def __init__(self,
encoder_path,
word_index,
index_word,
encoder_trainable=False,
hidden_size=768,
dropout=.25,
n_layers=1,
**kwargs):
super(ClassificationModel, self).__init__(**kwargs)
self.encoder_trainable = encoder_trainable
self.to_token = ToTokens(word_index)
self.to_tags = ToTags(index_word)
self.encoder_layer = get_encoder(encoder_path, encoder_trainable)
self.masking = tf.keras.layers.Masking()
self.dropout_layer = tf.keras.layers.Dropout(dropout)
if not self.encoder_trainable:
self.rnn_layers = []
for _ in range(n_layers):
rnn = tf.keras.layers.Bidirectional(
tf.keras.layers.LSTM(hidden_size,
return_sequences=False))
self.rnn_layers.append(rnn)
self.norm_layer = tf.keras.layers.LayerNormalization(epsilon=1e-9)
self.project_layer = tf.keras.layers.Dense(len(word_index))
self._set_inputs(
tf.keras.backend.placeholder((None, None), dtype='string'))
def __init__(self,
encoder_path,
word_index,
index_word,
encoder_trainable,
hidden_size,
dropout,
n_layers,
rnn,
n_additional_features,
bidirection=True,
use_crf=True,
**kwargs):
super(TaggerModel, self).__init__(**kwargs)
self.to_token = ToTokens(word_index)
self.to_tags = ToTags(index_word)
self.encoder_layer = get_encoder(encoder_path, encoder_trainable)
self.masking = tf.keras.layers.Masking()
self.dropout_layer = tf.keras.layers.Dropout(dropout)
self.n_additional_features = n_additional_features
self.use_crf = use_crf
self.rnn_layers = []
for _ in range(n_layers):
rnn_layer = rnn(hidden_size, return_sequences=True)
if bidirection:
rnn_layer = tf.keras.layers.Bidirectional(rnn_layer)
self.rnn_layers.append(rnn_layer)
self.project_layer = tf.keras.layers.Dense(len(word_index))
if use_crf:
self.crf_layer = CRF(len(word_index))
self._set_inputs(
tf.keras.backend.placeholder((None, None), dtype='string'))
def __init__(self,
encoder_path,
word_index,
index_word,
encoder_trainable=False,
hidden_size=768,
dropout=.25,
n_layers=1,
**kwargs):
super(TaggerPointerModel, self).__init__(**kwargs)
self.to_token = ToPointerTokens(word_index)
self.to_tags = ToPointerTags(index_word)
self.encoder_layer = get_encoder(encoder_path, encoder_trainable)
self.masking = tf.keras.layers.Masking()
self.dropout_layer = tf.keras.layers.Dropout(dropout)
self.rnn_layers = []
for _ in range(n_layers):
rnn = tf.keras.layers.Bidirectional(
tf.keras.layers.LSTM(hidden_size, return_sequences=True))
self.rnn_layers.append(rnn)
b_keys = sorted([x for x in word_index.keys() if x.startswith('B')])
i_keys = sorted([x for x in word_index.keys() if x.startswith('I')])
assert len(b_keys) == len(i_keys)
self.project_hidden_layers = []
self.norm_layers = []
# 2 layers hidden
for _ in range(2):
layers = []
norm_layers = []
for _ in range(len(b_keys)):
layers.append(
tf.keras.layers.Dense(hidden_size, activation='tanh'))
norm_layers.append(
tf.keras.layers.LayerNormalization(epsilon=1e-9))
self.project_hidden_layers.append(layers)
self.norm_layers.append(norm_layers)
self.project_begin_layers = []
for _ in range(len(b_keys)):
self.project_begin_layers.append(tf.keras.layers.Dense(2))
self.project_end_layers = []
for _ in range(len(i_keys)):
self.project_end_layers.append(tf.keras.layers.Dense(2))
self._set_inputs(
tf.keras.backend.placeholder((None, None), dtype='string'))
def __init__(self,
encoder_path,
tag0_size,
tag1_size,
proj0_size,
proj1_size,
word_index,
index_word,
pos_word_index,
pos_index_word,
encoder_trainable=False,
rnn=tf.keras.layers.LSTM,
bidirection=True,
dropout=0.33,
n_layers=3,
hidden_size=400,
**kwargs):
super(ParserModel, self).__init__(**kwargs)
self.encoder_layer = get_encoder(encoder_path,
encoder_trainable=encoder_trainable)
self.masking = tf.keras.layers.Masking()
self.dropout_layer = tf.keras.layers.Dropout(dropout)
self.rnn_layers0 = []
for _ in range(n_layers):
rnn_layer = rnn(hidden_size, return_sequences=True)
if bidirection:
rnn_layer = tf.keras.layers.Bidirectional(rnn_layer)
self.rnn_layers0.append(rnn_layer)
self.rnn_layers1 = []
for _ in range(n_layers):
rnn_layer = rnn(hidden_size, return_sequences=True)
if bidirection:
rnn_layer = tf.keras.layers.Bidirectional(rnn_layer)
self.rnn_layers1.append(rnn_layer)
self.dep_mlp = tf.keras.layers.Dense(hidden_size)
self.dep_concat = tf.keras.layers.Concatenate(axis=-1)
self.dep_proj = tf.keras.layers.Dense(hidden_size)
self.ph0 = tf.keras.layers.Dense(proj0_size)
self.pd0 = tf.keras.layers.Dense(proj0_size)
self.ph1 = tf.keras.layers.Dense(proj1_size)
self.pd1 = tf.keras.layers.Dense(proj1_size)
self.b0 = Biaffine(proj0_size, proj0_size, tag0_size)
self.b1 = Biaffine(proj1_size, proj1_size, tag1_size)
self.to_tokens = ToTokens(word_index)
self.pos_to_tokens = PosToTokens(pos_word_index)
self.to_tags = ToTags(index_word)
self.pos_to_tags = PosToTags(pos_index_word)
self.pos_mlp = tf.keras.layers.Dense(hidden_size)
self.pos_proj = tf.keras.layers.Dense(len(pos_word_index))
self.leaky_relu = tf.keras.layers.LeakyReLU(alpha=0.3)
self._set_inputs(
tf.keras.backend.placeholder((None, None), dtype='string'))