def __init__(self, cl_logits_input_dim=None):
self.layers = {}
self.initialize_vocab()
self.layers['embedding'] = layers_lib.Embedding(
self.vocab_size,
FLAGS.embedding_dims,
FLAGS.normalize_embeddings,
self.vocab_freqs,
FLAGS.keep_prob_emb,
name='embedding')
self.layers['embedding_1'] = layers_lib.Embedding(
self.vocab_size,
FLAGS.embedding_dims,
FLAGS.normalize_embeddings,
self.vocab_freqs,
FLAGS.keep_prob_emb,
name='embedding_1')
self.layers['cnn'] = layers_lib.CNN(FLAGS.embedding_dims,
FLAGS.keep_prob_emb)
self.layers['lstm_1'] = layers_lib.BiLSTM(FLAGS.rnn_cell_size,
FLAGS.rnn_num_layers,
name="Bilstm")
action_type = 5 if FLAGS.action == 'all' else 4
self.layers['action_select'] = layers_lib.Actionselect(
action_type, FLAGS.keep_prob_dense, name='action_output')
self.layers['cl_logits'] = layers_lib.Project_layer(
FLAGS.num_classes, FLAGS.keep_prob_dense, name='project_layer')
def _init_embed(self):
self.word_embed_matrix = tf.Variable(initial_value=self.word_embed_pl,
trainable=False)
self.pos_embed_matrix = tf.Variable(initial_value=np.random.normal(
0, 0.1, (2 * self.args.sentence_length + 1, self.args.pos_dims)),
trainable=True,
dtype=tf.float32)
self.layers = {}
self.layers['BiLSTM'] = layers_lib.BiLSTM(self.args.rnn_size)
self.layers['att_weights'] = {
'h1':
tf.Variable(
tf.truncated_normal([self.args.rnn_size, 1], stddev=0.01)),
}
self.input_data_embed = tf.nn.embedding_lookup(self.word_embed_matrix,
self.input_data)
self.input_sents_embed_un = tf.nn.embedding_lookup(
self.word_embed_matrix, self.input_sents)
self.input_pos_embed = tf.nn.embedding_lookup(self.word_embed_matrix,
self.input_sents)
self.input_sents_embed = tf.concat(
[self.input_sents_embed_un, self.input_pos_embed], 2)
def _init_embed(self):
self.char_embed = tf.Variable(initial_value=tf.truncated_normal(
[
len(self.model_args.characterEmbed.character2id),
self.args.char_rnn_size
],
stddev=0.01),
trainable=True,
name="char_embeddings",
dtype=tf.float32)
self.word_embed_matrix = tf.Variable(initial_value=self.word_embed_pl,
trainable=False)
self.layers = {}
self.layers['BiLSTM'] = layers_lib.BiLSTM(self.args.rnn_size)
self.layers['LSTM'] = layers_lib.LSTM(
self.args.char_rnn_size) #随便设置的char的
self.label_embedding = tf.get_variable('label_embeddings',
initializer=tf.truncated_normal(
[500, self.args.class_size],
stddev=0.01),
trainable=True)
self.input_data_embed = tf.nn.embedding_lookup(self.word_embed_matrix,
self.input_data)
def _init_embed(self):
self.word_embed_matrix = tf.Variable(initial_value=self.word_embed_pl,
trainable=False)
self.layers = {}
self.layers['BiLSTM'] = layers_lib.BiLSTM(self.args.rnn_size)
self.feature_dims = self.args.attention_hidden_size
self.layers['att_weights'] = {
'h1':
tf.Variable(
tf.truncated_normal(
[2 * self.args.rnn_size, self.args.attention_hidden_size],
stddev=0.01)),
'h2':
tf.Variable(
tf.truncated_normal([self.args.attention_hidden_size, 1],
stddev=0.01)),
}
self.input_data_embed = tf.nn.embedding_lookup(self.word_embed_matrix,
self.input_data)
#time:2019/1/21
self.rel_type_embedding = tf.Variable(
initial_value=tf.truncated_normal(
[self.args.class_size, self.args.type_dim], stddev=0.01),
trainable=True,
name="rel_type_embedding",
dtype=tf.float32)
self.img_type_embedding = tf.Variable(
initial_value=tf.truncated_normal(
[self.args.class_size, self.args.type_dim], stddev=0.01),
trainable=True,
name="img_type_embedding",
dtype=tf.float32)
self.rel_isa_embedding_diag = tf.Variable(
initial_value=tf.truncated_normal([self.args.type_dim],
stddev=0.01),
trainable=True,
name="img_type_embedding",
dtype=tf.float32)
self.rel_isa_embedding = tf.diag(self.rel_isa_embedding_diag)
self.img_isa_embedding_diag = tf.Variable(
initial_value=tf.truncated_normal([self.args.type_dim],
stddev=0.01),
trainable=True,
name="img_isa_embedding_diag",
dtype=tf.float32)
self.img_isa_embedding = tf.diag(self.img_isa_embedding_diag)
def _init_embed(self):
self.word_embed_matrix = tf.Variable(initial_value=self.word_embed_pl,
trainable=False)
self.layers = {}
self.layers['BiLSTM'] = layers_lib.BiLSTM(self.args.rnn_size)
#This place is very important!
self.feature_dims = self.args.attention_hidden_size
self.layers['att_weights'] = {
'h_m':
tf.Variable(
tf.truncated_normal([self.args.word_dim, self.feature_dims],
stddev=0.01)),
'h1':
tf.Variable(
tf.truncated_normal(
[2 * self.args.rnn_size, self.feature_dims], stddev=0.01)),
'h2':
tf.Variable(
tf.truncated_normal([self.feature_dims, 1], stddev=0.01)),
}
self.type_token_embeddings = tf.Variable(
initial_value=tf.truncated_normal(
[self.type_tree.typeToken_num, self.args.type_dim],
stddev=0.01),
trainable=True,
name="type_token_embeddings",
dtype=tf.float32)
self.type_token_embeddings_add_padding = tf.concat([
self.type_token_embeddings,
tf.constant([[0.0] * self.args.type_dim])
], 0)
self.type_path_embedding = tf.nn.embedding_lookup(
self.type_token_embeddings_add_padding, self.type_path)
self.type_path_embedding = tf.reduce_sum(self.type_path_embedding, 1)
self.type_embeddings_norm = tf.nn.l2_normalize(
self.type_path_embedding, -1)
self.type_embeddings_norm_add_padding = tf.concat([
self.type_embeddings_norm,
tf.constant([[0.0] * self.args.type_dim])
], 0)
def _init_embed(self):
self.word_embed_matrix = tf.Variable(initial_value=self.word_embed_pl,
trainable=False)
self.layers = {}
self.layers['BiLSTM'] = layers_lib.BiLSTM(self.args.rnn_size)
self.layers['att_weights'] = {
'h1':
tf.Variable(
tf.truncated_normal(
[2 * self.args.rnn_size, self.args.attention_hidden_size],
stddev=0.01)),
'h2':
tf.Variable(
tf.truncated_normal([self.args.attention_hidden_size, 1],
stddev=0.01)),
}
self.input_data_embed = tf.nn.embedding_lookup(self.word_embed_matrix,
self.input_data)
