def _wordvec_embedding(self, inputs):
wordvec = load_pretrained_vec()
embedding = tf.get_variable('embedding',
[wordvec.shape[0], wordvec.shape[1]],
initializer=tf.constant_initializer(
wordvec, tf.float32))
output = tf.nn.embedding_lookup(embedding, inputs)
return output
def __init__(self, **kwargs):
super(Embedding_Layer, self).__init__(**kwargs)
wordvec = load_pretrained_vec()
self.word_embedding = tf.keras.layers.Embedding(
wordvec.shape[0],
wordvec.shape[1],
tf.constant_initializer(wordvec, tf.float32),
name='word_embedding')
def _build_layers(self):
self.p_embedding = tf.keras.layers.Embedding(
Config.model.pos_num,
Config.model.pos_embedding_size,
name='pos_embedding')
self.rel_embedding = tf.keras.layers.Embedding(
Config.model.dep_num * 2,
Config.model.comp_action_embedding_size,
name='rel_embedding')
self.history_a_embedding = tf.keras.layers.Embedding(
3 + Config.model.dep_num * 4,
Config.model.history_action_embedding_size,
name='action_embedding')
wordvec = load_pretrained_vec()
self.w_embedding = tf.keras.layers.Embedding(wordvec.shape[0],
wordvec.shape[1],
tf.constant_initializer(
wordvec, tf.float32),
name='word_embedding')
self.embedding_dense = tf.keras.layers.Dense(
Config.model.embedding_fc_unit, tf.nn.relu, name='embedding_fc')
self.recurse_dense = tf.keras.layers.Dense(
Config.model.embedding_fc_unit, tf.nn.tanh, name='recurse_fc')
lstm_cells = [
tf.nn.rnn_cell.LSTMCell(Config.model.lstm_unit,
initializer=tf.orthogonal_initializer())
for _ in range(Config.model.lstm_layer_num)
]
self.stack_lstm = tf.nn.rnn_cell.MultiRNNCell(lstm_cells)
lstm_cells = [
tf.nn.rnn_cell.LSTMCell(Config.model.lstm_unit,
initializer=tf.orthogonal_initializer())
for _ in range(Config.model.lstm_layer_num)
]
self.buff_lstm = tf.nn.rnn_cell.MultiRNNCell(lstm_cells)
lstm_cells = [
tf.nn.rnn_cell.LSTMCell(Config.model.lstm_unit,
initializer=tf.orthogonal_initializer())
for _ in range(Config.model.lstm_layer_num)
]
self.action_lstm = tf.nn.rnn_cell.MultiRNNCell(lstm_cells)
self.final_dense = tf.keras.layers.Dense(2 + 2 * Config.model.dep_num,
name='softmax_fc')
def __init__(self, **kwargs):
super(Embedding_Layer, self).__init__(**kwargs)
self.dropout = tf.keras.layers.Dropout(
1 - tf.pow(Config.train.dropout_decay, Config.train.epoch))
wordvec = load_pretrained_vec()
self.word_embedding = tf.keras.layers.Embedding(
wordvec.shape[0],
wordvec.shape[1],
tf.constant_initializer(wordvec, tf.float32),
name='word_embedding')
self.pos_embedding = tf.keras.layers.Embedding(
Config.model.pos_num,
Config.model.pos_embedding_size,
name='pos_embedding')
self.conv3d_0 = tf.keras.layers.Conv3D(16, [1, 3, 3],
padding='SAME',
activation=tf.nn.relu,
name='conv3d_0')
self.conv3d_1 = tf.keras.layers.Conv3D(16, [1, 3, 3],
padding='SAME',
activation=tf.nn.relu,
name='conv3d_1')
self.maxpool3d_0 = tf.keras.layers.MaxPool3D([1, 1, 2], [1, 1, 2],
'SAME') # pool width only
self.maxpool3d_1 = tf.keras.layers.MaxPool3D([1, 2, 2], [1, 2, 2],
'SAME')
self.conv3d_2 = tf.keras.layers.Conv3D(32, [1, 3, 3],
padding='SAME',
activation=tf.nn.relu,
name='conv3d_2')
self.conv3d_3 = tf.keras.layers.Conv3D(32, [1, 3, 3],
padding='SAME',
activation=tf.nn.relu,
name='conv3d_3')
self.conv3d_4 = tf.keras.layers.Conv3D(64, [1, 3, 3],
padding='SAME',
activation=tf.nn.relu,
name='conv3d_4')
self.conv3d_5 = tf.keras.layers.Conv3D(64, [1, 3, 3],
padding='SAME',
activation=tf.nn.relu,
name='conv3d_5')
self.cnn_dense = tf.keras.layers.Dense(Config.model.cnn_dense_units,
tf.nn.relu,
name='cnn_dense')
def _embedding(self, word_id, pos_id):
wordvec = load_pretrained_vec()
embedding = tf.get_variable('word_embedding',
[wordvec.shape[0], wordvec.shape[1]],
initializer=tf.constant_initializer(
wordvec, tf.float32))
word_embedding = tf.nn.embedding_lookup(embedding, word_id)
embedding = tf.get_variable(
'pos_embedding',
[Config.model.pos_num, Config.model.pos_embedding_size])
pos_embedding = tf.nn.embedding_lookup(embedding, pos_id)
embedding = tf.concat([word_embedding, pos_embedding], -1)
outputs = slim.dropout(embedding,
Config.model.embedding_keep_prob,
is_training=self.is_training)
return outputs
def _build_layers(self):
self.p_embedding = tf.keras.layers.Embedding(
Config.model.pos_num,
Config.model.pos_embedding_size,
name='pos_embedding')
self.history_a_embedding = tf.keras.layers.Embedding(
4 + Config.model.dep_num * 4,
Config.model.history_action_embedding_size,
name='action_embedding')
wordvec = load_pretrained_vec()
self.w_embedding = tf.keras.layers.Embedding(wordvec.shape[0],
wordvec.shape[1],
tf.constant_initializer(
wordvec, tf.float32),
name='word_embedding')
self.learned_word_dense = tf.keras.layers.Dense(
Config.model.embedding_fc_unit, tf.nn.relu, name='learned_word_fc')
self.tree_lstm_cell = TreeLSTMCell(Config.model.lstm_unit)
lstm_cells = [
tf.nn.rnn_cell.LSTMCell(Config.model.lstm_unit,
initializer=tf.orthogonal_initializer())
for _ in range(2)
]
self.stack_lstm = tf.nn.rnn_cell.MultiRNNCell(lstm_cells)
self.fw_lstm_cell0 = tf.nn.rnn_cell.LSTMCell(
Config.model.lstm_unit,
initializer=tf.orthogonal_initializer(),
name='fw_cell0')
self.bw_lstm_cell0 = tf.nn.rnn_cell.LSTMCell(
Config.model.lstm_unit,
initializer=tf.orthogonal_initializer(),
name='bw_cell0')
self.fw_lstm_cell1 = tf.nn.rnn_cell.LSTMCell(
Config.model.lstm_unit,
initializer=tf.orthogonal_initializer(),
name='fw_cell1')
self.bw_lstm_cell1 = tf.nn.rnn_cell.LSTMCell(
Config.model.lstm_unit,
initializer=tf.orthogonal_initializer(),
name='bw_cell1')
lstm_cells = [
tf.nn.rnn_cell.LSTMCell(Config.model.lstm_unit,
initializer=tf.orthogonal_initializer())
for _ in range(2)
]
self.action_lstm = tf.nn.rnn_cell.MultiRNNCell(lstm_cells)
lstm_cells = [
tf.nn.rnn_cell.LSTMCell(Config.model.lstm_unit,
initializer=tf.orthogonal_initializer())
for _ in range(2)
]
self.deque_lstm = tf.nn.rnn_cell.MultiRNNCell(lstm_cells)
self.final_dense = tf.keras.layers.Dense(4 + 4 * Config.model.dep_num,
tf.nn.relu,
name='softmax_fc')
