def build_model(self, embedding, **kwargs) -> tf.keras.Model:
embed = build_embedding(embedding, self.transform.word_vocab,
self.transform)
self.transform.map_x = embed.dtype != tf.string
model = CNNTaggingModel(num_tags=len(self.transform.tag_vocab),
embed=embed,
**kwargs)
# model.build((None, None, 3))
return model
def build_model(self, word_embed, ngram_embed, window_size, weight_norm,
filters, kernel_size, dropout_embed, dropout_hidden,
**kwargs) -> tf.keras.Model:
word_vocab, ngram_vocab, tag_vocab = self.transform.word_vocab, self.transform.ngram_vocab, \
self.transform.tag_vocab
word_embed = build_embedding(word_embed, word_vocab, self.transform)
if 'map_x' in self.config:
self.config.map_word_feature = self.config.map_x
del self.config.map_x
else:
self.config.map_word_feature = True
if window_size:
ngram_embed = build_embedding(ngram_embed, ngram_vocab,
self.transform)
else:
ngram_embed = None
model = NgramConvTaggingModel(word_embed, ngram_embed, filters,
kernel_size, dropout_embed,
dropout_hidden, weight_norm,
len(tag_vocab))
return model
def build_model(self, embeddings, embedding_trainable, rnn_input_dropout, rnn_output_dropout, rnn_units,
loss,
**kwargs) -> tf.keras.Model:
model = tf.keras.Sequential()
embeddings = build_embedding(embeddings, self.transform.word_vocab, self.transform)
model.add(embeddings)
if rnn_input_dropout:
model.add(tf.keras.layers.Dropout(rnn_input_dropout, name='rnn_input_dropout'))
model.add(
tf.keras.layers.Bidirectional(tf.keras.layers.LSTM(units=rnn_units, return_sequences=True), name='bilstm'))
if rnn_output_dropout:
model.add(tf.keras.layers.Dropout(rnn_output_dropout, name='rnn_output_dropout'))
model.add(tf.keras.layers.TimeDistributed(tf.keras.layers.Dense(len(self.transform.tag_vocab)), name='dense'))
return model
def build_model(self, pretrained_embed, n_embed, training, **kwargs) -> tf.keras.Model:
if training:
self.config.n_words = len(self.transform.form_vocab)
else:
self.config.lstm_dropout = 0. # keras will use cuda lstm when config.lstm_dropout is 0
self.config.n_feats = len(self.transform.cpos_vocab)
self.config.n_rels = len(self.transform.rel_vocab)
self.config.pad_index = self.transform.form_vocab.pad_idx
self.config.unk_index = self.transform.form_vocab.unk_idx
self.config.bos_index = 2
pretrained: tf.keras.layers.Embedding = build_embedding(pretrained_embed, self.transform.form_vocab,
self.transform) if pretrained_embed else None
if pretrained_embed:
self.config.n_embed = pretrained.output_dim
model = BiaffineModel(self.config, pretrained)
return model