def cnn_model(num_steps, num_context, num_event, keep_rate_input, dae_weight,
phase_indicator, autoencoder_length, filter_num, filter_size,
autoencoder_initializer):
with tf.name_scope('data_source'):
# 标准输入规定为TBD
batch_size = tf.placeholder(tf.int32, [], name='batch_size')
event_placeholder = tf.placeholder(tf.float32,
[num_steps, None, num_event],
name='event_placeholder')
context_placeholder = tf.placeholder(tf.float32,
[num_steps, None, num_context],
name='context_placeholder')
time_list = tf.placeholder(tf.int32, [None, num_steps],
name='time_list')
y_placeholder = tf.placeholder(tf.float32, [None, 1],
name='y_placeholder')
with tf.name_scope('autoencoder'):
# input_x 用于计算重构原始向量时产生的误差
processed_input, autoencoder_weight = autoencoder.denoising_autoencoder(
phase_indicator, context_placeholder, keep_rate_input,
autoencoder_length, autoencoder_initializer)
with tf.name_scope('cnn'):
predictions, scores = _cnn_detail(processed_input=processed_input,
event_list=event_placeholder,
time_list=time_list,
filter_num=filter_num,
filter_size_list=filter_size,
scope='cnn')
with tf.name_scope('loss'):
with tf.name_scope('pred_loss'):
loss_pred = tf.losses.sigmoid_cross_entropy(
logits=scores, multi_class_labels=y_placeholder)
with tf.name_scope('dae_loss'):
if autoencoder_length > 0:
loss_dae = autoencoder.autoencoder_loss(
embedding=processed_input,
origin_input=context_placeholder,
weight=autoencoder_weight)
else:
loss_dae = 0
with tf.name_scope('loss_sum'):
loss = loss_pred + loss_dae * dae_weight
return loss, predictions, event_placeholder, context_placeholder, y_placeholder, batch_size, phase_indicator, \
time_list
def vanilla_rnn_model(cell,
num_steps,
num_hidden,
num_context,
num_event,
keep_rate_input,
dae_weight,
phase_indicator,
auto_encoder_value,
auto_encoder_initializer=tf.initializers.orthogonal()):
"""
:param cell:
:param num_steps:
:param num_hidden:
:param num_context: 要求Context变量全部变为二值变量
:param num_event:
:param dae_weight:
:param keep_rate_input:
:param phase_indicator:
:param auto_encoder_value: 大于0时执行对输入的自编码，值即为最终降到的维度
:param auto_encoder_initializer:
:return:
loss, prediction, x_placeholder, y_placeholder, batch_size, phase_indicator
其中 phase_indicator>0代表是测试期，<=0代表是训练期
"""
with tf.name_scope('vanilla_rnn'):
with tf.name_scope('data_source'):
batch_size = tf.placeholder(tf.int32, [], name='batch_size')
# 标准输入规定为BTD
event_placeholder = tf.placeholder(tf.float32,
[None, num_steps, num_event],
name='event_placeholder')
context_placeholder = tf.placeholder(
tf.float32, [None, num_steps, num_context],
name='context_placeholder')
y_placeholder = tf.placeholder(tf.float32, [None, 1],
name='y_placeholder')
# input_x 用于计算重构原始向量时产生的误差
processed_input, input_x, autoencoder_weight = autoencoder.denoising_autoencoder(
phase_indicator, context_placeholder, event_placeholder,
keep_rate_input, auto_encoder_value, auto_encoder_initializer)
output_list = __vanilla_rnn(batch_size=batch_size,
rnn_cell=cell,
input_x=processed_input)
with tf.variable_scope('output_layer', reuse=tf.AUTO_REUSE):
output_weight = tf.get_variable(
"weight", [num_hidden, 1],
initializer=tf.initializers.orthogonal())
bias = tf.get_variable('bias', [])
with tf.name_scope('loss'):
unnormalized_prediction = tf.matmul(output_list[-1],
output_weight) + bias
loss_pred = tf.losses.sigmoid_cross_entropy(
logits=unnormalized_prediction, multi_class_labels=y_placeholder)
if auto_encoder_value > 0:
loss_dae = autoencoder.autoencoder_loss(embedding=processed_input,
origin_input=input_x,
weight=autoencoder_weight)
else:
loss_dae = 0
loss = loss_pred + loss_dae * dae_weight
with tf.name_scope('prediction'):
prediction = tf.sigmoid(unnormalized_prediction)
return loss, prediction, event_placeholder, context_placeholder, y_placeholder, batch_size, phase_indicator
def hawkes_rnn_model(cell,
num_steps,
num_hidden,
num_context,
num_event,
keep_rate_input,
dae_weight,
phase_indicator,
autoencoder_length,
autoencoder_initializer=tf.initializers.orthogonal()):
"""
:param cell:
:param num_steps:
:param num_hidden:
:param num_context: 要求Context变量全部变为二值变量
:param num_event:
:param dae_weight:
:param keep_rate_input:
:param phase_indicator:
:param autoencoder_length: 大于0时执行对输入的自编码，值即为最终降到的维度
:param autoencoder_initializer:
:return:
loss, prediction, x_placeholder, y_placeholder, batch_size, phase_indicator
其中 phase_indicator>0代表是测试期，<=0代表是训练期
"""
with tf.name_scope('data_source'):
# 标准输入规定为TBD
batch_size = tf.placeholder(tf.int32, [], name='batch_size')
event_placeholder = tf.placeholder(tf.float32,
[num_steps, None, num_event],
name='event_placeholder')
context_placeholder = tf.placeholder(tf.float32,
[num_steps, None, num_context],
name='context_placeholder')
base_intensity = tf.placeholder(tf.float32, [num_event, 1],
name='base_intensity')
mutual_intensity = tf.placeholder(tf.float32, [num_event, num_event],
name='mutual_intensity')
time_list = tf.placeholder(tf.int32, [None, num_steps],
name='time_list')
task_index = tf.placeholder(tf.int32, [], name='task_index')
sequence_length = tf.placeholder(tf.int32, [None],
name='sequence_length')
y_placeholder = tf.placeholder(tf.float32, [None, 1],
name='y_placeholder')
initial_state = cell.get_initial_state(batch_size)
with tf.name_scope('autoencoder'):
processed_input, autoencoder_weight = autoencoder.denoising_autoencoder(
phase_indicator, context_placeholder, keep_rate_input,
autoencoder_length, autoencoder_initializer)
with tf.name_scope('hawkes_rnn'):
input_ = tf.concat([processed_input, event_placeholder], axis=2)
outputs, final_state = _hawkes_dynamic_rnn(
cell,
input_,
sequence_length,
initial_state,
base_intensity=base_intensity,
task_index=task_index,
mutual_intensity=mutual_intensity,
time_list=time_list,
event_list=event_placeholder)
# 在使用时LSTM时比较麻烦，因为state里同时包含了hidden state和cell state，只有后者是需要输出的
# 因此需要额外需要做一个split。这种写法非常不优雅，但是我想了一想，也没什么更好的办法
# 做split时，需要特别注意一下state里到底谁前谁后
output_length = outputs.shape[2].value
state_length = final_state.shape[1].value
if output_length == state_length:
# 不需要做任何事情
pass
elif output_length * 2 == state_length:
final_state = tf.split(final_state, 2, axis=1)[0]
else:
raise ValueError('Invalid Size')
with tf.variable_scope('output_para'):
output_weight = tf.get_variable(
"weight", [num_hidden, 1],
initializer=tf.initializers.orthogonal())
bias = tf.get_variable('bias', [])
with tf.name_scope('prediction'):
unnormalized_prediction = tf.matmul(final_state, output_weight) + bias
prediction = tf.sigmoid(unnormalized_prediction)
with tf.name_scope('loss'):
with tf.name_scope('pred_loss'):
loss_pred = tf.losses.sigmoid_cross_entropy(
logits=unnormalized_prediction,
multi_class_labels=y_placeholder)
with tf.name_scope('dae_loss'):
if autoencoder_length > 0:
loss_dae = autoencoder.autoencoder_loss(
embedding=processed_input,
origin_input=context_placeholder,
weight=autoencoder_weight)
else:
loss_dae = 0
with tf.name_scope('loss_sum'):
loss = loss_pred + loss_dae * dae_weight
return loss, prediction, event_placeholder, context_placeholder, y_placeholder, batch_size, phase_indicator, \
base_intensity, mutual_intensity, time_list, task_index, sequence_length, final_state
def concat_hawkes_model(cell_context,
cell_event,
num_steps,
num_hidden,
num_context,
num_event,
keep_rate_input,
dae_weight,
phase_indicator,
autoencoder_length,
autoencoder_initializer=tf.initializers.orthogonal()):
with tf.name_scope('data_source'):
# 标准输入规定为TBD
batch_size = tf.placeholder(tf.int32, [], name='batch_size')
event_placeholder = tf.placeholder(tf.float32,
[num_steps, None, num_event],
name='event_placeholder')
context_placeholder = tf.placeholder(tf.float32,
[num_steps, None, num_context],
name='context_placeholder')
base_intensity = tf.placeholder(tf.float32, [num_event, 1],
name='base_intensity')
mutual_intensity = tf.placeholder(tf.float32, [num_event, num_event],
name='mutual_intensity')
time_list = tf.placeholder(tf.int32, [None, num_steps],
name='time_list')
task_index = tf.placeholder(tf.int32, [], name='task_index')
sequence_length = tf.placeholder(tf.int32, [None],
name='sequence_length')
y_placeholder = tf.placeholder(tf.float32, [None, 1],
name='y_placeholder')
event_initial_state = cell_event.get_initial_state(batch_size)
context_initial_state = cell_context.get_initial_state(batch_size)
with tf.name_scope('autoencoder'):
# input_x 用于计算重构原始向量时产生的误差
processed_input, autoencoder_weight = autoencoder.denoising_autoencoder(
phase_indicator, context_placeholder, keep_rate_input,
autoencoder_length, autoencoder_initializer)
with tf.name_scope('context_hawkes_rnn'):
outputs, final_state = _hawkes_dynamic_rnn(
cell_context,
processed_input,
sequence_length,
context_initial_state,
base_intensity=base_intensity,
task_index=task_index,
mutual_intensity=mutual_intensity,
time_list=time_list,
event_list=event_placeholder,
scope='context_hawkes')
output_length = outputs.shape[2].value
state_length = final_state.shape[1].value
if output_length == state_length:
# 不需要做任何事情
context_final_state = final_state
elif output_length * 2 == state_length:
context_final_state = tf.split(final_state, 2, axis=1)[0]
else:
raise ValueError('Invalid Size')
with tf.name_scope('event_hawkes_rnn'):
outputs, final_state = _hawkes_dynamic_rnn(
cell_event,
event_placeholder,
sequence_length,
event_initial_state,
base_intensity=base_intensity,
task_index=task_index,
mutual_intensity=mutual_intensity,
time_list=time_list,
event_list=event_placeholder,
scope='event_hawkes')
output_length = outputs.shape[2].value
state_length = final_state.shape[1].value
if output_length == state_length:
# 不需要做任何事情
event_final_state = final_state
elif output_length * 2 == state_length:
event_final_state = tf.split(final_state, 2, axis=1)[0]
else:
raise ValueError('Invalid Size')
with tf.name_scope('state_concat'):
concat_final_state = tf.concat(
[event_final_state, context_final_state], axis=1)
with tf.variable_scope('output_para'):
output_weight = tf.get_variable(
"weight", [num_hidden * 2, 1],
initializer=tf.initializers.orthogonal())
bias = tf.get_variable('bias', [])
with tf.name_scope('prediction'):
unnormalized_prediction = tf.matmul(concat_final_state,
output_weight) + bias
prediction = tf.sigmoid(unnormalized_prediction)
with tf.name_scope('loss'):
with tf.name_scope('pred_loss'):
loss_pred = tf.losses.sigmoid_cross_entropy(
logits=unnormalized_prediction,
multi_class_labels=y_placeholder)
with tf.name_scope('dae_loss'):
if autoencoder_length > 0:
loss_dae = autoencoder.autoencoder_loss(
embedding=processed_input,
origin_input=context_placeholder,
weight=autoencoder_weight)
else:
loss_dae = 0
with tf.name_scope('loss_sum'):
loss = loss_pred + loss_dae * dae_weight
return loss, prediction, event_placeholder, context_placeholder, y_placeholder, batch_size, phase_indicator, \
base_intensity, mutual_intensity, time_list, task_index, sequence_length, concat_final_state