def build_model(cls, args, src_dict, tgt_dict):
base_architecture(args)
if args.encoder_embed_path:
encoder_pretrained_embedding = utils.load_embedding(
args.encoder_embed_path, src_dict)
if args.decoder_embed_path:
decoder_pretrained_embedding = utils.load_embedding(
args.decoder_embed_path, tgt_dict)
encoder = LSTMEncoder(
dictionary=src_dict,
embed_dim=args.encoder_embed_dim,
hidden_size=args.encoder_hidden_size,
num_layers=args.encoder_num_layers,
bidirectional=args.encoder_bidirectional,
dropout_in=args.encoder_dropout_in,
dropout_out=args.encoder_dropout_out,
pretrained_embedding=encoder_pretrained_embedding
if args.encoder_embed_path else None,
)
decoder = LSTMDecoder(
dictionary=tgt_dict,
embed_dim=args.decoder_embed_dim,
hidden_size=args.decoder_hidden_size,
num_layers=args.decoder_num_layers,
dropout_in=args.decoder_dropout_in,
dropout_out=args.decoder_dropout_out,
pretrained_embedding=decoder_pretrained_embedding
if args.decoder_embed_path else None,
use_attention=bool(eval(args.decoder_use_attention)),
)
return cls(encoder, decoder)
def build_embedding(dictionary, embed_dim, path=None):
embedding = nn.Embedding(len(dictionary),
embed_dim,
padding_idx=dictionary.pad_idx)
if path is not None:
utils.load_embedding(path, dictionary, embedding)
return embedding
def build_model(cls, args, src_dict, tgt_dict):
""" Constructs the model. """
base_architecture(args)
encoder_pretrained_embedding = None
decoder_pretrained_embedding = None
# Load pre-trained embeddings, if desired
if args.encoder_embed_path:
encoder_pretrained_embedding = utils.load_embedding(args.encoder_embed_path, src_dict)
if args.decoder_embed_path:
decoder_pretrained_embedding = utils.load_embedding(args.decoder_embed_path, tgt_dict)
# Construct the encoder
encoder = LSTMEncoder(dictionary=src_dict,
embed_dim=args.encoder_embed_dim,
hidden_size=args.encoder_hidden_size,
num_layers=args.encoder_num_layers,
bidirectional=bool(args.encoder_bidirectional),
dropout_in=args.encoder_dropout_in,
dropout_out=args.encoder_dropout_out,
pretrained_embedding=encoder_pretrained_embedding)
# Construct the decoder
decoder = LSTMDecoder(dictionary=tgt_dict,
embed_dim=args.decoder_embed_dim,
hidden_size=args.decoder_hidden_size,
num_layers=args.decoder_num_layers,
dropout_in=args.decoder_dropout_in,
dropout_out=args.decoder_dropout_out,
pretrained_embedding=decoder_pretrained_embedding,
use_attention=bool(eval(args.decoder_use_attention)),
use_lexical_model=bool(eval(args.decoder_use_lexical_model)))
return cls(encoder, decoder)
def build_embedding(dictionary, embed_dim, path=None):
embedding = nn.Embedding(len(dictionary),
embed_dim,
padding_idx=dictionary.pad_idx)
nn.init.kaiming_normal_(embedding.weight)
nn.init.constant_(embedding.weight[dictionary.pad_idx], 0)
if path is not None:
utils.load_embedding(path, dictionary, embedding)
return embedding
def build_model(cls, args, src_dict, tgt_dict):
""" Constructs the model. """
base_architecture(args)
encoder_pretrained_embedding = None
decoder_pretrained_embedding = None
# select the device to run the model on
use_cuda = torch.cuda.is_available() and args.device == 'cuda'
device = torch.device("cuda" if use_cuda else "cpu")
# Load pre-trained embeddings, if desired
if args.encoder_embed_path:
encoder_pretrained_embedding = utils.load_embedding(
args.encoder_embed_path, src_dict)
if args.decoder_embed_path:
decoder_pretrained_embedding = utils.load_embedding(
args.decoder_embed_path, tgt_dict)
# Construct the encoder
encoder = LSTMEncoder(
dictionary=src_dict,
device=device,
embed_dim=args.encoder_embed_dim,
hidden_size=args.encoder_hidden_size,
num_layers=args.encoder_num_layers,
bidirectional=args.encoder_bidirectional,
dropout_in=args.encoder_dropout_in,
dropout_out=args.encoder_dropout_out,
pretrained_embedding=encoder_pretrained_embedding)
# Construct the decoder
decoder = LSTMDecoder(
dictionary=tgt_dict,
device=device,
embed_dim=args.decoder_embed_dim,
hidden_size=args.decoder_hidden_size,
num_layers=args.decoder_num_layers,
dropout_in=args.decoder_dropout_in,
dropout_out=args.decoder_dropout_out,
pretrained_embedding=decoder_pretrained_embedding,
use_attention=bool(eval(args.decoder_use_attention)),
use_lexical_model=bool(eval(args.decoder_use_lexical_model)))
return cls(encoder, decoder)
