def gen_episode(self, memory, q_vec, fact_vecs, i):
def gen_attn(fact_vec, _reuse=tf.AUTO_REUSE):
with tf.variable_scope('attention', reuse=_reuse):
features = [
fact_vec * q_vec, fact_vec * memory,
tf.abs(fact_vec - q_vec),
tf.abs(fact_vec - memory)
]
feature_vec = tf.concat(features, 1)
attention = tf.layers.dense(feature_vec,
args.embed_dim,
tf.tanh,
reuse=_reuse,
name='fc1')
attention = tf.layers.dense(attention,
1,
reuse=_reuse,
name='fc2')
return tf.squeeze(attention, 1)
# Gates (attentions) are activated, if sentence relevant to the question or memory
attns = tf.map_fn(gen_attn, tf.transpose(fact_vecs, [1, 0, 2]))
attns = tf.transpose(attns) # (B, n_fact)
attns = tf.nn.softmax(attns) # (B, n_fact)
attns = tf.expand_dims(attns, -1) # (B, n_fact, 1)
# The relevant facts are summarized in another GRU
reuse = i > 0
with tf.variable_scope('attention_gru', reuse=reuse):
_, episode = tf.nn.dynamic_rnn(
AttentionGRUCell(args.hidden_size, reuse=reuse),
tf.concat([fact_vecs, attns], 2), # (B, n_fact, D+1)
self.placeholders['inputs_len'],
dtype=np.float32)
return episode # (B, D)
def gen_episode(features, memory, q_vec, fact_vecs, is_training):
def gen_attn(fact_vec):
features = [
fact_vec * q_vec, fact_vec * memory,
tf.abs(fact_vec - q_vec),
tf.abs(fact_vec - memory)
]
feature_vec = tf.concat(features, 1)
attention = tf.layers.dense(feature_vec,
args.embed_dim,
tf.tanh,
name='attn_proj_1')
attention = tf.layers.dense(attention, 1, name='attn_proj_2')
return tf.squeeze(attention, 1)
# Gates (attentions) are activated, if sentence relevant to the question or memory
attns = tf.map_fn(gen_attn, tf.transpose(fact_vecs, [1, 0, 2]))
attns = tf.transpose(attns) # (B, n_fact)
attns = tf.nn.softmax(attns) # (B, n_fact)
attns = tf.expand_dims(attns, -1) # (B, n_fact, 1)
# The relevant facts are summarized in another GRU
_, episode = tf.nn.dynamic_rnn(
AttentionGRUCell(args.hidden_size, name='attn_gru'),
tf.concat([fact_vecs, attns], 2), # (B, n_fact, D+1)
features['inputs_len'],
dtype=np.float32)
return episode # (B, D)
def gen_episode(self, memory, q_vec, fact_vecs, i):
# Gates (attentions) are activated if sentence relevant to the question or memory
attentions = [self.gen_attention(memory, q_vec, fact_vec, bool(i) or bool(n_fact))
for n_fact, fact_vec in enumerate(tf.unstack(fact_vecs, axis=1))] # n_fact list of (B,)
attentions = tf.transpose(tf.stack(attentions)) # (B, n_fact)
attentions = tf.nn.softmax(attentions) # (B, n_fact)
attentions = tf.expand_dims(attentions, -1) # (B, n_fact, 1)
# The relevant facts are summarized in another GRU
reuse = i > 0
with tf.variable_scope('attention_gru', reuse=reuse):
_, episode = tf.nn.dynamic_rnn(
AttentionGRUCell(args.hidden_size, reuse=reuse),
tf.concat([fact_vecs, attentions], 2), # (B, n_fact, D+1)
self.placeholders['inputs_len'],
dtype=np.float32)
return episode # (B, D)