_layer.reset(keys=outputs[-1])
if self.training:
self.decoder.pr_force = self.pr_force
nb = 1 if self.pr_force < 1 else 0
else:
nb = num_beams
outputs_dec = self.decoder(decoder_inputs,
hidden=state,
num_beams=nb)
predictions = outputs_dec[:decoder_inputs.size(
0)] if num_beams == 0 else self.decoder.beam_outputs
return predictions, [*outputs, *outputs_dec]
class HREDAttention(nn.Module):
"""Basic HRED model
paper: A Hierarchical Latent Variable Encoder-Decoder Model for Generating Dialogues. Iulian Vlad Serban et al. 2016a.
github: https://github.com/julianser/hed-dlg-truncated
arxiv: http://arxiv.org/abs/1605.06069
"""
BPTT_MAX_UTTERANCES = 1000
def __init__(self,
ntoken: int,
emb_sz: HParam,
nhid: HParam,
nlayers: HParam,
att_nhid: int,
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
att_nhid (int): Number of hidden dims for the attention Module
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 = get_kwarg(kwargs, name="dropoutd",
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)
self.cell_type = "gru"
self.nt = ntoken[-1]
self.pr_force = 1.0
self.nlayers = nlayers
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,
)
self.query_encoder = Encoder(embedding_layer=encoder_embedding_layer,
encoder_layer=encoder_rnn)
self.session_encoder = RNNLayers(
input_size=encoder_rnn.output_size,
nhid=nhid[1],
output_size=kwargs.get("output_size", emb_sz[0]),
nlayers=1,
bidir=False,
cell_type=self.cell_type,
wdrop=wdrop[1],
dropouth=dropouth[1],
)
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] * 2),
output_size=kwargs.get("output_size",
emb_sz[-1]),
nhid=nhid[-1],
bidir=False,
dropouth=dropouth[2],
wdrop=wdrop[2],
nlayers=nlayers[-1],
cell_type=self.cell_type)
projection_layer = AttentionProjection(
output_size=ntoken[-1],
input_size=emb_sz[-1],
dropout=dropoutd,
att_nhid=att_nhid,
att_type="SDP",
tie_encoder=decoder_embedding_layer if tie_decoder else None)
self.decoder = AttentionDecoder(
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 forward(self, *inputs, num_beams=0):
encoder_inputs, decoder_inputs = assert_dims(
inputs, [2, None, None]) # dims: [sl, bs] for encoder and decoder
# reset the states for the new batch
bs = encoder_inputs.size(2)
self.session_encoder.reset(bs)
self.decoder.reset(bs)
query_encoder_outputs = []
outputs = []
num_utterances, max_sl, *_ = encoder_inputs.size()
for index, context in enumerate(encoder_inputs):
self.query_encoder.reset(bs)
outputs = self.query_encoder(context) # context has size [sl, bs]
# BPTT if the dialogue is too long repackage the first half of the outputs to decrease
# the gradient backpropagation and fit it into memory
# to test before adding back
out = repackage_var(outputs[-1][
-1]) if max_sl * num_utterances > self.BPTT_MAX_UTTERANCES and index <= num_utterances // 2 else \
outputs[-1][-1]
query_encoder_outputs.append(
out) # get the last sl output of the query_encoder
query_encoder_outputs = torch.stack(query_encoder_outputs,
dim=0) # [cl, bs, nhid]
session_outputs = self.session_encoder(query_encoder_outputs)
self.decoder.projection_layer.reset(keys=session_outputs[-1])
if self.training:
self.decoder.pr_force = self.pr_force
nb = 1 if self.pr_force < 1 else 0
else:
nb = num_beams
state = self.decoder.hidden
outputs_dec = self.decoder(decoder_inputs, hidden=state, num_beams=nb)
predictions = outputs_dec[-1][:decoder_inputs.size(
0)] if num_beams == 0 else self.decoder.beam_outputs
return predictions, [*outputs, *outputs_dec]