def build_models(self):
'''
build all models.
in this model source and target share embeddings
'''
self.train_models['embedding'] = natsEmbedding(
vocab_size = self.batch_data['vocab_size'],
emb_dim = self.args.emb_dim,
share_emb_weight = True
).to(self.args.device)
self.train_models['encoder'] = EncoderRNN(
self.args.emb_dim, self.args.src_hidden_dim,
self.args.nLayers, 'lstm',
device = self.args.device
).to(self.args.device)
self.train_models['encoder2decoder'] = natsEncoder2Decoder(
src_hidden_size = self.args.src_hidden_dim,
trg_hidden_size = self.args.trg_hidden_dim,
rnn_network = 'lstm',
device = self.args.device
).to(self.args.device)
self.train_models['decoderRNN'] = torch.nn.LSTMCell(
self.args.emb_dim+self.args.trg_hidden_dim,
self.args.trg_hidden_dim
).to(self.args.device)
self.train_models['attnEncoder'] = AttentionEncoder(
self.args.src_hidden_dim,
self.args.trg_hidden_dim,
attn_method='luong_general',
repetition='temporal'
).to(self.args.device)
self.train_models['attnDecoder'] = AttentionDecoder(
self.args.trg_hidden_dim,
attn_method='luong_general'
).to(self.args.device)
self.train_models['wrapDecoder'] = torch.nn.Linear(
self.args.src_hidden_dim*2+self.args.trg_hidden_dim*2,
self.args.trg_hidden_dim, bias=True
).to(self.args.device)
self.train_models['genPrb'] = torch.nn.Linear(
self.args.emb_dim+self.args.src_hidden_dim*2+self.args.trg_hidden_dim, 1
).to(self.args.device)
# decoder to vocab
self.train_models['decoder2proj'] = torch.nn.Linear(
self.args.trg_hidden_dim, self.args.emb_dim, bias=False
).to(self.args.device)
def build_models(self):
'''
build all models.
in this model source and target share embeddings
'''
self.base_models['embedding_base'] = natsEmbedding(
vocab_size=self.batch_data['vocab_size'],
emb_dim=self.args.emb_dim,
share_emb_weight=self.args.share_emb_weight
).to(self.args.device)
self.base_models['encoder_base'] = natsEncoder(
emb_dim=self.args.emb_dim,
hidden_size=self.args.src_hidden_dim,
rnn_network=self.args.rnn_network,
device=self.args.device
).to(self.args.device)
self.train_models['encoder'] = natsEncoder(
emb_dim=self.args.src_hidden_dim*2,
hidden_size=self.args.src_hidden_dim,
rnn_network=self.args.rnn_network,
device=self.args.device
).to(self.args.device)
self.train_models['encoder2decoder'] = natsEncoder2Decoder(
src_hidden_size=self.args.src_hidden_dim,
trg_hidden_size=self.args.trg_hidden_dim,
rnn_network=self.args.rnn_network
).to(self.args.device)
self.train_models['pgdecoder'] = PointerGeneratorDecoder(
input_size=self.args.emb_dim,
src_hidden_size=self.args.src_hidden_dim,
trg_hidden_size=self.args.trg_hidden_dim,
attn_method=self.args.attn_method,
repetition=self.args.repetition,
pointer_net=self.args.pointer_net,
attn_decoder=self.args.attn_decoder,
rnn_network=self.args.rnn_network,
device=self.args.device
).to(self.args.device)
# decoder to vocab
if self.args.share_emb_weight:
self.train_models['decoder2proj'] = torch.nn.Linear(
self.args.trg_hidden_dim, self.args.emb_dim, bias=False).to(self.args.device)
else:
self.train_models['decoder2vocab'] = torch.nn.Linear(
self.args.trg_hidden_dim, self.batch_data['vocab_size']).to(self.args.device)
