def generate_episode(self, memory, q_vec, fact_vecs, speaker_info, hop_index):
"""Generate episode by applying attention to current fact vectors through a modified GRU"""
attentions = [tf.squeeze(
self.get_attention(q_vec, memory, fv, speaker_info, bool(hop_index) or bool(i)), axis=1) # reuse is false only for 0th hop, and first fv.
for i, fv in enumerate(tf.unstack(fact_vecs, axis=1))]
attentions = tf.transpose(tf.stack(attentions))
self.attentions.append(attentions)
attentions = tf.nn.softmax(attentions)
attentions = tf.expand_dims(attentions, axis=-1)
# print("hello")
# print(fact_vecs, attentions)
# print("done")
if self.config.GRU == "true" :
reuse = True if hop_index > 0 else False
if self.config.speaker_gru is not None:
speaker_info_sentence = tf.expand_dims(tf.ones([self.max_sentences,1]), 1) * speaker_info
speaker_info_sentence = tf.transpose(speaker_info_sentence, perm=[1,0,2]) # assigning the speaker to each sentence spoken
print("speaker_info_sentence", speaker_info_sentence)
# concatenate fact vectors and attentions for input into attGRU
gru_inputs = tf.concat([fact_vecs, speaker_info_sentence, attentions], 2)
# gru_inputs = tf.concat([fact_vecs, attentions], 2)
with tf.variable_scope('attention_gru', reuse=reuse):
_, episode = tf.nn.dynamic_rnn(AttentionGRUCell(2*self.config.hidden_size),
gru_inputs,
dtype=np.float32,
sequence_length=self.input_len_placeholder
)
else :
# concatenate fact vectors and attentions for input into attGRU
gru_inputs = tf.concat([fact_vecs, attentions], 2)
# gru_inputs = tf.concat([fact_vecs, attentions], 2)
with tf.variable_scope('attention_gru', reuse=reuse):
_, episode = tf.nn.dynamic_rnn(AttentionGRUCell(self.config.hidden_size),
gru_inputs,
dtype=np.float32,
sequence_length=self.input_len_placeholder
)
else :
fact_trans = tf.transpose(fact_vecs, perm=[0,2,1])
final = tf.matmul(fact_trans, attentions)
episode = tf.squeeze(final)
episode = fact_vecs * attentions
# print(episode)
return episode
def generate_episode(self, memory, q_vec, fact_vecs, hop_index):
"""Generate episode by applying attention to current fact vectors through a modified GRU"""
attentions = [tf.squeeze(
self.get_attention(q_vec, memory, fv, bool(hop_index) or bool(i)), axis=1)
for i, fv in enumerate(tf.unstack(fact_vecs, axis=1))]
attentions = tf.transpose(tf.stack(attentions))
self.attentions.append(attentions)
attentions = tf.nn.softmax(attentions)
attentions = tf.expand_dims(attentions, axis=-1)
reuse = True if hop_index > 0 else False
# concatenate fact vectors and attentions for input into attGRU
gru_inputs = tf.concat([fact_vecs, attentions], 2)
with tf.variable_scope('attention_gru', reuse=reuse):
_, episode = tf.nn.dynamic_rnn(AttentionGRUCell(self.config.hidden_size),
gru_inputs,
dtype=np.float32,
sequence_length=self.input_len_placeholder
)
return episode
def generate_episode(self, memory, q_vec, fact_vecs, hop_index):
attentions = [
tf.squeeze(self.attention(q_vec, memory, fact_vec,
bool(hop_index) or bool(i)),
axis=1)
for i, fact_vec in enumerate(tf.unstack(fact_vecs, axis=1))
]
attentions = tf.transpose(tf.stack(attentions))
self.attentions.append(attentions)
attentions = tf.nn.softmax(attentions)
attentions = tf.expand_dims(attentions, axis=-1)
if hop_index > 0:
reuse = True
else:
reuse = False
gru_inputs = tf.concat([fact_vecs, attentions], 2)
with tf.variable_scope('attention_gru', reuse=reuse):
h, episode = tf.nn.dynamic_rnn(
AttentionGRUCell(self.config.hidden_size),
gru_inputs,
dtype=np.float32,
sequence_length=self.input_len_placeholder)
return episode
def generate_episode(self, memory, q, facts, mask, hop_index):
attentions = [
tf.squeeze(self.get_attention(q, memory, fact,
bool(hop_index) or bool(i)),
axis=1)
for i, fact in enumerate(tf.unstack(facts, axis=1))
] # list [num_max_sents] * [batch_size,]
attentions = tf.transpose(tf.stack(attentions)) + (
mask * -1e9) # tensor [batch_size, num_max_sents]
attentions = tf.nn.softmax(attentions)
self.attentions.append(attentions) # for visualization
attentions = tf.expand_dims(
attentions, axis=-1) # tensor [batch_size, num_max_sents, 1]
reuse = True if hop_index > 0 else False
gru_inputs = tf.concat(
[facts, attentions],
2) # tensor [batch_size, num_max_sents, hid_size + 1]
with tf.variable_scope("episode", reuse=reuse):
_, episode = tf.nn.dynamic_rnn(cell=AttentionGRUCell(
self.hid_size),
inputs=gru_inputs,
sequence_length=self.num_sents,
dtype=tf.float32)
return episode