def _build_model(cls, model_config, optim_config, data_config):
"""Build Editor.
Args:
model_config (Config): Editor config
optim_config (Config): optimization config
data_config (Config): dataset config
Returns:
Editor
"""
file_path = join(data.workspace.word_vectors, model_config.wvec_path)
word_embeddings = load_embeddings(file_path, model_config.word_dim,
model_config.vocab_size,
model_config.num_copy_tokens)
word_dim = word_embeddings.embed_dim
source_token_embedder = TokenEmbedder(word_embeddings,
model_config.train_source_embeds)
target_token_embedder = TokenEmbedder(word_embeddings,
model_config.train_target_embeds)
# number of input channels
num_inputs = len(data_config.source_cols)
decoder_cell = AttentionDecoderCell(
target_token_embedder,
2 *
word_dim, # 2 * word_dim because we concat base and copy vectors
model_config.agenda_dim,
model_config.hidden_dim,
model_config.hidden_dim,
model_config.attention_dim,
num_layers=model_config.decoder_layers,
num_inputs=num_inputs,
dropout_prob=model_config.decoder_dropout_prob,
disable_attention=False)
encoder = Encoder(word_dim, model_config.agenda_dim,
model_config.hidden_dim, model_config.encoder_layers,
num_inputs, model_config.encoder_dropout_prob, False)
copy_len = [5, 5, 40]
model = Editor(source_token_embedder,
encoder,
decoder_cell,
copy_lens=copy_len)
model = try_gpu(model)
optimizer = optim.Adam(model.parameters(),
lr=optim_config.learning_rate)
return model, optimizer
def _build_editor(cls, model_config, data_config, word_embeddings,
word_dim, vae_mode):
source_token_embedder = TokenEmbedder(word_embeddings,
model_config.train_source_embeds)
target_token_embedder = TokenEmbedder(word_embeddings,
model_config.train_target_embeds)
# number of input channels
if vae_mode:
num_inputs = len(data_config.source_cols)
else: #edit model uses num_inputs + num_inputs + 1
num_inputs = len(data_config.source_cols) * 2 + 1
decoder_cell = AttentionDecoderCell(
target_token_embedder,
2 * word_dim,
# 2 * word_dim because we concat base and copy vectors
model_config.agenda_dim,
model_config.hidden_dim,
model_config.hidden_dim,
model_config.attention_dim,
num_layers=model_config.decoder_layers,
num_inputs=num_inputs,
dropout_prob=model_config.decoder_dropout_prob,
disable_attention=vae_mode)
if vae_mode:
encoder = Encoder(word_dim, model_config.agenda_dim,
model_config.hidden_dim,
model_config.encoder_layers, num_inputs,
model_config.encoder_dropout_prob, vae_mode,
model_config.vae_kappa)
else:
encoder = Encoder(word_dim, model_config.agenda_dim,
model_config.hidden_dim,
model_config.encoder_layers, num_inputs,
model_config.encoder_dropout_prob, vae_mode)
vae_copy_len = [5, 10, 185]
editor_copy_len = [5, 10, 10, 5, 10, 10, 150]
if vae_mode:
model = Editor(source_token_embedder, encoder, decoder_cell,
vae_copy_len)
else:
model = Editor(source_token_embedder, encoder, decoder_cell,
editor_copy_len)
model = try_gpu(model)
return model
def _build_model(cls, model_config, optim_config, data_config):
"""Build Editor.
Args:
model_config (Config): Editor config
optim_config (Config): optimization config
data_config (Config): dataset config
Returns:
Editor
"""
file_path = join(data.workspace.word_vectors, model_config.wvec_path)
word_embeddings = load_embeddings(file_path, model_config.word_dim,
model_config.vocab_size,
model_config.num_copy_tokens)
word_dim = word_embeddings.embed_dim
edit_model = cls._build_editor(model_config,
data_config,
word_embeddings,
word_dim,
vae_mode=False)
#VAEretreiver
vocab_dict = word_embeddings.vocab._word2index
encoder = Encoder(word_dim,
model_config.agenda_dim,
model_config.hidden_dim,
model_config.encoder_layers,
len(data_config.source_cols),
model_config.encoder_dropout_prob,
use_vae=True,
kappa=model_config.vae_kappa,
use_target=False)
source_token_embedder = TokenEmbedder(word_embeddings,
model_config.train_source_embeds)
target_token_embedder = TokenEmbedder(word_embeddings,
model_config.train_target_embeds)
ret_copy_len = [5, 10, 165]
num_inputs = len(data_config.source_cols)
decoder_cell = AttentionDecoderCell(
target_token_embedder,
2 * word_dim,
# 2 * word_dim because we concat base and copy vectors
model_config.agenda_dim,
model_config.hidden_dim,
model_config.hidden_dim,
model_config.attention_dim,
num_layers=model_config.decoder_layers,
num_inputs=num_inputs,
dropout_prob=model_config.decoder_dropout_prob,
disable_attention=True)
vae_model = VAERetriever(source_token_embedder, encoder, decoder_cell,
ret_copy_len)
ret_model = vae_model
vae_ret_model = EditRetriever(vae_model, ret_model, edit_model)
vae_ret_model = try_gpu(vae_ret_model)
optimizer = optim.Adam(vae_ret_model.parameters(),
lr=optim_config.learning_rate)
#optimizer = optim.SGD(vae_ret_model.parameters(), lr=optim_config.learning_rate)
return vae_ret_model, optimizer