def build_graph(self):
with self._graph.as_default():
rand_seed = self._conf['rand_seed']
tf.set_random_seed(rand_seed)
#word embedding
if self._word_embedding_init is not None:
word_embedding_initializer = tf.constant_initializer(
self._word_embedding_init)
else:
word_embedding_initializer = tf.random_normal_initializer(
stddev=0.1)
self._word_embedding = tf.get_variable(
name='word_embedding',
shape=[self._conf['vocab_size'] + 1, self._conf['emb_size']],
dtype=tf.float32,
initializer=word_embedding_initializer)
#define placehloders
self.turns = tf.placeholder(tf.int32,
shape=[
self._conf["batch_size"],
self._conf["max_turn_num"],
self._conf["max_turn_len"]
])
self.tt_turns_len = tf.placeholder(
tf.int32, shape=[self._conf["batch_size"]])
self.every_turn_len = tf.placeholder(
tf.int32,
shape=[self._conf["batch_size"], self._conf["max_turn_num"]])
self.response = tf.placeholder(
tf.int32,
shape=[self._conf["batch_size"], self._conf["max_turn_len"]])
self.response_len = tf.placeholder(
tf.int32, shape=[self._conf["batch_size"]])
self.label = tf.placeholder(tf.float32,
shape=[self._conf["batch_size"]])
#define operations
#response part
Hr = tf.nn.embedding_lookup(self._word_embedding, self.response)
if self._conf['is_positional'] and self._conf['stack_num'] > 0:
with tf.variable_scope('positional'):
Hr = op.positional_encoding_vector(Hr, max_timescale=10)
for index in range(self._conf['stack_num']):
with tf.variable_scope('self_stack_' + str(index)):
Hr = layers.block(Hr,
Hr,
Hr,
Q_lengths=self.response_len,
K_lengths=self.response_len)
#context part
#a list of length max_turn_num, every element is a tensor with shape [batch, max_turn_len]
list_turn_t = tf.unstack(self.turns, axis=1)
list_turn_length = tf.unstack(self.every_turn_len, axis=1)
sim_turns = []
#for every turn_t calculate matching vector
for turn_t, t_turn_length in zip(list_turn_t, list_turn_length):
Hu = tf.nn.embedding_lookup(
self._word_embedding,
turn_t) #[batch, max_turn_len, emb_size]
if self._conf['is_positional'] and self._conf['stack_num'] > 0:
with tf.variable_scope('positional', reuse=True):
Hu = op.positional_encoding_vector(Hu,
max_timescale=10)
for index in range(self._conf['stack_num']):
with tf.variable_scope('self_stack_' + str(index),
reuse=True):
Hu = layers.block(Hu,
Hu,
Hu,
Q_lengths=t_turn_length,
K_lengths=t_turn_length)
with tf.variable_scope('u_attentd_r_' + str(index)):
try:
u_a_r = layers.block(Hu,
Hr,
Hr,
Q_lengths=t_turn_length,
K_lengths=self.response_len)
except ValueError:
tf.get_variable_scope().reuse_variables()
u_a_r = layers.block(Hu,
Hr,
Hr,
Q_lengths=t_turn_length,
K_lengths=self.response_len)
with tf.variable_scope('r_attend_u_' + str(index)):
try:
r_a_u = layers.block(Hr,
Hu,
Hu,
Q_lengths=self.response_len,
K_lengths=t_turn_length)
except ValueError:
tf.get_variable_scope().reuse_variables()
r_a_u = layers.block(Hr,
Hu,
Hu,
Q_lengths=self.response_len,
K_lengths=t_turn_length)
u_a_r = tf.stack([u_a_r, Hu], axis=-1)
r_a_u = tf.stack([r_a_u, Hr], axis=-1)
#calculate similarity matrix
with tf.variable_scope('similarity'):
# sim shape [batch, max_turn_len, max_turn_len, 2*stack_num+1]
# divide sqrt(200) to prevent gradient explosion
sim = tf.einsum('biks,bjks->bijs', r_a_u,
u_a_r) / tf.sqrt(200.0)
sim_turns.append(sim)
#cnn and aggregation
sim = tf.stack(sim_turns, axis=1)
print('sim shape: %s' % sim.shape)
with tf.variable_scope('cnn_aggregation'):
final_info = layers.CNN_3d(sim, 32, 16)
#for douban
#final_info = layers.CNN_3d(sim, 16, 16)
#loss and train
with tf.variable_scope('loss'):
self.loss, self.logits = layers.loss(final_info, self.label)
self.global_step = tf.Variable(0, trainable=False)
initial_learning_rate = self._conf['learning_rate']
self.learning_rate = tf.train.exponential_decay(
initial_learning_rate,
global_step=self.global_step,
decay_steps=400,
decay_rate=0.9,
staircase=True)
Optimizer = tf.train.AdamOptimizer(self.learning_rate)
self.optimizer = Optimizer.minimize(self.loss)
self.init = tf.global_variables_initializer()
self.saver = tf.train.Saver(
max_to_keep=self._conf["max_to_keep"])
self.all_variables = tf.global_variables()
self.all_operations = self._graph.get_operations()
self.grads_and_vars = Optimizer.compute_gradients(self.loss)
for grad, var in self.grads_and_vars:
if grad is None:
print var
self.capped_gvs = [(tf.clip_by_value(grad, -1, 1), var)
for grad, var in self.grads_and_vars]
self.g_updates = Optimizer.apply_gradients(
self.capped_gvs, global_step=self.global_step)
return self._graph
def build_graph(self):
with self._graph.as_default():
if self._conf['rand_seed'] is not None:
rand_seed = self._conf['rand_seed']
tf.set_random_seed(rand_seed)
print('set tf random seed: %s' % self._conf['rand_seed'])
#word embedding
if self._word_embedding_init is not None:
word_embedding_initializer = tf.constant_initializer(
self._word_embedding_init)
else:
word_embedding_initializer = tf.random_normal_initializer(
stddev=0.1)
self._word_embedding = tf.get_variable(
name='word_embedding',
shape=[self._conf['vocab_size'] + 1, self._conf['emb_size']],
dtype=tf.float32,
initializer=word_embedding_initializer)
#define placehloders
#config max_turn_history_num
self.turns_history = tf.placeholder(
tf.int32,
shape=[
self._conf["batch_size"],
self._conf["max_turn_history_num"],
self._conf["max_turn_len"]
])
self.turns = tf.placeholder(tf.int32,
shape=[
self._conf["batch_size"],
self._conf["max_turn_num"],
self._conf["max_turn_len"]
])
self.tt_turns_len = tf.placeholder(
tf.int32, shape=[self._conf["batch_size"]])
self.every_turn_len = tf.placeholder(
tf.int32,
shape=[self._conf["batch_size"], self._conf["max_turn_num"]])
self.response = tf.placeholder(
tf.int32,
shape=[self._conf["batch_size"], self._conf["max_turn_len"]])
self.response_len = tf.placeholder(
tf.int32, shape=[self._conf["batch_size"]])
self.label = tf.placeholder(tf.float32,
shape=[self._conf["batch_size"]])
#define operations
#response part
Hr = tf.nn.embedding_lookup(self._word_embedding, self.response)
turns_history_embedding = tf.nn.embedding_lookup(
self._word_embedding, self.turns_history)
if self._conf['is_positional'] and self._conf['stack_num'] > 0:
with tf.variable_scope('positional'):
Hr = op.positional_encoding_vector(Hr, max_timescale=10)
Hr_stack = [Hr]
_batch_size, _turn_nums, _turn_words, _emb_size = turns_history_embedding.get_shape(
).as_list()
turns_history_embedding = tf.reshape(turns_history_embedding,
[-1, _turn_words, _emb_size])
for index in range(self._conf['stack_num']):
turns_history_embedding, _ = self._multihead(
turns_history_embedding, turns_history_embedding,
turns_history_embedding)
turns_history_embedding = tf.reshape(
turns_history_embedding,
[_batch_size, _turn_nums, _turn_words, _emb_size])
for index in range(self._conf['stack_num']):
with tf.variable_scope('self_stack_' + str(index)):
Hr = layers.block(Hr,
Hr,
Hr,
Q_lengths=self.response_len,
K_lengths=self.response_len)
Hr_stack.append(Hr)
with tf.variable_scope('respone_extraction_history'):
turn_important_inf = []
#需要增加一个全链接层
for _t in tf.split(turns_history_embedding,
self._conf['max_turn_history_num'], 1):
_t = tf.squeeze(_t)
#_match_result = layers.attention(Hr_stack[-1], _t, _t, self.response_len, self.response_len)
_match_result = layers.attention(
self._dense1(Hr_stack[-1]), _t, _t, self.response_len,
self.response_len)
turn_important_inf.append(tf.expand_dims(_match_result, 1))
best_turn_match = tf.concat(turn_important_inf, 1)
with tf.variable_scope('response_extraciton_best_information'):
#best_information,_ = self._multihead(Hr_stack[-1], best_turn_match, best_turn_match)
best_information, _ = self._multihead(
self._dense2(Hr_stack[-1]), best_turn_match,
best_turn_match)
best_information = layers.FFN(best_information)
#context part
#a list of length max_turn_num, every element is a tensor with shape [batch, max_turn_len]
list_turn_t = tf.unstack(self.turns, axis=1)
list_turn_length = tf.unstack(self.every_turn_len, axis=1)
sim_turns = []
#for every turn_t calculate matching vector
for turn_t, t_turn_length in zip(list_turn_t, list_turn_length):
Hu = tf.nn.embedding_lookup(
self._word_embedding,
turn_t) #[batch, max_turn_len, emb_size]
if self._conf['is_positional'] and self._conf['stack_num'] > 0:
with tf.variable_scope('positional', reuse=True):
Hu = op.positional_encoding_vector(Hu,
max_timescale=10)
Hu_stack = [Hu]
for index in range(self._conf['stack_num']):
with tf.variable_scope('self_stack_' + str(index),
reuse=True):
Hu = layers.block(Hu,
Hu,
Hu,
Q_lengths=t_turn_length,
K_lengths=t_turn_length)
Hu_stack.append(Hu)
r_a_t_stack = []
t_a_r_stack = []
for index in range(self._conf['stack_num'] + 1):
with tf.variable_scope('t_attend_r_' + str(index)):
try:
t_a_r = layers.block(tf.add(
Hu_stack[index], best_information),
Hr_stack[index],
Hr_stack[index],
Q_lengths=t_turn_length,
K_lengths=self.response_len)
except ValueError:
tf.get_variable_scope().reuse_variables()
t_a_r = layers.block(tf.add(
Hu_stack[index], best_information),
Hr_stack[index],
Hr_stack[index],
Q_lengths=t_turn_length,
K_lengths=self.response_len)
with tf.variable_scope('r_attend_t_' + str(index)):
try:
r_a_t = layers.block(
Hr_stack[index],
tf.add(Hu_stack[index], best_information),
tf.add(Hu_stack[index], best_information),
Q_lengths=self.response_len,
K_lengths=t_turn_length)
except ValueError:
tf.get_variable_scope().reuse_variables()
r_a_t = layers.block(
Hr_stack[index],
tf.add(Hu_stack[index], best_information),
tf.add(Hu_stack[index], best_information),
Q_lengths=self.response_len,
K_lengths=t_turn_length)
t_a_r_stack.append(t_a_r)
r_a_t_stack.append(r_a_t)
t_a_r_stack.extend(Hu_stack)
r_a_t_stack.extend(Hr_stack)
t_a_r = tf.stack(t_a_r_stack, axis=-1)
r_a_t = tf.stack(r_a_t_stack, axis=-1)
#calculate similarity matrix
with tf.variable_scope('similarity'):
# sim shape [batch, max_turn_len, max_turn_len, 2*stack_num+1]
# divide sqrt(200) to prevent gradient explosion
sim = tf.einsum('biks,bjks->bijs', t_a_r,
r_a_t) / tf.sqrt(200.0)
sim_turns.append(sim)
#cnn and aggregation
sim = tf.stack(sim_turns, axis=1)
print('sim shape: %s' % sim.shape)
with tf.variable_scope('cnn_aggregation'):
final_info = layers.CNN_3d(sim, 32, 16)
#final_info_dim = final_info.get_shape().as_list()[-1]
#for douban
#final_info = layers.CNN_3d(sim, 16, 16)
# _x = self._conv1d(best_information)
# _x = self._pool1d(_x)
#final_info = tf.concat([final_info,best_information],-1)
#loss and train
with tf.variable_scope('loss'):
self.loss, self.logits = layers.loss(final_info, self.label)
self.global_step = tf.Variable(0, trainable=False)
initial_learning_rate = self._conf['learning_rate']
self.learning_rate = tf.train.exponential_decay(
initial_learning_rate,
global_step=self.global_step,
decay_steps=400,
decay_rate=0.9,
staircase=True)
Optimizer = tf.train.AdamOptimizer(self.learning_rate)
self.optimizer = Optimizer.minimize(
self.loss, global_step=self.global_step)
self.init = tf.global_variables_initializer()
self.saver = tf.train.Saver(
max_to_keep=self._conf["max_to_keep"])
self.all_variables = tf.global_variables()
self.all_operations = self._graph.get_operations()
self.grads_and_vars = Optimizer.compute_gradients(self.loss)
for grad, var in self.grads_and_vars:
if grad is None:
print(var)
self.capped_gvs = [(tf.clip_by_value(grad, -1, 1), var)
for grad, var in self.grads_and_vars]
self.g_updates = Optimizer.apply_gradients(
self.capped_gvs, global_step=self.global_step)
return self._graph
def self_cross_attention_block(config, Hu, every_turn_len, Hr, response_len):
"""
:param config:
:param Hu: shape = (batch_size, max_turn_num, sentence_len, emb_size)
:param every_turn_len: shape = (batch_size, max_turn_num )
:param Hr: shape = (batch_size, sentence_len, emb_size)
:param response_len: shape = (batch_size)
:return:
"""
if config['is_positional'] and config['stack_num'] > 0:
with tf.variable_scope('positional', reuse=tf.AUTO_REUSE):
Hr = op.positional_encoding_vector(Hr, max_timescale=10)
Hr_stack = [Hr]
for index in range(config['stack_num']):
with tf.variable_scope('self_stack_' + str(index),
reuse=tf.AUTO_REUSE):
# Hr.shape = (batch_size, max_turn_len, emb_size)
Hr = layers.block(Hr,
Hr,
Hr,
Q_lengths=response_len,
K_lengths=response_len)
Hr_stack.append(Hr)
# context part
# a list of length max_turn_num, every element is a tensor with shape [batch, max_turn_len, emb_size]
list_turn_t = tf.unstack(Hu, axis=1)
list_turn_length = tf.unstack(every_turn_len, axis=1)
sim_turns = []
# for every Hu calculate matching vector
for Hu, t_turn_length in zip(list_turn_t, list_turn_length):
if config['is_positional'] and config['stack_num'] > 0:
with tf.variable_scope('positional', reuse=tf.AUTO_REUSE):
Hu = op.positional_encoding_vector(Hu, max_timescale=10)
Hu_stack = [Hu]
for index in range(config['stack_num']):
with tf.variable_scope('self_stack_' + str(index),
reuse=tf.AUTO_REUSE):
Hu = layers.block(Hu,
Hu,
Hu,
Q_lengths=t_turn_length,
K_lengths=t_turn_length)
Hu_stack.append(Hu)
r_a_t_stack = []
t_a_r_stack = []
for index in range(config['stack_num'] + 1):
with tf.variable_scope('t_attend_r_' + str(index),
reuse=tf.AUTO_REUSE):
try:
t_a_r = layers.block(Hu_stack[index],
Hr_stack[index],
Hr_stack[index],
Q_lengths=t_turn_length,
K_lengths=response_len)
except ValueError:
tf.get_variable_scope().reuse_variables()
t_a_r = layers.block(Hu_stack[index],
Hr_stack[index],
Hr_stack[index],
Q_lengths=t_turn_length,
K_lengths=response_len)
with tf.variable_scope('r_attend_t_' + str(index),
reuse=tf.AUTO_REUSE):
try:
r_a_t = layers.block(Hr_stack[index],
Hu_stack[index],
Hu_stack[index],
Q_lengths=response_len,
K_lengths=t_turn_length)
except ValueError:
tf.get_variable_scope().reuse_variables()
r_a_t = layers.block(Hr_stack[index],
Hu_stack[index],
Hu_stack[index],
Q_lengths=response_len,
K_lengths=t_turn_length)
t_a_r_stack.append(t_a_r)
r_a_t_stack.append(r_a_t)
t_a_r_stack.extend(Hu_stack)
r_a_t_stack.extend(Hr_stack)
t_a_r = tf.stack(t_a_r_stack, axis=-1)
r_a_t = tf.stack(r_a_t_stack, axis=-1)
# calculate similarity matrix
with tf.variable_scope('similarity', reuse=tf.AUTO_REUSE):
# sim shape [batch, max_turn_len, max_turn_len, 2*stack_num+1]
# divide sqrt(200) to prevent gradient explosion
sim = tf.einsum('biks,bjks->bijs', t_a_r, r_a_t) / tf.sqrt(200.0)
sim_turns.append(sim)
# cnn and aggregation
sim = tf.stack(sim_turns, axis=1)
print('sim shape: %s' % sim.shape)
with tf.variable_scope('cnn_aggregation', reuse=tf.AUTO_REUSE):
final_info = layers.CNN_3d(sim, 32, 16)
with tf.variable_scope('linear', reuse=tf.AUTO_REUSE):
W = tf.get_variable(name='weights',
shape=[final_info.shape[-1], 1],
initializer=tf.orthogonal_initializer())
bias = tf.get_variable(name='bias',
shape=[1],
initializer=tf.zeros_initializer())
logits = tf.reshape(tf.matmul(final_info, W) + bias, [-1])
return logits
def build_graph(self):
with self._graph.as_default():
if self._conf['rand_seed'] is not None:
rand_seed = self._conf['rand_seed']
tf.set_random_seed(rand_seed)
print('set tf random seed: %s' % self._conf['rand_seed'])
# word embedding
if self._word_embedding_init is not None:
word_embedding_initializer = tf.constant_initializer(
self._word_embedding_init)
else:
word_embedding_initializer = tf.random_normal_initializer(
stddev=0.1)
self._word_embedding = tf.get_variable(
name='word_embedding',
shape=[self._conf['vocab_size'] + 1, self._conf['emb_size']],
dtype=tf.float32,
initializer=word_embedding_initializer)
# define placehloders
self.turns = tf.placeholder(
tf.int32,
shape=[self._conf["batch_size"], self._conf["max_turn_num"],
self._conf["max_turn_len"]])
self.tt_turns_len = tf.placeholder( # turn_num
tf.int32,
shape=[self._conf["batch_size"]])
self.every_turn_len = tf.placeholder(
tf.int32,
shape=[self._conf["batch_size"], self._conf["max_turn_num"]])
self.turns_intent = tf.placeholder(
tf.float32,
shape=[self._conf["batch_size"], self._conf["max_turn_num"],
self._conf["intent_size"]])
self.response = tf.placeholder(
tf.int32,
shape=[self._conf["batch_size"], self._conf["max_turn_len"]])
self.response_len = tf.placeholder(
tf.int32,
shape=[self._conf["batch_size"]])
self.response_intent = tf.placeholder(
tf.float32,
shape=[self._conf["batch_size"], self._conf["intent_size"]])
self.label = tf.placeholder(
tf.float32,
shape=[self._conf["batch_size"]])
# define operations
# response part
Hr = tf.nn.embedding_lookup(self._word_embedding, self.response)
# [batch_size, max_turn_len, embed_size]
# print('[after embedding_lookup] Hr shape: %s' % Hr.shape)
if self._conf['is_positional'] and self._conf['stack_num'] > 0:
with tf.variable_scope('positional'):
Hr = op.positional_encoding_vector(Hr, max_timescale=10)
Hr_stack = [Hr] # 1st element of Hr_stack is the orginal embedding
# lyang comments: self attention
for index in range(self._conf['stack_num']):
# print('[self attention for response] stack index: %d ' % index)
with tf.variable_scope('self_stack_' + str(index)):
# [batch, max_turn_len, emb_size]
Hr = layers.block( # attentive module
Hr, Hr, Hr,
Q_lengths=self.response_len,
K_lengths=self.response_len)
# print('[after layers.block] Hr shape: %s' % Hr.shape)
# Hr is still [batch_size, max_turn_len, embed_size]
Hr_stack.append(Hr)
# print('[after self attention of response] len(Hr_stack)',
# len(Hr_stack)) # 1+stack_num
# context part
# a list of length max_turn_num, every element is a tensor with shape [batch, max_turn_len]
list_turn_t = tf.unstack(self.turns, axis=1)
list_turn_length = tf.unstack(self.every_turn_len, axis=1)
list_turn_intent = tf.unstack(self.turns_intent, axis=1)
sim_turns = []
attention_turns = [] # intent based attention on each turn
# for every turn_t calculate matching vector
turn_index = 0
for turn_t, t_turn_length, t_intent in zip(list_turn_t, list_turn_length, list_turn_intent):
print('current turn_index : ', turn_index)
turn_index += 1
Hu = tf.nn.embedding_lookup(self._word_embedding,
turn_t) # [batch, max_turn_len, emb_size]
# print('[after embedding_lookup] Hu shape: %s' % Hu.shape)
if self._conf['is_positional'] and self._conf['stack_num'] > 0:
with tf.variable_scope('positional', reuse=True):
Hu = op.positional_encoding_vector(Hu,
max_timescale=10)
Hu_stack = [Hu] # 1st element of Hu_stack is the orginal embedding
# lyang comments: self attention
for index in range(self._conf['stack_num']):
# print('[self attention for context turn] stack index: %d ' % index)
with tf.variable_scope('self_stack_' + str(index),
reuse=True):
# [batch, max_turn_len, emb_size]
Hu = layers.block( # attentive module
Hu, Hu, Hu,
Q_lengths=t_turn_length, K_lengths=t_turn_length)
# print('[after layers.block] Hu shape: %s' % Hu.shape)
Hu_stack.append(Hu)
# print('[after self attention of context turn] len(Hu_stack)',
# len(Hu_stack)) # 1+stack_num
# lyang comments: cross attention
# print('[cross attention ...]')
r_a_t_stack = []
t_a_r_stack = []
# cross attention
for index in range(self._conf['stack_num'] + 1):
# print('[cross attention] stack index = ', index)
with tf.variable_scope('t_attend_r_' + str(index)):
try:
# [batch, max_turn_len, emb_size]
t_a_r = layers.block( # attentive module
Hu_stack[index], Hr_stack[index],
Hr_stack[index],
Q_lengths=t_turn_length,
K_lengths=self.response_len)
except ValueError:
tf.get_variable_scope().reuse_variables()
t_a_r = layers.block(
# [batch, max_turn_len, emb_size]
Hu_stack[index], Hr_stack[index],
Hr_stack[index],
Q_lengths=t_turn_length,
K_lengths=self.response_len)
# print('[cross attention t_attend_r_] stack index: %d, t_a_r.shape: %s' % (
# index, t_a_r.shape))
with tf.variable_scope('r_attend_t_' + str(index)):
try:
# [batch, max_turn_len, emb_size]
r_a_t = layers.block( # attentive module
Hr_stack[index], Hu_stack[index],
Hu_stack[index],
Q_lengths=self.response_len,
K_lengths=t_turn_length)
except ValueError:
tf.get_variable_scope().reuse_variables()
r_a_t = layers.block(
Hr_stack[index], Hu_stack[index],
Hu_stack[index],
Q_lengths=self.response_len,
K_lengths=t_turn_length)
# print('[cross attention r_a_t_] stack index: %d, r_a_t.shape: %s' % (
# index, r_a_t.shape))
t_a_r_stack.append(t_a_r)
r_a_t_stack.append(r_a_t)
# print('[cross attention] len(t_a_r_stack):', len(t_a_r_stack))
# print('[cross attention] len(r_a_t_stack):', len(r_a_t_stack))
# print('[before extend] len(t_a_r_stack):', len(t_a_r_stack))
# print('[before extend] len(r_a_t_stack):', len(r_a_t_stack))
# lyang comments: 3D aggregation
t_a_r_stack.extend(
Hu_stack) # half from self-attention; half from cross-attention
r_a_t_stack.extend(
Hr_stack) # half from self-attention; half from cross-attention
# after extend, len(t_a_r_stack)) = 2*(stack_num+1)
# print('[after extend] len(t_a_r_stack):', len(t_a_r_stack))
# print('[after extend] len(r_a_t_stack):', len(r_a_t_stack))
t_a_r = tf.stack(t_a_r_stack, axis=-1)
r_a_t = tf.stack(r_a_t_stack, axis=-1)
# print('after stack along the last dimension: ')
# print('t_a_r shape: %s' % t_a_r.shape)
# print('r_a_t shape: %s' % r_a_t.shape)
# after stack, t_a_r and r_a_t are (batch, max_turn_len, embed_size, 2*(stack_num+1))
with tf.variable_scope('intent_based_attention',
reuse=tf.AUTO_REUSE): # share parameter across different turns
# there are 3 different ways to implement intent based attention
# implement these three different variations and compare the
# effectiveness as model abalation analysis
# let I_u_t and I_r_k are intent vector in [12,1]
# 1. dot: w * [I_u_t, I_r_k], where w is [24,1]
# 2. biliear: I_u_t' * w * I_r_k, where w is [12,12]
# 3. outprod: I_u_t * I_r_k' -> [12,12] out product ->
# flaten to [144,1] outprod -> w*outprod
# where w is [1,144]
attention_logits = layers.attention_intent(t_intent,
self.response_intent,
self._conf['intent_attention_type'])
# print('[intent_based_attention] attention_logits.shape: %s' % attention_logits.shape)
attention_turns.append(attention_logits)
# calculate similarity matrix
with tf.variable_scope('similarity'):
# sim shape [batch, max_turn_len, max_turn_len, 2*(stack_num+1)]
# divide sqrt(200) to prevent gradient explosion
# A_biks * B_bjks -> C_bijs
sim = tf.einsum('biks,bjks->bijs', t_a_r, r_a_t) / tf.sqrt(
200.0)
# (batch, max_turn_len, embed_size, 2*(stack_num+1)) *
# (batch, max_turn_len, embed_size, 2*(stack_num+1)) ->
# [batch, max_turn_len, max_turn_len, 2*(stack_num+1)]
# where k is corresponding to the dimension of embed_size,
# which can be eliminated by dot product with einsum
# print('[similarity] after einsum dot prod sim shape: %s' % sim.shape)
# [batch, max_turn_len, max_turn_len, 2*(stack_num+1)]
# ! Here we multipy sim by intent based attention weights before
# append sim into sim_turns in order to generate the weighted
# stack in the next step
sim_turns.append(sim)
# print('[similarity] after append, len(sim_turns):', len(sim_turns))
attention_logits = tf.stack(attention_turns, axis=1) # [batch, max_turn_num]
print('[attention_logits] after stack attention_logits.shape: %s' % attention_logits.shape)
# add mask in attention following the way in BERT
# real turn_num is in self.tt_turns_len [batch]
# return a mask tensor with shape [batch, conf['max_turn_num']]
attention_mask = tf.sequence_mask(self.tt_turns_len, self._conf['max_turn_num'],
dtype=tf.float32)
print('[attention_mask] attention_mask.shape: %s' % attention_mask.shape)
# Since attention_mask is 1.0 for positions we want to attend and 0.0 for
# masked positions, this operation will create a tensor which is 0.0 for
# positions we want to attend and -10000.0 for masked positions.
adder = (1.0 - attention_mask) * -10000.0
# Since we are adding it to the raw scores before the softmax, this is
# effectively the same as removing these entirely.
attention_logits += adder
attention = tf.nn.softmax(attention_logits) # by default softmax along dim=-1 [batch, max_turn_num]
print('[attention] attention.shape: %s' % attention_mask.shape)
self.attention = attention # will print it for visualization
# cnn and aggregation
# lyang comments aggregation by 3D CNN layer
# [3d cnn aggregation] sim shape: (32, 9, 180, 180, 10)
# conv_0 shape: (32, 9, 180, 180, 16)
# pooling_0 shape: (32, 3, 60, 60, 16)
# conv_1 shape: (32, 3, 60, 60, 16)
# pooling_1 shape: (32, 1, 20, 20, 16)
# [3d cnn aggregation] final_info: (32, 6400) # [batch * feature_size]
# [batch, max_turn_num, max_turn_len, max_turn_len, 2*(stack_num+1)]
# (32, 9, 180, 180, 10)
sim = tf.stack(sim_turns, axis=1)
# multipy sim by attention score
sim = tf.einsum('bijks,bi->bijks', sim, attention)
print('[3d cnn aggregation] sim shape: %s' % sim.shape)
with tf.variable_scope('cnn_aggregation'):
final_info = layers.CNN_3d(sim, self._conf['cnn_3d_oc0'],
self._conf['cnn_3d_oc1'])
# for udc
# final_info = layers.CNN_3d(sim, 32, 16)
# for douban
# final_info = layers.CNN_3d(sim, 16, 16)
print('[3d cnn aggregation] final_info: %s' % final_info.shape)
# loss and train
with tf.variable_scope('loss'):
self.loss, self.logits = layers.loss(final_info, self.label)
self.global_step = tf.Variable(0, trainable=False)
initial_learning_rate = self._conf['learning_rate']
self.learning_rate = tf.train.exponential_decay(
initial_learning_rate,
global_step=self.global_step,
decay_steps=400,
decay_rate=0.9,
staircase=True)
Optimizer = tf.train.AdamOptimizer(self.learning_rate)
self.optimizer = Optimizer.minimize(
self.loss,
global_step=self.global_step)
self.init = tf.global_variables_initializer()
self.saver = tf.train.Saver(
max_to_keep=self._conf["max_to_keep"])
self.all_variables = tf.global_variables()
self.all_operations = self._graph.get_operations()
self.grads_and_vars = Optimizer.compute_gradients(self.loss)
for grad, var in self.grads_and_vars:
if grad is None:
print var
self.capped_gvs = [(tf.clip_by_value(grad, -1, 1), var) for
grad, var in self.grads_and_vars]
self.g_updates = Optimizer.apply_gradients(
self.capped_gvs,
global_step=self.global_step)
return self._graph
def dam_model(input_x,
input_x_mask,
input_y,
input_y_mask,
word_emb,
keep_rate,
conf,
x_len=None,
y_len=None):
Hr = tf.nn.embedding_lookup(word_emb, input_y)
if conf['is_positional'] and conf['stack_num'] > 0:
with tf.variable_scope('positional'):
Hr = op.positional_encoding_vector(Hr, max_timescale=10)
Hr_stack = [Hr]
for index in range(conf['stack_num']):
with tf.variable_scope('self_stack_cr_' + str(index)):
Hr = layers.block(Hr, Hr, Hr, Q_lengths=y_len, K_lengths=y_len)
Hr_stack.append(Hr)
#context part
#a list of length max_turn_num, every element is a tensor with shape [batch, max_turn_len]
list_turn_t = tf.unstack(input_x, axis=1)
list_turn_length = tf.unstack(x_len, axis=1)
sim_turns = []
#for every turn_t calculate matching vector
for turn_t, t_turn_length in zip(list_turn_t, list_turn_length):
Hu = tf.nn.embedding_lookup(word_emb,
turn_t) #[batch, max_turn_len, emb_size]
if conf['is_positional'] and conf['stack_num'] > 0:
with tf.variable_scope('positional', reuse=True):
Hu = op.positional_encoding_vector(Hu, max_timescale=10)
Hu_stack = [Hu]
for index in range(conf['stack_num']):
with tf.variable_scope('self_stack_cr_' + str(index), reuse=True):
Hu = layers.block(Hu,
Hu,
Hu,
Q_lengths=t_turn_length,
K_lengths=t_turn_length)
Hu_stack.append(Hu)
r_a_t_stack = []
t_a_r_stack = []
for index in range(conf['stack_num'] + 1):
with tf.variable_scope('t_attend_r_cr_' + str(index)):
try:
t_a_r = layers.block(Hu_stack[index],
Hr_stack[index],
Hr_stack[index],
Q_lengths=t_turn_length,
K_lengths=y_len)
except ValueError:
tf.get_variable_scope().reuse_variables()
t_a_r = layers.block(Hu_stack[index],
Hr_stack[index],
Hr_stack[index],
Q_lengths=t_turn_length,
K_lengths=y_len)
with tf.variable_scope('r_attend_t_cr_' + str(index)):
try:
r_a_t = layers.block(Hr_stack[index],
Hu_stack[index],
Hu_stack[index],
Q_lengths=y_len,
K_lengths=t_turn_length)
except ValueError:
tf.get_variable_scope().reuse_variables()
r_a_t = layers.block(Hr_stack[index],
Hu_stack[index],
Hu_stack[index],
Q_lengths=y_len,
K_lengths=t_turn_length)
t_a_r_stack.append(t_a_r)
r_a_t_stack.append(r_a_t)
t_a_r_stack.extend(Hu_stack)
r_a_t_stack.extend(Hr_stack)
t_a_r = tf.stack(t_a_r_stack, axis=-1)
r_a_t = tf.stack(r_a_t_stack, axis=-1)
#calculate similarity matrix
with tf.variable_scope('similarity'):
# sim shape [batch, max_turn_len, max_turn_len, 2*stack_num+1]
# divide sqrt(200) to prevent gradient explosion
sim = tf.einsum('biks,bjks->bijs', t_a_r, r_a_t) / tf.sqrt(200.0)
sim_turns.append(sim)
#cnn and aggregation
sim = tf.stack(sim_turns, axis=1)
print('sim shape: %s' % sim.shape)
with tf.variable_scope('cnn_aggregation'):
final_info = layers.CNN_3d(sim, 32, 16)
#for douban
#final_info = layers.CNN_3d(sim, 16, 16)
return final_info
