def __init__(self, l1_vocab_size, l2_vocab_size, embedding_size,
hidden_size, rnn_type, n_layers_encoder, n_layers_decoder,
dropout):
super(Model, self).__init__()
self.l1_vocab_size = l1_vocab_size
self.l2_vocab_size = l2_vocab_size
self.embedding_size = embedding_size
self.hidden_size = hidden_size
self.rnn_type = rnn_type
self.n_layers_encoder = n_layers_encoder
self.n_layers_decoder = n_layers_decoder
self.dropout = dropout
# Embed
self.l1_embeddings = FixedEmbedding(l1_vocab_size, embedding_size)
self.l2_embeddings = FixedEmbedding(l2_vocab_size, embedding_size)
# Encode
self.encoder = RNNEncoder(rnn_type=rnn_type,
bidirectional=True,
num_layers=n_layers_encoder,
hidden_size=hidden_size,
dropout=dropout,
l1_embeddings=self.l1_embeddings,
l2_embeddings=self.l2_embeddings)
# Decode
self.l1_decoder = InputFeedRNNDecoder(
rnn_type=rnn_type,
bidirectional_encoder=True,
num_layers=n_layers_decoder,
hidden_size=hidden_size,
attn_type='general',
coverage_attn=None, # Not supported.
context_gate=None,
copy_attn=None, # Not supported.
dropout=dropout,
embeddings=self.l1_embeddings)
self.l2_decoder = InputFeedRNNDecoder(
rnn_type=rnn_type,
bidirectional_encoder=True,
num_layers=n_layers_decoder,
hidden_size=hidden_size,
attn_type='general',
coverage_attn=None, # Not supported.
context_gate=None,
copy_attn=None, # Not supported.
dropout=dropout,
embeddings=self.l2_embeddings)
# Project
self.l1_to_vocab = nn.Linear(hidden_size, l1_vocab_size)
self.l2_to_vocab = nn.Linear(hidden_size, l2_vocab_size)
def make_decoder(opt, embeddings):
"""
Various decoder dispatcher function.
Args:
opt: the option in current environment.
embeddings (Embeddings): vocab embeddings for this decoder.
"""
if opt.decoder_type == "transformer":
return TransformerDecoder(opt.dec_layers, opt.rnn_size,
opt.global_attention, opt.copy_attn,
opt.dropout, embeddings)
elif opt.decoder_type == "cnn":
return CNNDecoder(opt.dec_layers, opt.rnn_size, opt.global_attention,
opt.copy_attn, opt.cnn_kernel_width, opt.dropout,
embeddings)
elif opt.input_feed:
return InputFeedRNNDecoder(opt.rnn_type, opt.brnn, opt.dec_layers,
opt.rnn_size, opt.global_attention,
opt.coverage_attn, opt.context_gate,
opt.copy_attn, opt.dropout, embeddings)
else:
return StdRNNDecoder(opt.rnn_type, opt.brnn, opt.dec_layers,
opt.rnn_size, opt.global_attention,
opt.coverage_attn, opt.context_gate,
opt.copy_attn, opt.dropout, embeddings)
def make_decoder(opt, embeddings):
"""
Various decoder dispatcher function.
Args:
opt: the option in current environment.
embeddings (Embeddings): vocab embeddings for this decoder.
"""
if opt.decoder_type == "transformer":
return TransformerDecoder(opt.dec_layers, opt.rnn_size,
opt.global_attention, opt.copy_attn,
opt.dropout, embeddings)
elif opt.decoder_type == "cnn":
return CNNDecoder(opt.dec_layers, opt.rnn_size, opt.global_attention,
opt.copy_attn, opt.cnn_kernel_width, opt.dropout,
embeddings)
elif opt.input_feed:
return InputFeedRNNDecoder(
opt.rnn_type, opt.brnn, opt.dec_layers, opt.rnn_size,
opt.global_attention, opt.coverage_attn, opt.context_gate,
opt.copy_attn, opt.dropout, embeddings, opt.affective_attention,
opt.affective_attn_strength) # Add affective attention options px
else:
return StdRNNDecoder(
opt.rnn_type, opt.brnn, opt.dec_layers, opt.rnn_size,
opt.global_attention, opt.coverage_attn, opt.context_gate,
opt.copy_attn, opt.dropout, embeddings, opt.affective_attention,
opt.affective_attn_strength,
opt.word_freq if "word_freq" in opt else "",
opt.local_weights if "local_weights" in opt else False)
def make_decoder(opt, embeddings):
"""
Various decoder dispatcher function.
Args:
opt: the option in current environment.
embeddings (Embeddings): vocab embeddings for this decoder.
"""
if opt.decoder_type == "transformer":
return TransformerDecoder(opt.dec_layers, opt.rnn_size,
opt.global_attention, opt.copy_attn,
opt.dropout, embeddings)
elif opt.decoder_type == "cnn":
return CNNDecoder(opt.dec_layers, opt.rnn_size, opt.global_attention,
opt.copy_attn, opt.cnn_kernel_width, opt.dropout,
embeddings)
elif opt.input_feed and opt.inference_network_type == "none":
print("input feed")
return InputFeedRNNDecoder(opt.rnn_type, opt.brnn, opt.dec_layers,
opt.memory_size, opt.decoder_rnn_size,
opt.attention_size, opt.global_attention,
opt.coverage_attn, opt.context_gate,
opt.copy_attn, opt.dropout, embeddings,
opt.reuse_copy_attn)
elif opt.input_feed and opt.inference_network_type != "none":
print("VARIATIONAL DECODER")
scoresFstring = opt.alpha_transformation
scoresF = scoresF_dict[scoresFstring]
return ViRNNDecoder(
opt.rnn_type,
opt.brnn,
opt.dec_layers,
memory_size=opt.memory_size,
hidden_size=opt.decoder_rnn_size,
attn_size=opt.attention_size,
attn_type=opt.global_attention,
coverage_attn=opt.coverage_attn,
context_gate=opt.context_gate,
copy_attn=opt.copy_attn,
dropout=opt.dropout,
embeddings=embeddings,
reuse_copy_attn=opt.reuse_copy_attn,
p_dist_type=opt.p_dist_type,
q_dist_type=opt.q_dist_type,
use_prior=opt.use_generative_model > 0,
scoresF=scoresF,
n_samples=opt.n_samples,
mode=opt.mode,
)
else:
return StdRNNDecoder(opt.rnn_type, opt.brnn, opt.dec_layers,
opt.rnn_size, opt.global_attention,
opt.coverage_attn, opt.context_gate,
opt.copy_attn, opt.dropout, embeddings,
opt.reuse_copy_attn)
def make_decoder(opt, embeddings):
"""
Various decoder dispatcher function.
Args:
opt: the option in current environment.
embeddings (Embeddings): vocab embeddings for this decoder.
"""
if opt.decoder_type == "transformer":
return TransformerDecoder(opt.dec_layers, opt.rnn_size,
opt.global_attention, opt.copy_attn,
opt.dropout, embeddings)
elif opt.decoder_type == "cnn":
return CNNDecoder(opt.dec_layers, opt.rnn_size,
opt.global_attention, opt.copy_attn,
opt.cnn_kernel_width, opt.dropout,
embeddings)
elif opt.model_type == "hierarchical_text":
print("ModelConstructor line:96 hierarchical text decoder")
return HierarchicalInputFeedRNNDecoder(opt.rnn_type, opt.brnn,
opt.dec_layers, opt.rnn_size,
opt.global_attention,
opt.coverage_attn,
opt.context_gate,
opt.copy_attn,
opt.dropout,
embeddings,
opt.reuse_copy_attn,
opt.model_type,
opt.hier_add_word_enc_input )
elif opt.input_feed:
return InputFeedRNNDecoder(opt.rnn_type, opt.brnn,
opt.dec_layers, opt.rnn_size,
opt.global_attention,
opt.coverage_attn,
opt.context_gate,
opt.copy_attn,
opt.dropout,
embeddings,
opt.reuse_copy_attn)
else:
return StdRNNDecoder(opt.rnn_type, opt.brnn,
opt.dec_layers, opt.rnn_size,
opt.global_attention,
opt.coverage_attn,
opt.context_gate,
opt.copy_attn,
opt.dropout,
embeddings,
opt.reuse_copy_attn)
def make_decoder(opt, embeddings, mmod_dcap=False):
"""
Various decoder dispatcher function.
Args:
opt: the option in current environment.
embeddings (Embeddings): vocab embeddings for this decoder.
"""
if opt.decoder_type == "transformer":
if mmod_dcap:
return multimodal.CapsuleTransformerDecoder(
opt.dec_layers, opt.rnn_size,
opt.global_attention, opt.copy_attn,
opt.dropout, embeddings,
opt.num_iterations, opt.num_capsules, opt.num_regions
)
else:
return TransformerDecoder(opt.dec_layers, opt.rnn_size,
opt.global_attention, opt.copy_attn,
opt.dropout, embeddings)
elif opt.decoder_type == "cnn":
return CNNDecoder(opt.dec_layers, opt.rnn_size,
opt.global_attention, opt.copy_attn,
opt.cnn_kernel_width, opt.dropout,
embeddings)
elif opt.input_feed:
return InputFeedRNNDecoder(opt.rnn_type, opt.brnn,
opt.dec_layers, opt.rnn_size,
opt.global_attention,
opt.coverage_attn,
opt.context_gate,
opt.copy_attn,
opt.dropout,
embeddings,
opt.reuse_copy_attn)
else:
return StdRNNDecoder(opt.rnn_type, opt.brnn,
opt.dec_layers, opt.rnn_size,
opt.global_attention,
opt.coverage_attn,
opt.context_gate,
opt.copy_attn,
opt.dropout,
embeddings,
opt.reuse_copy_attn)
def make_decoder(opt, embeddings):
"""
Various decoder dispatcher function.
Args:
opt: the option in current environment.
embeddings (Embeddings): vocab embeddings for this decoder.
"""
if opt.decoder_type == "transformer":
return TransformerDecoder(opt.dec_layers, opt.rnn_size,
opt.global_attention, opt.copy_attn,
opt.dropout, embeddings)
elif opt.decoder_type == "cnn":
return CNNDecoder(opt.dec_layers, opt.rnn_size, opt.global_attention,
opt.copy_attn, opt.cnn_kernel_width, opt.dropout,
embeddings)
elif opt.input_feed:
print("input feed")
if opt.inference_network_type == 'none':
return InputFeedRNNDecoder(opt.rnn_type, opt.brnn, opt.dec_layers,
opt.rnn_size, opt.global_attention,
opt.coverage_attn, opt.context_gate,
opt.copy_attn, opt.dropout, embeddings,
opt.reuse_copy_attn, opt.dist_type)
else:
return ViInputFeedRNNDecoder(opt.rnn_type,
opt.brnn,
opt.dec_layers,
opt.rnn_size,
opt.global_attention,
opt.coverage_attn,
opt.context_gate,
opt.copy_attn,
opt.dropout,
embeddings,
opt.reuse_copy_attn,
dist_type=opt.dist_type,
normalization=opt.prior_normalization)
else:
return StdRNNDecoder(opt.rnn_type, opt.brnn, opt.dec_layers,
opt.rnn_size, opt.global_attention,
opt.coverage_attn, opt.context_gate,
opt.copy_attn, opt.dropout, embeddings,
opt.reuse_copy_attn, opt.dist_type)
def make_decoder(opt, embeddings, stage1, basic_enc_dec):
"""
Various decoder dispatcher function.
Args:
opt: the option in current environment.
embeddings (Embeddings): vocab embeddings for this decoder.
stage1: stage1 decoder
"""
return InputFeedRNNDecoder(opt.rnn_type,
opt.brnn2,
opt.dec_layers2,
opt.rnn_size,
opt.global_attention,
opt.coverage_attn,
opt.context_gate,
True,
opt.dropout,
embeddings,
opt.reuse_copy_attn,
hier_attn=True)
def make_decoder(opt, embeddings):
if opt.decoder_type == "transformer":
return TransformerDecoder(opt.dec_layers, opt.rnn_size,
opt.global_attention, opt.copy_attn,
opt.dropout, embeddings)
elif opt.decoder_type == "cnn":
return CNNDecoder(opt.dec_layers, opt.rnn_size, opt.global_attention,
opt.copy_attn, opt.cnn_kernel_width, opt.dropout,
embeddings)
elif opt.input_feed:
return InputFeedRNNDecoder(opt.rnn_type, opt.brnn, opt.dec_layers,
opt.rnn_size, opt.global_attention,
opt.coverage_attn, opt.context_gate,
opt.copy_attn, opt.dropout, embeddings,
opt.reuse_copy_attn)
else:
return StdRNNDecoder(opt.rnn_type, opt.brnn, opt.dec_layers,
opt.rnn_size, opt.global_attention,
opt.coverage_attn, opt.context_gate,
opt.copy_attn, opt.dropout, embeddings,
opt.reuse_copy_attn)
def make_decoder(opt, embeddings):
"""
Various decoder dispatcher function.
Args:
opt: the option in current environment.
embeddings (Embeddings): vocab embeddings for this decoder.
"""
if opt.decoder_type == "transformer":
return TransformerDecoder(opt.dec_layers, opt.rnn_size,
opt.global_attention, opt.copy_attn,
opt.dropout, embeddings)
elif opt.decoder_type == "cnn":
return CNNDecoder(opt.dec_layers, opt.rnn_size, opt.global_attention,
opt.copy_attn, opt.cnn_kernel_width, opt.dropout,
embeddings)
# NOTE: THIS IS WHAT GETS TRIGGERED IN DEFAULT EXPERIMENTS
elif opt.decoder_type == "charrnn":
print(f"opt.rnn_type={opt.rnn_type}")
print(f"opt.brnn={opt.brnn}")
print(f"opt.dec_lay={opt.dec_layers}")
print(f"opt.rnn_size={opt.rnn_size}")
print(f"opt.global_attention={opt.global_attention}")
print(f"opt.coverage_attn={opt.coverage_attn}")
print(f"opt.context_gate={opt.context_gate}")
print(f"opt.copy_attn={opt.copy_attn}")
print(f"opt.dropout={opt.dropout}")
print(f"embeddings={embeddings}")
print(f"opt.reuse_copy_attn={opt.reuse_copy_attn}")
"""
rnn_type,
bidirectional_encoder,
num_layers,
hidden_size,
attn_type="general",
coverage_attn=False,
context_gate=None,
copy_attn=False,
dropout=0.0,
embeddings=None,
reuse_copy_attn=False):
"""
return [
StdWordRNNDecoder(opt.rnn_type, opt.brnn, opt.dec_layers,
opt.rnn_size, opt.global_attention,
opt.coverage_attn, opt.context_gate,
opt.copy_attn, opt.dropout, embeddings,
opt.reuse_copy_attn),
StdCharRNNDecoder(opt.rnn_type, opt.brnn, opt.dec_layers,
opt.rnn_size, opt.global_attention,
opt.coverage_attn, opt.context_gate,
opt.copy_attn, opt.dropout, embeddings,
opt.reuse_copy_attn)
]
elif opt.input_feed:
return InputFeedRNNDecoder(opt.rnn_type, opt.brnn, opt.dec_layers,
opt.rnn_size, opt.global_attention,
opt.coverage_attn, opt.context_gate,
opt.copy_attn, opt.dropout, embeddings,
opt.reuse_copy_attn)
else:
return StdRNNDecoder(opt.rnn_type, opt.brnn, opt.dec_layers,
opt.rnn_size, opt.global_attention,
opt.coverage_attn, opt.context_gate,
opt.copy_attn, opt.dropout, embeddings,
opt.reuse_copy_attn)
