def selector_fn(outputs):
if scope_name is not None:
with variable_scope.variable_scope("%s/selector" % scope_name,
reuse=reuse) as scope:
selector_logits = layers.linear(outputs, 1, scope=scope)
else:
with variable_scope.variable_scope("selector",
reuse=reuse) as scope:
selector_logits = layers.linear(outputs, 1, scope=scope)
return selector_logits
def prepare_attention(attention_states,
attention_option,
num_units,
imem=None,
output_alignments=False,
reuse=False):
# Prepare attention keys / values from attention_states
with variable_scope.variable_scope("attention_keys", reuse=reuse) as scope:
attention_keys = layers.linear(
attention_states, num_units, biases_initializer=None, scope=scope)
attention_values = attention_states
if imem is not None:
if type(imem) is tuple:
with variable_scope.variable_scope("imem_graph", reuse=reuse) as scope:
attention_keys2, attention_states2 = array_ops.split(layers.linear(
imem[0], num_units*2, biases_initializer=None, scope=scope), [num_units, num_units], axis=2)
with variable_scope.variable_scope("imem_triple", reuse=reuse) as scope:
attention_keys3, attention_states3 = array_ops.split(layers.linear(
imem[1], num_units*2, biases_initializer=None, scope=scope), [num_units, num_units], axis=3)
attention_keys = (attention_keys, attention_keys2, attention_keys3)
attention_values = (attention_states, attention_states2, attention_states3)
else:
with variable_scope.variable_scope("imem", reuse=reuse) as scope:
attention_keys2, attention_states2 = array_ops.split(layers.linear(
imem, num_units*2, biases_initializer=None, scope=scope), [num_units, num_units], axis=2)
attention_keys = (attention_keys, attention_keys2)
attention_values = (attention_states, attention_states2)
# Attention score function
if imem is None:
attention_score_fn = _create_attention_score_fn("attention_score", num_units,
attention_option, reuse)
else:
attention_score_fn = (_create_attention_score_fn("attention_score", num_units,
attention_option, reuse),
_create_attention_score_fn("imem_score", num_units,
"luong", reuse, output_alignments=output_alignments))
# Attention construction function
attention_construct_fn = _create_attention_construct_fn("attention_construct",
num_units,
attention_score_fn,
reuse)
return (attention_keys, attention_values, attention_score_fn,
attention_construct_fn)
def output_fn(outputs):
if scope_name is not None:
with variable_scope.variable_scope("%s/output_projection" %
scope_name,
reuse=reuse) as scope:
output_logits = layers.linear(outputs,
num_symbols,
scope=scope)
else:
with variable_scope.variable_scope("output_projection",
reuse=reuse) as scope:
output_logits = layers.linear(outputs,
num_symbols,
scope=scope)
return output_logits
def project_fn(input):
output = layers.linear(input,
vocabulary_count,
scope="projection_layer")
softmaxed_probability = tf.nn.softmax(
output) # batch_size*decoder_len*vocabulary_count
return softmaxed_probability
def condition_tensor(tensor, conditioning):
"""Condition the value of a tensor.
Conditioning scheme based on https://arxiv.org/abs/1609.03499.
Args:
tensor: A minibatch tensor to be conditioned.
conditioning: A minibatch Tensor of to condition on. Must be 2D, with first
dimension the same as `tensor`.
Returns:
`tensor` conditioned on `conditioning`.
Raises:
ValueError: If the non-batch dimensions of `tensor` aren't fully defined.
ValueError: If `conditioning` isn't at least 2D.
ValueError: If the batch dimension for the input Tensors don't match.
"""
tensor.shape[1:].assert_is_fully_defined()
num_features = tensor.shape[1:].num_elements()
if conditioning.shape.ndims < 2:
raise ValueError('conditioning must be at least 2D, but saw shape: %s'
% conditioning.shape)
mapped_conditioning = layers.linear(
layers.flatten(conditioning), num_features)
if not mapped_conditioning.shape.is_compatible_with(tensor.shape):
mapped_conditioning = array_ops.reshape(
mapped_conditioning, _get_shape(tensor))
return tensor + mapped_conditioning
def sequence_loss(
outputs, # 解码器输出
targets, # 标签
masks): # 对标签的 mask
with tf.variable_scope('decoder_rnn'):
# 预测值
logits = layers.linear(
outputs, num_symbols,
scope=name) # [batch_size, decoder_len, num_symbols]
logits = tf.reshape(
logits,
[-1, num_symbols]) # [batch_size*decoder_len, num_symbols]
# 标签
local_labels = tf.reshape(targets,
[-1]) # [batch_size*decoder_len]
local_masks = tf.reshape(masks, [-1]) # [batch_size*decoder_len]
# 计算损失
local_loss = tf.nn.sparse_softmax_cross_entropy_with_logits(
labels=local_labels, logits=logits) # [batch_size*decoder_len]
# 序列长度外的部分不计算损失
local_loss = local_loss * local_masks
loss = tf.reduce_sum(local_loss) # 序列的总损失 [batch_size*decoder_len]
total_size = tf.reduce_sum(local_masks) # 序列的总长度
total_size += 1e-12 # 避免总长度为0
return loss, loss / total_size # 一个 batch 的数据的每个单词的平均损失
def construct_fn(attention_query, attention_keys, attention_values):
context = attention_score_fn(attention_query, attention_keys,
attention_values)
concat_input = array_ops.concat(1, [attention_query, context])
attention = layers.linear(
concat_input, num_units, biases_initializer=None, scope=scope)
return attention
def loss_fn(decoder_output, label_id, mask):
'''
:param decoder_output: [batch_size decoder_len num_units]
:param label_id: batch_size,decoder_len
:param mask: [batch_size,decoder_len]
:return:
'''
with tf.variable_scope("decoder_rnn"):
softmaxed_probability = layers.linear(
decoder_output, vocabulary_count, scope="projection_layer"
) # batch_size decoder_len vocabulary_count
logits = tf.reshape(
softmaxed_probability,
[-1, vocabulary_count
]) # 二维[batch_size*decoder_len, vovabulary_count]
labels = tf.reshape(label_id, [-1]) # [batch_size*decoder_len]
label_mask = tf.reshape(mask, [-1]) # [batch_size*decoder_len]
'''
logits是神经网络输出层的输出，shape为[batch_size,num_classes]
label是一个一维向量，长度为batch_size，每个元素取值区间是[0,num_classes)，其实每一个值就是代表了batch中对应样本的类别
tf.nn.sparse_softmax_cross_entropy_with_logits该函数先计算logits的softmax值，再计算softmax与label的交叉熵损失
因此传入的logits无须提前softmax
'''
local_loss = tf.nn.sparse_softmax_cross_entropy_with_logits(
labels=labels, logits=logits) # [batch_size*decoder_len]
total_size = tf.reduce_sum(
label_mask) # batch_size个response的总长度（不算padding部分）
total_size += 1e-12 # 避免总长度为0
loss = tf.reduce_sum(local_loss) # batch_size个response的总损失
avg_loss = loss / total_size # 每个单词的平均损失
return loss, avg_loss
def prepare_attention(
encoder_output, # 编码器输出 [batch_size, encoder_len, num_units]
num_units,
attention_option="bahdanau",
output_alignments=False,
reuse=False):
# 根据编码器的输出，构造注意力的 keys 和 values
with tf.variable_scope("attention_keys", reuse=reuse) as scope:
attention_keys = layers.linear(encoder_output,
num_units,
biases_initializer=None,
scope=scope)
attention_values = encoder_output
attention_score_fn = create_attention_score_fn(
num_units,
attention_option=attention_option,
output_alignments=output_alignments,
reuse=reuse)
attention_construct_fn = create_attention_construct_fn(num_units,
attention_score_fn,
reuse=reuse)
return (attention_keys, attention_values, attention_score_fn,
attention_construct_fn)
def construct_fn(attention_query, attention_keys, attention_values):
alignments = None
if type(attention_score_fn) is tuple:
context0 = attention_score_fn[0](attention_query, attention_keys[0],
attention_values[0])
if len(attention_keys) == 2:
context1 = attention_score_fn[1](attention_query, attention_keys[1],
attention_values[1])
elif len(attention_keys) == 3:
context1 = attention_score_fn[1](attention_query, attention_keys[1:],
attention_values[1:])
if type(context1) is tuple:
if len(context1) == 2:
context1, alignments = context1
concat_input = array_ops.concat([attention_query, context0, context1], 1)
elif len(context1) == 3:
context1, context2, alignments = context1
concat_input = array_ops.concat([attention_query, context0, context1, context2], 1)
else:
concat_input = array_ops.concat([attention_query, context0, context1], 1)
else:
context = attention_score_fn(attention_query, attention_keys,
attention_values)
concat_input = array_ops.concat([attention_query, context], 1)
attention = layers.linear(
concat_input, num_units, biases_initializer=None, scope=scope)
if alignments is None:
return attention
else:
return attention, alignments
def condition_tensor(tensor, conditioning):
"""Condition the value of a tensor.
Conditioning scheme based on https://arxiv.org/abs/1609.03499.
Args:
tensor: A minibatch tensor to be conditioned.
conditioning: A minibatch Tensor of to condition on. Must be 2D, with first
dimension the same as `tensor`.
Returns:
`tensor` conditioned on `conditioning`.
Raises:
ValueError: If the non-batch dimensions of `tensor` aren't fully defined.
ValueError: If `conditioning` isn't at least 2D.
ValueError: If the batch dimension for the input Tensors don't match.
"""
tensor.shape[1:].assert_is_fully_defined()
num_features = tensor.shape[1:].num_elements()
if conditioning.shape.ndims < 2:
raise ValueError(
'conditioning must be at least 2D, but saw shape: %s' %
conditioning.shape)
mapped_conditioning = layers.linear(layers.flatten(conditioning),
num_features)
if not mapped_conditioning.shape.is_compatible_with(tensor.shape):
mapped_conditioning = array_ops.reshape(mapped_conditioning,
_get_shape(tensor))
return tensor + mapped_conditioning
def construct_fn(attention_query, attention_keys, attention_values):
context = attention_score_fn(attention_query, attention_keys,
attention_values)
concat_input = array_ops.concat([attention_query, context], 1)
attention = layers.linear(
concat_input, num_units, biases_initializer=None, scope=scope)
return attention
def total_loss(outputs, targets, masks, alignments, triples_embedding, use_entities, entity_targets):
'''
outputs: batch_size * decoder_len * num_units
targets: batch_size * decoder_len
masks: batch_size * decoder_len
alignments: [batch_size, decoder_len, triple_num, triple_len] 注意力系数
triples_embedding: [batch_size, triple_num, triple_len, 3*num_trans_units]
use_entities: batch_size * decoder_len 用 1 标注了回复的哪个时间步用了三元组
entity_targets: [batch_size, decoder_len, triple_num, triple_len] 用 1 标注了每个 batch 每个时间步用了哪个图的哪个三元组
'''
batch_size = tf.shape(outputs)[0]
local_masks = tf.reshape(masks, [-1])
logits = layers.linear(outputs, num_symbols, scope='decoder_rnn/%s' % name) # batch_size * decoder_len * num_symbols
one_hot_targets = tf.one_hot(targets, num_symbols) # batch_size * decoder_len * num_symbols
# 每一步的单词预测为 target 的概率
word_prob = tf.reduce_sum(tf.nn.softmax(logits) * one_hot_targets, axis=2) # batch_size * decoder_len
# 对三元组选择概率的系数，论文中的 gamma_t
selector = tf.squeeze(tf.sigmoid(layers.linear(outputs, 1, scope='decoder_rnn/selector'))) # batch_size * decoder_len
# 每一步对的三元组注意力为实际采用的三元组的概率
triple_prob = tf.reduce_sum(alignments * entity_targets, axis=[2, 3]) # batch_size * decoder_len
#
ppx_prob = word_prob * (1 - use_entities) + triple_prob * use_entities # batch_size * decoder_len
# 论文中的 P(y_t)
final_prob = word_prob * (1 - selector) * (1 - use_entities) + triple_prob * selector * use_entities # batch_size * decoder_len
final_loss = tf.reduce_sum(tf.reshape( - tf.log(1e-12 + final_prob), [-1]) * local_masks) # value
#
ppx_loss = tf.reduce_sum(tf.reshape( - tf.log(1e-12 + ppx_prob), [-1]) * local_masks) # value
# 每个 batch 的 ppx
sentence_ppx = tf.reduce_sum(tf.reshape(tf.reshape( - tf.log(1e-12 + ppx_prob), [-1]) * local_masks, [batch_size, -1]), axis=1) # [batch_size]
selector_loss = tf.reduce_sum(tf.reshape( - tf.log(1e-12 + selector * use_entities + (1 - selector) * (1 - use_entities)), [-1]) * local_masks)
loss = final_loss + selector_loss
total_size = tf.reduce_sum(local_masks)
total_size += 1e-12 # to avoid division by 0 for all-0 weights
# 每个词的平均损失，每个词的平均 ppx，每个句子的平均 ppx
return loss / total_size, ppx_loss / total_size, sentence_ppx / tf.reduce_sum(masks, axis=1)
def total_loss(outputs, # [batch_size, decoder_len, num_units]
targets, # [batch_size, decoder_len]
masks, # [batch_size, decoder_len]
alignments, # [batch_size, decoder_len, triple_num, triple_len]
triples_embedding,
use_entities, # [batch_size, decoder_len] 用1标注了回复的哪个时间步用了三元组
entity_targets): # [batch_size, decoder_len, triple_num, triple_len] 用1标注了每个batch每个时间步用了哪个图的哪个三元组
batch_size = tf.shape(outputs)[0]
local_masks = tf.reshape(masks, [-1]) # [batch_size*decoder_len]
logits = layers.linear(outputs, num_symbols, scope='decoder_rnn/%s' % name) # [batch_size, decoder_len, num_symbols]
one_hot_targets = tf.one_hot(targets, num_symbols) # [batch_size, decoder_len, num_symbols]
# 每一步的单词预测正确的概率
word_prob = tf.reduce_sum(tf.nn.softmax(logits) * one_hot_targets, axis=2) # [batch_size, decoder_len]
# 每一步使用实体词的概率预测
selector = tf.squeeze(tf.sigmoid(layers.linear(outputs, 1, scope='decoder_rnn/selector'))) # [batch_size, decoder_len]
# 每一步对使用的三元组的注意力
triple_prob = tf.reduce_sum(alignments * entity_targets, axis=[2, 3]) # [batch_size, decoder_len]
# 每一步正确的概率
ppx_prob = word_prob * (1 - use_entities) + triple_prob * use_entities # [batch_size, decoder_len]
# 加上选择器选择正确的概率
final_prob = word_prob * (1 - selector) * (1 - use_entities) + triple_prob * selector * use_entities # [batch_size, decoder_len]
# 加上选择器选择的损失
final_loss = tf.reduce_sum(tf.reshape(- tf.log(1e-12 + final_prob), [-1]) * local_masks)
# 不加选择器的ppx
ppx_loss = tf.reduce_sum(tf.reshape( - tf.log(1e-12 + ppx_prob), [-1]) * local_masks)
# 每个batch的ppx
sentence_ppx = tf.reduce_sum(tf.reshape(tf.reshape(-tf.log(1e-12 + ppx_prob), [-1]) * local_masks, [batch_size, -1]), axis=1) # [batch_size]
# 选择器的损失
selector_loss = tf.reduce_sum(tf.reshape(-tf.log(1e-12 + selector * use_entities + (1 - selector) * (1 - use_entities)), [-1]) * local_masks) # [batch_size]
loss = final_loss + selector_loss
total_size = tf.reduce_sum(local_masks)
total_size += 1e-12
# 每个词的平均损失、每个词的平均ppx、每个样本的每个词的ppx[batch_size]
return loss / total_size, ppx_loss / total_size, sentence_ppx / tf.reduce_sum(masks, axis=1)
def prepare_multistep_attention(encoder_states,
decoder_reprs,
kd_states1,
kd_states2,
attention_option,
num_units,
reuse=False):
# Prepare attention keys / values from attention_states
with variable_scope.variable_scope("attn_keys", reuse=reuse) as scope:
attention_keys1 = layers.linear(encoder_states,
num_units,
biases_initializer=None,
scope=scope)
attention_values1 = encoder_states
# Attention scoring function
attention_score_fn1 = _create_attention_score_fn(
"attn_score", num_units, attention_option, reuse)
with variable_scope.variable_scope("attn_reprs", reuse=reuse) as scope:
if decoder_reprs is not None:
attention_keys2 = layers.linear(decoder_reprs,
num_units,
biases_initializer=None,
scope=scope)
else:
attention_keys2 = None
attention_values2 = decoder_reprs
# Attention scoring function
attention_score_fn2 = _create_attention_score_fn(
"attn_score", num_units, attention_option, reuse)
attention_keys = (attention_keys1, attention_keys2)
if kd_states1 is not None and kd_states2 is not None:
attention_values = (attention_values1, attention_values2, kd_states1,
kd_states2)
else:
attention_values = (attention_values1, attention_values2, None, None)
attention_score_fn = (attention_score_fn1, attention_score_fn2)
# Attention construction function
attention_construct_fn = _create_attention_construct_fn(
"attn_construct_multi", num_units, attention_score_fn, reuse)
return attention_keys, attention_values, attention_construct_fn
def inference_fn(inference_output):
with tf.variable_scope("decoder_rnn"):
inference_softmaxed_probability = tf.nn.softmax(
layers.linear(inference_output,
vocabulary_count,
scope="projection_layer")
) # 词汇表softmaxed后的概率 [batch_size decoder_len vovabulary_count]
inference_maximum_likelihood_id = tf.argmax(
inference_softmaxed_probability, axis=2)
return inference_maximum_likelihood_id
def construct_fn(attention_query, attention_keys, attention_values):
context, alignments = attention_score_fn(attention_query,
attention_keys,
attention_values)
concat_input = array_ops.concat([attention_query, context], 1)
concat_input = array_ops.reshape(concat_input, [-1, 1024])
attention = layers.linear(concat_input,
num_units,
biases_initializer=None,
scope=scope)
return attention, alignments
def project_fn(input):
'''
如果input是batch_size*num_units,创建权重矩阵 W:num_units*vocabulary_count
output = input * W batch_size * vocabulary_count
如果input是batch_size, decoder-len, num_units，则output是batch_size, decoder-len, vocabulary_count
'''
output = layers.linear(
input, vocabulary_count,
scope="projection_layer") # batch_size * vocabulary_count
softmaxed_probability = tf.nn.softmax(
output) # batch_size * vocabulary_count
return softmaxed_probability
def construct_fn(attention_query, attention_keys, attention_values):
'''拼接几个注意力
返回:
attention: [batch_size, num_units] 拼接完几个注意力并做一次线性变化得到的输出
alignments: 注意力系数
'''
alignments = None
# 如果有静态图或者三元组
if type(attention_score_fn) is tuple:
# 计算编码器每一步输出的注意力
context0 = attention_score_fn[0](attention_query,
attention_keys[0],
attention_values[0])
# 如果只有静态图
if len(attention_keys) == 2:
context1 = attention_score_fn[1](attention_query,
attention_keys[1],
attention_values[1])
# 如果既有静态图还有三元组
elif len(attention_keys) == 3:
context1 = attention_score_fn[1](attention_query,
attention_keys[1:],
attention_values[1:])
if type(context1) is tuple:
# 只有静态图且要求输出对齐
if len(context1) == 2:
context1, alignments = context1
concat_input = array_ops.concat(
[attention_query, context0, context1], 1)
# 既有静态图还有三元组
elif len(context1) == 3:
context1, context2, alignments = context1
concat_input = array_ops.concat(
[attention_query, context0, context1, context2], 1)
else: # 存在静态图没有三元组且不要求输出静态图的对齐的情况
concat_input = array_ops.concat(
[attention_query, context0, context1], 1)
# 如果没有静态图或者三元组
else:
context = attention_score_fn(attention_query, attention_keys,
attention_values)
concat_input = array_ops.concat([attention_query, context], 1)
# 给拼接完的注意力做一个线性变化 [batch_size, num_units]
attention = layers.linear(concat_input,
num_units,
biases_initializer=None,
scope=scope)
if alignments is None:
return attention
else:
return attention, alignments
def total_loss(outputs, targets, masks, alignments, triples_embedding, use_entities, entity_targets):
local_masks = tf.reshape(masks, [-1])
logits = layers.linear(outputs, num_symbols, scope='decoder_rnn/%s' % name)
one_hot_targets = tf.one_hot(targets, num_symbols)
word_prob = tf.reduce_sum(tf.nn.softmax(logits) * one_hot_targets, axis=2)
selector = tf.squeeze(tf.sigmoid(layers.linear(outputs, 1, scope='decoder_rnn/selector')))
triple_prob = tf.reduce_sum(alignments * entity_targets, axis=[2])
cast_selector = tf.cast(tf.reduce_sum(alignments, axis=2) > tf.reduce_sum(tf.nn.softmax(logits), axis=2), tf.float32)
final_prob = word_prob * (1 - selector) + triple_prob * selector
ppx_prob = word_prob * (1 - use_entities) + triple_prob * use_entities
final_loss = tf.reshape( - tf.log(1e-12 + final_prob), [-1]) * local_masks
ppx_loss = tf.reshape( - tf.log(1e-12 + ppx_prob), [-1]) * local_masks
sentence_ppx = tf.reduce_sum( - tf.log(1e-12 + ppx_prob) * masks, axis=1)
loss = tf.reduce_sum(final_loss)
#loss = tf.Print(loss, ['use_entity', tf.reduce_min(use_entities), tf.reduce_max(use_entities), 'triple_prob',tf.reduce_min(triple_prob), 'word_prob', tf.reduce_min(word_prob), 'final_prob', tf.reduce_min(final_prob), 'final_loss', tf.reduce_min(final_loss)], summarize=1e6)
total_size = tf.reduce_sum(local_masks)
total_size += 1e-12 # to avoid division by 0 for all-0 weights
return loss / total_size, tf.reduce_sum(ppx_loss) / total_size, sentence_ppx / tf.reduce_sum(masks, axis=1)
def total_loss(outputs, targets, masks, alignments, triples_embedding, use_entities, entity_targets):
batch_size = tf.shape(outputs)[0]
local_masks = tf.reshape(masks, [-1])
logits = layers.linear(outputs, num_symbols, scope='decoder_rnn/%s' % name)
one_hot_targets = tf.one_hot(targets, num_symbols)
word_prob = tf.reduce_sum(tf.nn.softmax(logits) * one_hot_targets, axis=2)
selector = tf.squeeze(tf.sigmoid(layers.linear(outputs, 1, scope='decoder_rnn/selector')))
triple_prob = tf.reduce_sum(alignments * entity_targets, axis=[2, 3])
ppx_prob = word_prob * (1 - use_entities) + triple_prob * use_entities
final_prob = word_prob * (1 - selector) * (1 - use_entities) + triple_prob * selector * use_entities
final_loss = tf.reduce_sum(tf.reshape( - tf.log(1e-12 + final_prob), [-1]) * local_masks)
ppx_loss = tf.reduce_sum(tf.reshape( - tf.log(1e-12 + ppx_prob), [-1]) * local_masks)
sentence_ppx = tf.reduce_sum(tf.reshape(tf.reshape( - tf.log(1e-12 + ppx_prob), [-1]) * local_masks, [batch_size, -1]), axis=1)
selector_loss = tf.reduce_sum(tf.reshape( - tf.log(1e-12 + selector * use_entities + (1 - selector) * (1 - use_entities)), [-1]) * local_masks)
loss = final_loss + selector_loss
total_size = tf.reduce_sum(local_masks)
total_size += 1e-12 # to avoid division by 0 for all-0 weights
return loss / total_size, ppx_loss / total_size, sentence_ppx / tf.reduce_sum(masks, axis=1)
def construct_fn(attention_query, attention_keys, attention_values):
'''拼接计算完的上下文
返回:
attention: [batch_size, num_units] 注意力的上下文拼接，拼接完做了一次线性变化的
alignments: 注意力系数
'''
alignments = None
# 如果有图向量和三元组嵌入
if type(attention_score_fn) is tuple:
# 用 bahdanau 计算带注意力的上下文
context0 = attention_score_fn[0](attention_query,
attention_keys[0],
attention_values[0])
# 训练没走这个分支
if len(attention_keys) == 2:
context1 = attention_score_fn[1](attention_query,
attention_keys[1],
attention_values[1])
# 如果有图向量和三元组嵌入
elif len(attention_keys) == 3:
context1 = attention_score_fn[1](attention_query,
attention_keys[1:],
attention_values[1:])
if type(context1) is tuple:
if len(context1) == 2:
context1, alignments = context1
concat_input = array_ops.concat(
[attention_query, context0, context1], 1)
elif len(context1) == 3: # 训练走的这个分支
context1, context2, alignments = context1
concat_input = array_ops.concat(
[attention_query, context0, context1, context2], 1
) # [batch_size, num_units*4] 解码器输出，上下文，图上下文，三元组上下文拼接
else:
concat_input = array_ops.concat(
[attention_query, context0, context1], 1)
else:
context = attention_score_fn(attention_query, attention_keys,
attention_values)
concat_input = array_ops.concat([attention_query, context], 1)
# 给拼接完的分布做一个线性变化，将最后一个维度转为 num_units
attention = layers.linear(concat_input,
num_units,
biases_initializer=None,
scope=scope) # [batch_size, num_units]
if alignments is None:
return attention
else:
return attention, alignments
def construct_fn(attention_query, attention_keys, attention_values):
if isinstance(attention_score_fn, tuple): # multi-step decoding
attention_score_fn1, attention_score_fn2 = attention_score_fn
attention_keys1, attention_keys2 = attention_keys
attention_values1, decoder_reprs, kd_states1, kd_states2 = attention_values
context1 = attention_score_fn1(attention_query,
attention_keys1,
attention_values1)
if kd_states1 is None or kd_states2 is None:
context2 = attention_score_fn2(attention_query,
attention_keys2,
decoder_reprs)
concat_input = array_ops.concat(
[attention_query, context1, context2], 1)
else:
if decoder_reprs is None:
print("concat=3")
concat_input = array_ops.concat([
attention_query, context1, kd_states1, kd_states2
], 1)
else:
print("concat=4")
context2 = attention_score_fn2(attention_query,
attention_keys2,
decoder_reprs)
concat_input = array_ops.concat([
attention_query, context1, context2, kd_states1,
kd_states2
], 1)
else: # only one step decoding
if isinstance(attention_values, tuple):
attention_values1, kd_state = attention_values
context1 = attention_score_fn(attention_query,
attention_keys,
attention_values1)
concat_input = array_ops.concat(
[attention_query, context1, kd_state], 1)
else:
context = attention_score_fn(attention_query,
attention_keys,
attention_values)
concat_input = array_ops.concat([attention_query, context],
1)
attention = layers.linear(concat_input,
num_units,
biases_initializer=None,
scope=scope)
return attention
def sequence_loss(outputs, targets, masks):
with variable_scope.variable_scope('decoder_rnn'):
logits = layers.linear(outputs, num_symbols, scope=name) # [batch_size, decoder_len, num_symbols]
logits = tf.reshape(logits, [-1, num_symbols]) # [batch_size*decoder_len, num_symbols]
local_labels = tf.reshape(targets, [-1]) # [batch_size*decoder_len]
local_masks = tf.reshape(masks, [-1]) # [batch_size*decoder_len]
local_loss = tf.nn.sparse_softmax_cross_entropy_with_logits(labels=local_labels, logits=logits)
local_loss = local_loss * local_masks # 序列长度外的部分不计算损失
loss = tf.reduce_sum(local_loss) # 序列的总损失
total_size = tf.reduce_sum(local_masks) # 序列的总长度
total_size += 1e-12 # 避免总长度为0
return loss / total_size # 返回平均损失
def construct_fn(attention_query, attention_keys, attention_values):
context, scores, alignments = attention_score_fn(
attention_query, attention_keys, attention_values)
concat_input = array_ops.concat([attention_query, context], 1)
#NOTICE! here pass scope which is outside construct_fn func! so always not affected by using env scope, be like
#seq2seq/main/decode/attention_construct/weights in rnn_decoder.py not seq2seq/main/decode/rnn/loop_function/weights in beam_decoder.py
attention = layers.linear(concat_input,
num_units,
biases_initializer=None,
scope=scope)
#add this to make it safe using in loop, if only used in dynamic decode do not need this since only call above linear once
#this kind of witting then add reuse seems safe and elegant
#http://stackoverflow.com/questions/38545362/tensorflow-variable-scope-reuse-if-variable-exists
scope.reuse_variables()
return attention, scores, alignments
def sequence_loss(outputs, targets, masks):
with variable_scope.variable_scope('decoder_rnn'):
logits = layers.linear(outputs, num_symbols, scope=name)
logits = tf.reshape(logits, [-1, num_symbols])
local_labels = tf.reshape(targets, [-1])
local_masks = tf.reshape(masks, [-1])
local_loss = tf.nn.sparse_softmax_cross_entropy_with_logits(labels=local_labels, logits=logits)
local_loss = local_loss * local_masks
loss = tf.reduce_sum(local_loss)
total_size = tf.reduce_sum(local_masks)
total_size += 1e-12 # to avoid division by 0 for all-0 weights
return loss / total_size
def attention_construct_fn(
query, # 解码器输出 [batch_size, num_units]
keys, # [batch_size, encoder_len, num_units]
values): # [batch_size, encoder_len, num_units]
alignments = None
context = attention_score_fn(query, keys, values)
if type(context) is tuple:
context, alignments = context
concat_input = tf.concat([query, context], axis=1)
attention = layers.linear(concat_input,
num_units,
biases_initializer=None,
scope=scope) # [batch_size, num_units]
if alignments is None:
return attention # [batch_size, num_units]
else:
return attention, alignments
def loss_fn(decoder_output, label_id, mask):
with tf.variable_scope("decoder_rnn"):
softmaxed_probability = tf.nn.softmax(
layers.linear(decoder_output,
vocabulary_count,
scope="projection_layer"))
logits = tf.reshape(
softmaxed_probability,
[-1, vocabulary_count
]) # [batch_size*decoder_len vovabulary_count]
labels = tf.reshape(label_id, [-1]) # [batch_size*decoder_len]
label_mask = tf.reshape(mask, [-1]) # [batch_size*decoder_len]
local_loss = tf.nn.sparse_softmax_cross_entropy_with_logits(
labels=labels, logits=logits) # [batch_size*decoder_len]
total_size = tf.reduce_sum(label_mask)
total_size += 1e-12
loss = tf.reduce_sum(local_loss * label_mask)
avg_loss = loss / total_size
return loss, avg_loss
def prepare_attention(attention_states,
attention_option,
num_units,
reuse=False):
"""Prepare keys/values/functions for attention.
Args:
attention_states: hidden states to attend over.
attention_option: how to compute attention, either "luong" or "bahdanau".
num_units: hidden state dimension.
reuse: whether to reuse variable scope.
Returns:
attention_keys: to be compared with target states.
attention_values: to be used to construct context vectors.
attention_score_fn: to compute similarity between key and target states.
attention_construct_fn: to build attention states.
"""
# Prepare attention keys / values from attention_states
with variable_scope.variable_scope("attention_keys", reuse=reuse) as scope:
attention_keys = layers.linear(attention_states,
num_units,
biases_initializer=None,
scope=scope)
#well here do not need ad scope reuse since you get attention_keys and use it not call this linear again,
#attention_construct_fn is different, since you call that function each step during loop
attention_values = attention_states
# Attention score function
attention_score_fn = _create_attention_score_fn("attention_score",
num_units,
attention_option, reuse)
# Attention construction function
attention_construct_fn = _create_attention_construct_fn(
"attention_construct", num_units, attention_score_fn, reuse)
return (attention_keys, attention_values, attention_score_fn,
attention_construct_fn)
def prepare_attention(attention_states,
attention_len,
attention_option,
num_units,
reuse=False):
# Prepare attention keys / values from attention_states
with variable_scope.variable_scope("attention_keys", reuse=reuse) as scope:
attention_keys = layers.linear(
attention_states, num_units, biases_initializer=None, scope=scope)
attention_values = attention_states
# Attention score function
attention_score_fn = _create_attention_score_fn(
"attention_score", attention_len, num_units, attention_option, reuse)
# Attention construction function
attention_construct_fn = _create_attention_construct_fn(
"attention_construct", num_units, attention_score_fn, reuse)
return (attention_keys, attention_values, attention_score_fn,
attention_construct_fn)
def prepare_attention(attention_states,
attention_option,
num_units,
reuse=False):
"""Prepare keys/values/functions for attention.
Args:
attention_states: hidden states to attend over.
attention_option: how to compute attention, either "luong" or "bahdanau".
num_units: hidden state dimension.
reuse: whether to reuse variable scope.
Returns:
attention_keys: to be compared with target states.
attention_values: to be used to construct context vectors.
attention_score_fn: to compute similarity between key and target states.
attention_construct_fn: to build attention states.
"""
# Prepare attention keys / values from attention_states
with variable_scope.variable_scope("attention_keys", reuse=reuse) as scope:
attention_keys = layers.linear(attention_states,
num_units,
biases_initializer=None,
scope=scope)
#attention_keys = tf.Print(attention_keys, ["attention_keys", tf.shape(attention_states), tf.shape(attention_keys), attention_keys])
attention_values = attention_states
# Attention score function
attention_score_fn = _create_attention_score_fn("attention_score",
num_units,
attention_option, reuse)
# Attention construction function
attention_construct_fn = _create_attention_construct_fn(
"attention_construct", num_units, attention_score_fn, reuse)
return (attention_keys, attention_values, attention_score_fn,
attention_construct_fn)
def prepare_attention(attention_states,
attention_option,
num_units,
reuse=False):
"""Prepare keys/values/functions for attention.
Args:
attention_states: hidden states to attend over.
attention_option: how to compute attention, either "luong" or "bahdanau".
num_units: hidden state dimension.
reuse: whether to reuse variable scope.
Returns:
attention_keys: to be compared with target states.
attention_values: to be used to construct context vectors.
attention_score_fn: to compute similarity between key and target states.
attention_construct_fn: to build attention states.
"""
# Prepare attention keys / values from attention_states
with variable_scope.variable_scope("attention_keys", reuse=reuse) as scope:
attention_keys = layers.linear(
attention_states, num_units, biases_initializer=None, scope=scope)
attention_values = attention_states
# Attention score function
attention_score_fn = _create_attention_score_fn("attention_score", num_units,
attention_option, reuse)
# Attention construction function
attention_construct_fn = _create_attention_construct_fn("attention_construct",
num_units,
attention_score_fn,
reuse)
return (attention_keys, attention_values, attention_score_fn,
attention_construct_fn)
def attention(query):
"""Point on hidden using hidden_features and query."""
with vs.variable_scope("Attention"):
v = vs.get_variable("AttnV", [num_units])
#[batch_size, num_units] -> [batch_size, num_units]
processed_query = layers.linear(query, num_units, scope="query_layer")
#processed_query = layers_core.dense(query, num_units, use_bias=False, name="query_layer")
#->[batch_size, 1, num_units]
processed_query = tf.expand_dims(processed_query, 1)
#[batch_size, attn_length, num_units] + [batch_size, 1, num_units] -> [batch_size, attn_length, num_units]
#reduce_sum -> [batch_size, attn_length]
scores = math_ops.reduce_sum(v * math_ops.tanh(keys + processed_query), [2])
if feed_prev:
alignments = nn_ops.softmax(scores)
#-> [batch_size, attn_length, 1]
alignments = array_ops.expand_dims(alignments, 2)
#context_vector = math_ops.reduce_sum(alignments * values, [1])
#return scores, alignments, context_vector
return scores, alignments
else:
return scores