def _preprocess(self,
Xw,
Xw_len,
Xc,
Xc_len,
scope="preprocess_layers",
reuse=False):
with tf.variable_scope(scope, reuse=reuse):
Xw_embedded, size_w = embedded(Xw,
self.embeddings_w[0],
self.embeddings_w[1],
self.config.wv_config["train_w"],
scope="embedded_w")
Xc_embedded, size_c = embedded(Xc,
self.embeddings_c[0],
self.embeddings_c[1],
self.config.wv_config["train_c"],
scope="embedded_c")
batch_size, seq_len = tf.shape(Xw)[0], tf.shape(Xw)[1]
Xc_embedded = tf.reshape(Xc_embedded,
shape=[batch_size * seq_len, -1, size_c])
Xc_len = tf.reshape(Xc_len, shape=[
batch_size * seq_len,
])
Xc_embedded, size_c = bi_gru(Xc_embedded, Xc_len,
(self.config.char_dim, ), 2,
self.initializer, 1.0, "bi_gru_c2w")
Xc_embedded = tf.reshape(Xc_embedded,
shape=[batch_size, seq_len, size_c])
X_embedded = tf.concat([Xw_embedded, Xc_embedded], axis=-1)
out_w, out_w_size = bi_gru(X_embedded, Xw_len,
(self.config.bi_dim, ), 1,
self.initializer, 1.0, "bi_gru__wc")
return out_w, out_w_size
def _encode(self, Xw, Xw_l, Xc, Xc_l, scope="encode_layers", reuse=False):
with tf.variable_scope(scope, reuse=reuse):
Xw_embedded, size_w = embedded(Xw,
self.embeddings_w[0],
self.embeddings_w[1],
self.config.wv_config["train_w"],
scope="embedded_w")
Xc_embedded, size_c = embedded(Xc,
self.embeddings_c[0],
self.embeddings_c[1],
self.config.wv_config["train_c"],
scope="embedded_c")
batch_size = tf.shape(Xw)[0]
# char
v0, v0_size = attention_han(Xc_embedded, self.config.un_dim,
self.initializer, "attention_han_c")
v1, v1_size = bi_gru(Xc_embedded, Xc_l, (self.config.bi_dim, ), 2,
self.initializer, 1.0, "bi_gru_c")
char_v = tf.reshape(tf.concat([v0, v1], axis=-1),
[batch_size, v0_size + v1_size])
# word
v0, v0_size = attention_han(Xw_embedded, self.config.un_dim,
self.initializer, "attention_han_w")
v1, v1_size = bi_gru(Xw_embedded, Xw_l, (self.config.bi_dim, ), 2,
self.initializer, 1.0, "bi_gru_w")
word_v = tf.reshape(tf.concat([v0, v1], axis=-1),
[batch_size, v0_size + v1_size])
# phrase
Xp_embedded, size_p = conv_with_max_pool(Xw_embedded, (2, 3, 4, 5),
size_w // 4, False,
tf.nn.selu,
self.initializer,
"conv_w2p")
v0, v0_size = attention_han(Xp_embedded, self.config.un_dim,
self.initializer, "attention_han_p")
v1, v1_size = bi_gru(Xp_embedded, Xw_l, (self.config.bi_dim, ), 2,
self.initializer, 1.0, "bi_gru_p")
phrase_v = tf.reshape(tf.concat([v0, v1], axis=-1),
[batch_size, v0_size + v1_size])
return char_v, word_v, phrase_v