def cqa_model(final_hidden):
final_hidden_shape = modeling.get_shape_list(final_hidden, expected_rank=3)
batch_size = final_hidden_shape[0]
seq_length = final_hidden_shape[1]
hidden_size = final_hidden_shape[2]
output_weights = tf.get_variable(
"cls/cqa/output_weights", [2, hidden_size],
initializer=tf.truncated_normal_initializer(stddev=0.02))
output_bias = tf.get_variable(
"cls/cqa/output_bias", [2], initializer=tf.zeros_initializer())
final_hidden_matrix = tf.reshape(final_hidden, [batch_size * seq_length, hidden_size])
logits = tf.matmul(final_hidden_matrix, output_weights, transpose_b=True)
logits = tf.nn.bias_add(logits, output_bias)
logits = tf.reshape(logits, [batch_size, seq_length, 2])
logits = tf.transpose(logits, [2, 0, 1])
unstacked_logits = tf.unstack(logits, axis=0)
(start_logits, end_logits) = (unstacked_logits[0], unstacked_logits[1])
return (start_logits, end_logits)
def gather_indexes(sequence_tensor, positions):
"""Gathers the vectors at the specific positions over a minibatch."""
sequence_shape = modeling.get_shape_list(sequence_tensor, expected_rank=3)
batch_size = sequence_shape[0]
seq_length = sequence_shape[1]
width = sequence_shape[2]
flat_offsets = tf.reshape(
tf.range(0, batch_size, dtype=tf.int32) * seq_length, [-1, 1])
flat_positions = tf.reshape(positions + flat_offsets, [-1])
flat_sequence_tensor = tf.reshape(sequence_tensor,
[batch_size * seq_length, width])
output_tensor = tf.gather(flat_sequence_tensor, flat_positions)
return output_tensor
def gather_indexes(sequence_tensor, positions):
"""Gathers the vectors at the specific positions over a minibatch."""
sequence_shape = modeling.get_shape_list(sequence_tensor, expected_rank = 3)
batch_size = sequence_shape[0]
seq_length = sequence_shape[1]
width = sequence_shape[2]
flat_offsets = tf.reshape(
tf.range(0, batch_size, dtype = tf.int32) * seq_length, [-1, 1]
)
flat_positions = tf.reshape(positions + flat_offsets, [-1])
flat_sequence_tensor = tf.reshape(
sequence_tensor, [batch_size * seq_length, width]
)
output_tensor = tf.gather(flat_sequence_tensor, flat_positions)
return output_tensor
def create_model(bert_config, is_training, input_ids, input_mask, segment_ids,
use_one_hot_embeddings, scope):
"""Creates a classification model."""
with tf.variable_scope('bert', reuse=tf.AUTO_REUSE) as real_scope:
model = modeling.BertModel(
config=bert_config,
is_training=is_training,
input_ids=input_ids,
input_mask=input_mask,
token_type_ids=segment_ids,
use_one_hot_embeddings=use_one_hot_embeddings,
scope=scope)
final_hidden = model.get_sequence_output()
final_hidden_shape = modeling.get_shape_list(final_hidden, expected_rank=3)
return final_hidden[:, 0, :]
def gather_indexes(sequence_tensor, positions):
"""Gathers the vectors at the specific positions over a minibatch."""
# sequence_tensor, [B, seq_len, hidden_dim]，bert最后一层输出
# positions, [B, masked_token_num]
sequence_shape = modeling.get_shape_list(sequence_tensor, expected_rank=3)
batch_size = sequence_shape[0]
seq_length = sequence_shape[1]
width = sequence_shape[2]
flat_offsets = tf.reshape(
tf.range(0, batch_size, dtype=tf.int32) * seq_length, [-1, 1])
flat_positions = tf.reshape(positions + flat_offsets, [-1]) # 为了下面展开后一次性查找
flat_sequence_tensor = tf.reshape(sequence_tensor,
[batch_size * seq_length, width])
output_tensor = tf.gather(flat_sequence_tensor,
flat_positions) # [B*mask_token_num, hidden_dim]
return output_tensor
def get_lm_output(config, input_tensor, output_weights, label_ids, label_mask):
"""Get loss and log probs for the LM."""
input_shape = modeling.get_shape_list(input_tensor, expected_rank=3)
input_tensor = tf.reshape(
input_tensor, [input_shape[0] * input_shape[1], input_shape[2]])
with tf.variable_scope("cls/predictions"):
# We apply one more non-linear transformation before the output layer.
# This matrix is not used after pre-training.
with tf.variable_scope("transform"):
input_tensor = tf.layers.dense(
input_tensor,
units=config.hidden_size,
activation=modeling.get_activation(config.hidden_act),
kernel_initializer=modeling.create_initializer(
config.initializer_range))
input_tensor = modeling.layer_norm(input_tensor)
# The output weights are the same as the input embeddings, but there is
# an output-only bias for each token.
output_bias = tf.get_variable("output_bias",
shape=[config.vocab_size],
initializer=tf.zeros_initializer())
logits = tf.matmul(input_tensor, output_weights, transpose_b=True)
logits = tf.nn.bias_add(logits, output_bias)
log_probs = tf.nn.log_softmax(logits, axis=-1)
label_ids = tf.reshape(label_ids, [-1])
one_hot_labels = tf.one_hot(label_ids,
depth=config.vocab_size,
dtype=tf.float32)
# The `positions` tensor might be zero-padded (if the sequence is too
# short to have the maximum number of predictions). The `label_weights`
# tensor has a value of 1.0 for every real prediction and 0.0 for the
# padding predictions.
per_example_loss = -tf.reduce_sum(log_probs * one_hot_labels,
axis=[-1])
label_mask = tf.reshape(label_mask, [input_shape[0] * input_shape[1]])
loss_mask = tf.dtypes.cast(label_mask, tf.float32)
per_example_loss = tf.math.multiply(per_example_loss, loss_mask)
loss = tf.reduce_mean(per_example_loss)
return (loss, per_example_loss, log_probs)
def create_model_start(bert_config, is_training, input_ids, input_mask, segment_ids,
use_one_hot_embeddings):
"""Creates a classification model."""
model = modeling.BertModel(
config=bert_config,
is_training=is_training,
input_ids=input_ids,
input_mask=input_mask,
token_type_ids=segment_ids,
use_one_hot_embeddings=use_one_hot_embeddings)
final_hidden = model.get_sequence_output()
final_hidden_shape = modeling.get_shape_list(final_hidden, expected_rank=3)
batch_size = final_hidden_shape[0]
seq_length = final_hidden_shape[1]
hidden_size = final_hidden_shape[2]
output_weights1 = tf.get_variable(
"cls/squad/output_weights1", [384, hidden_size],
initializer=tf.truncated_normal_initializer(stddev=0.02))
output_bias1 = tf.get_variable(
"cls/squad/output_bias1", [384], initializer=tf.zeros_initializer())
final_hidden_matrix = tf.reshape(final_hidden,
[batch_size * seq_length, hidden_size])
keep_prob = 1.0
if is_training:
keep_prob = 0.9
else:
keep_prob = 1.0
logits = tf.matmul(final_hidden_matrix, output_weights1, transpose_b=True)
logits = tf.nn.bias_add(logits, output_bias1)
logits = tf.nn.relu(logits)
logits = tf.nn.dropout(logits, keep_prob)
logits = tf.reshape(logits, [batch_size, seq_length, 384])
logits = tf.transpose(logits, [2, 0, 1])
unstacked_logits = tf.unstack(logits, axis=0)
s = tf.reduce_sum(unstacked_logits[0:384], 0)
return s
def metric_fn(per_example_loss, label_ids, logits, is_real_example):
# batch_size * sequence_len
shape_list = modeling.get_shape_list(label_ids, expected_rank=2)
label_ids = tf.reshape(label_ids, [-1])
is_real_example = tf.tile(is_real_example[:, tf.newaxis], [1, shape_list[1]])
is_real_example = tf.reshape(is_real_example, [-1])
per_example_loss = tf.reshape(per_example_loss, [-1])
logits = tf.reshape(logits, [-1])
predictions = tf.argmax(logits, axis=-1, output_type=tf.int32)
accuracy = tf.metrics.accuracy(
labels=label_ids, predictions=predictions, weights=is_real_example)
loss = tf.metrics.mean(values=per_example_loss, weights=is_real_example)
return {
"eval_accuracy": accuracy,
"eval_loss": loss,
}
def create_model(bert_config, is_training, input_ids, input_mask, segment_ids,
use_one_hot_embeddings):
"""Creates a classification model."""
model = modeling.BertModel(
config=bert_config,
is_training=is_training,
input_ids=input_ids,
input_mask=input_mask,
token_type_ids=segment_ids,
use_one_hot_embeddings=use_one_hot_embeddings)
#out1 model.
#pdb.set_trace()
final_hidden = model.get_sequence_output()
final_hidden_shape = modeling.get_shape_list(final_hidden, expected_rank=3)
batch_size = final_hidden_shape[0]
seq_length = final_hidden_shape[1]
hidden_size = final_hidden_shape[2]
#out2 output_weights
output_weights = tf.get_variable(
"cls/squad/output_weights", [2, hidden_size],
initializer=tf.truncated_normal_initializer(stddev=0.02))
#out3 output_bias
output_bias = tf.get_variable(
"cls/squad/output_bias", [2], initializer=tf.zeros_initializer())
final_hidden_matrix = tf.reshape(final_hidden,
[batch_size * seq_length, hidden_size])
#out4 logits * 2
logits = tf.matmul(final_hidden_matrix, output_weights, transpose_b=True)
logits = tf.nn.bias_add(logits, output_bias)
logits = tf.reshape(logits, [batch_size, seq_length, 2])
logits = tf.transpose(logits, [2, 0, 1])
unstacked_logits = tf.unstack(logits, axis=0)
#out5 start_logits end_logits
(start_logits, end_logits) = (unstacked_logits[0], unstacked_logits[1])
return (start_logits, end_logits)
def create_model(bert_config, is_training, input_ids, input_mask, segment_ids,
labels, num_labels, use_one_hot_embeddings):
"""Creates a classification model."""
model = modeling.BertModel(
config=bert_config,
is_training=is_training,
input_ids=input_ids,
input_mask=input_mask,
token_type_ids=segment_ids,
use_one_hot_embeddings=use_one_hot_embeddings)
# In the demo, we are doing a simple classification task on the entire
# segment.
#
# If you want to use the token-level output, use model.get_sequence_output()
# instead.
output_layer = model.get_sequence_output() #Pooler前一层输出
final_hidden_shape = modeling.get_shape_list(output_layer, expected_rank=3)
hidden_size = final_hidden_shape[-1]
output_weight = tf.get_variable(
"output_weights", [num_labels, hidden_size],
initializer=tf.truncated_normal_initializer(stddev=0.02)
)
output_bias = tf.get_variable(
"output_bias", [num_labels], initializer=tf.zeros_initializer()
)
with tf.variable_scope("loss"):
if is_training:
output_layer = tf.nn.dropout(output_layer, keep_prob=0.9)
output_layer_matrix = tf.reshape(output_layer, [-1, hidden_size]) # [batch_size*seq_length, hidden_size]
logits = tf.matmul(output_layer_matrix, output_weight, transpose_b=True) # [batch_size*seg_length, num_label]
logits = tf.nn.bias_add(logits, output_bias)
logits = tf.reshape(logits, [-1, FLAGS.max_seq_length, num_labels]) # 最后一层输出[batch_size, seq_length, num_label]
# mask = tf.cast(input_mask,tf.float32)
# loss = tf.contrib.seq2seq.sequence_loss(logits,labels,mask)
# return (loss, logits, predict)
##########################################################################
log_probs = tf.nn.log_softmax(logits, axis=-1) # 对输出的softmax归一化
one_hot_labels = tf.one_hot(labels, depth=num_labels, dtype=tf.float32)
per_example_loss = -tf.reduce_sum(one_hot_labels * log_probs, axis=-1) # 每个样本的损失值
loss = tf.reduce_mean(per_example_loss) # 总的损失值
predict = tf.argmax(log_probs, axis=-1) # 预测值
return (loss, per_example_loss, logits, predict)
def create_model(bert_config, is_training, input_ids, input_mask, segment_ids,
use_one_hot_embeddings):
"""Creates a classification model."""
#初始化bert模型参数
model = modeling.BertModel(
config=bert_config,
is_training=is_training,
input_ids=input_ids, #输入格式转换后的队列 “”
input_mask=input_mask, #屏蔽
token_type_ids=segment_ids, #输入的句子分段ID
use_one_hot_embeddings=use_one_hot_embeddings)
#从bert模型读取序列输出
final_hidden = model.get_sequence_output()
final_hidden_shape = modeling.get_shape_list(final_hidden, expected_rank=3)
batch_size = final_hidden_shape[0] #批次数量大小
seq_length = final_hidden_shape[1] #序列长度
hidden_size = final_hidden_shape[2] #隐藏单元大小
#初始化权重矩阵
output_weights = tf.get_variable(
"cls/squad/output_weights", [2, hidden_size],
initializer=tf.truncated_normal_initializer(stddev=0.02))
#初始化偏置项
output_bias = tf.get_variable("cls/squad/output_bias", [2],
initializer=tf.zeros_initializer())
final_hidden_matrix = tf.reshape(final_hidden,
[batch_size * seq_length, hidden_size])
#回归训练，求解y=ax+b ，这种模型结构，a=output_weights，b=output_bias，y=logits
logits = tf.matmul(final_hidden_matrix, output_weights,
transpose_b=True) #矩阵相乘
logits = tf.nn.bias_add(logits, output_bias) #加上偏置项
logits = tf.reshape(
logits,
[batch_size, seq_length, 2]) #把输出转换成[每批数量，序列长度，2] 这种格式[2维数组数,行数，列数]
logits = tf.transpose(logits, [2, 0, 1]) #对上一步的结果转置，[2,0,1]代表[列数，2维数组数，行数]
unstacked_logits = tf.unstack(logits, axis=0) #对矩阵在行上拆分
(start_logits, end_logits) = (unstacked_logits[0], unstacked_logits[1])
return (start_logits, end_logits) #返回起始位置，结束位置
def create_model(bert_config, is_training, input_ids, input_mask, segment_ids,
use_one_hot_embeddings):
"""Creates a classification model."""
model = modeling.BertModel(
config=bert_config,
is_training=is_training,
input_ids=input_ids,
input_mask=input_mask,
token_type_ids=segment_ids,
use_one_hot_embeddings=use_one_hot_embeddings)
#self.sequence_output = self.all_encoder_layers[-1]
#取最后一层（batch_size，seq_length，hidden_size）
final_hidden = model.get_sequence_output()
final_hidden_shape = modeling.get_shape_list(final_hidden, expected_rank=3)
batch_size = final_hidden_shape[0]
seq_length = final_hidden_shape[1]
hidden_size = final_hidden_shape[2]
output_weights = tf.get_variable(
"cls/squad/output_weights", [2, hidden_size],
initializer=tf.truncated_normal_initializer(stddev=0.02))
output_bias = tf.get_variable(
"cls/squad/output_bias", [2], initializer=tf.zeros_initializer())
final_hidden_matrix = tf.reshape(final_hidden,
[batch_size * seq_length, hidden_size])
logits = tf.matmul(final_hidden_matrix, output_weights, transpose_b=True)
logits = tf.nn.bias_add(logits, output_bias)
logits = tf.reshape(logits, [batch_size, seq_length, 2])
logits = tf.transpose(logits, [2, 0, 1])
unstacked_logits = tf.unstack(logits, axis=0)
(start_logits, end_logits) = (unstacked_logits[0], unstacked_logits[1])
###如果这里只想做一个填空形式，例如只判断某个词是否为开始位置
###可以加个全连接层[hidden_size,1],相乘后得到[batch_size, seq_length, 1],sequeeze去掉维度为1的那个维度；
###最终得到[batch_size, seq_length]就可以用多分类判断seq_length的每个位置是否为答案的位置
return (start_logits, end_logits)
def create_model(bert_config, is_training, input_ids, input_mask, segment_ids,
labels, num_labels, use_one_hot_embeddings):
"""Creates a seq labelling model."""
model = modeling.BertModel(
config=bert_config,
is_training=is_training,
input_ids=input_ids,
input_mask=input_mask,
token_type_ids=segment_ids,
use_one_hot_embeddings=use_one_hot_embeddings)
with tf.variable_scope("finetune/seq"):
#get sequnece output
final_hidden = model.get_sequence_output()
final_hidden_shape = modeling.get_shape_list(final_hidden, expected_rank=3)
batch_size = final_hidden_shape[0]
seq_length = final_hidden_shape[1]
hidden_size = final_hidden_shape[2]
final_hidden = tf.reshape(final_hidden, [batch_size*seq_length, hidden_size])
output_weights = tf.get_variable(
"output_weights", [num_labels, hidden_size],
initializer=tf.truncated_normal_initializer(stddev=0.02))
output_bias = tf.get_variable(
"output_bias", [num_labels], initializer=tf.zeros_initializer())
logits = tf.matmul(final_hidden, output_weights, transpose_b=True)
logits = tf.nn.bias_add(logits, output_bias)
logits_out = tf.reshape(logits, [batch_size, seq_length, num_labels])
probabilities = tf.nn.softmax(logits_out)
log_probs = tf.nn.log_softmax(logits)
labels = tf.reshape(labels, [-1])
label_weights = tf.cast(tf.reshape(input_mask, [-1]), dtype=tf.float32)
one_hot_labels = tf.one_hot(
labels, depth=num_labels, dtype=tf.float32)
per_example_loss = -tf.reduce_sum(log_probs * one_hot_labels, axis=[-1]) * label_weights
numerator = tf.reduce_sum(per_example_loss)
denominator = tf.reduce_sum(label_weights) + 1e-5
loss = numerator/denominator
per_example_loss = tf.reshape(per_example_loss, [batch_size, seq_length])
return (loss, per_example_loss, logits_out, probabilities)
def create_model(bert_config, is_training, input_ids, input_mask, segment_ids,
use_one_hot_embeddings):
"""Creates a classification model."""
model = modeling.BertModel(
config=bert_config,#たぶんこの設定にしたがってbertを呼び出すということ
is_training=is_training,
input_ids=input_ids,
input_mask=input_mask,
token_type_ids=segment_ids,
use_one_hot_embeddings=use_one_hot_embeddings)
final_hidden = model.get_sequence_output()#Bertの最終層
final_hidden_shape = modeling.get_shape_list(final_hidden, expected_rank=3)
batch_size = final_hidden_shape[0]
seq_length = final_hidden_shape[1]
hidden_size = final_hidden_shape[2]
output_weights = tf.get_variable(
#/はスコープの区切りを表す。だからcls/squad/output_weightsはclsスコープのsquadスコープのoutput_weightsという変数を表す
#変数がないときは定義し、ある時はそれを呼び出す
"cls/squad/output_weights", [2, hidden_size],
initializer=tf.truncated_normal_initializer(stddev=0.02))
output_bias = tf.get_variable(
#/はスコープの区切りを表す。だからcls/squad/output_weightsはclsスコープのsquadスコープのoutput_weightsという変数を表す
#変数がないときは定義し、ある時はそれを呼び出す
"cls/squad/output_bias", [2], initializer=tf.zeros_initializer())
final_hidden_matrix = tf.reshape(final_hidden,
[batch_size * seq_length, hidden_size])
logits = tf.matmul(final_hidden_matrix, output_weights, transpose_b=True)
logits = tf.nn.bias_add(logits, output_bias)
logits = tf.reshape(logits, [batch_size, seq_length, 2])
logits = tf.transpose(logits, [2, 0, 1])
unstacked_logits = tf.unstack(logits, axis=0)
(start_logits, end_logits) = (unstacked_logits[0], unstacked_logits[1])
return (start_logits, end_logits)
def get_new_input_ids(input_ids,
masked_lm_probs,masked_lm_positions,masked_lm_weights):
#Reshaping the masked_lm_probs
sequence_shape = modeling.get_shape_list(masked_lm_positions, expected_rank=2)
#for getting the shape info
batch_size = sequence_shape[0]
seq_length = sequence_shape[1]
# Gathering the first masks only. shape = (batch_size,vocab_size)
mask_lab_pred = tf.gather(masked_lm_probs,
tf.range(0,batch_size*seq_length,delta=seq_length),axis=0)
#Getting the indexes for the predictions. shape = (batch_size,1)
#mask_lab_pred = tf.random.multinomial(mask_lab_pred,1,output_dtype=tf.int32)
mask_lab_pred = tf.reshape(tf.cast(tf.argmax(mask_lab_pred,
axis=-1),dtype=tf.int32),(batch_size,1))
# Gathering positions of the first mask. shape=(batch_size,)
mask_positions = tf.gather(masked_lm_positions,
tf.constant(0,dtype=tf.int32),axis=-1)
#Assigning the first masks in input_ids
#--First converting mask_positions to one hot, shape=(batch_size,seq_length)
mask_positions = tf.one_hot(mask_positions,depth=seq_length,axis=1,dtype=tf.int32)
#--Removing those positions there is no mask left
mask_positions = tf.cast(tf.reduce_max(masked_lm_positions,axis=-1,keepdims=True)>0
,dtype=tf.int32)*mask_positions
#--Next multiplying mask_positions_one_hot to mask_lab_pred and forming new_ids
input_ids = mask_positions*mask_lab_pred+(1-mask_positions)*input_ids
#Setting the first mask lm_weights to be zero
masked_lm_weights = tf.slice(masked_lm_weights,(0,1),(-1,-1))
masked_lm_weights = tf.concat([masked_lm_weights,tf.zeros(shape=(batch_size,1),
dtype=tf.float32)],axis=-1)
#Setting the masked_lm_positions first masks to be zero
masked_lm_positions = tf.slice(masked_lm_positions,(0,1),(-1,-1))
masked_lm_positions = tf.concat([masked_lm_positions,tf.zeros(shape=(batch_size,1),
dtype=tf.int32)],axis=-1)
return input_ids,masked_lm_positions,masked_lm_weights
def create_model(bert_config, is_training, input_ids, input_mask, segment_ids,
labels, num_labels, use_one_hot_embeddings):
"""Creates a classification model."""
model = modeling.BertModel(config=bert_config,
is_training=is_training,
input_ids=input_ids,
input_mask=input_mask,
token_type_ids=segment_ids,
use_one_hot_embeddings=use_one_hot_embeddings)
# In the demo, we are doing a simple classification task on the entire
# segment.
#
# If you want to use the token-level output, use model.get_sequence_output()
# instead.
final_hidden = model.get_sequence_output()
final_hidden_shape = modeling.get_shape_list(final_hidden, expected_rank=3)
batch_size = final_hidden_shape[0]
seq_length = final_hidden_shape[1]
hidden_size = final_hidden_shape[2]
output_weights = tf.get_variable(
"output_weights", [num_labels, hidden_size],
initializer=tf.truncated_normal_initializer(stddev=0.02))
output_bias = tf.get_variable("output_bias", [num_labels],
initializer=tf.zeros_initializer())
with tf.variable_scope("loss"):
final_hidden_matrix = tf.reshape(
final_hidden, [batch_size * seq_length, hidden_size])
logits = tf.matmul(final_hidden_matrix,
output_weights,
transpose_b=True)
logits = tf.nn.bias_add(logits, output_bias)
logits = tf.reshape(logits, [batch_size, seq_length, num_labels])
#logits = tf.transpose(logits, [2,0,1])
one_hot_labels = tf.one_hot(labels, depth=num_labels, dtype=tf.float32)
log_probs = tf.nn.log_softmax(logits, axis=-1)
per_example_loss = -tf.reduce_sum(one_hot_labels * log_probs, axis=-1)
loss = tf.reduce_mean(per_example_loss)
return (loss, per_example_loss, logits)
def gather_indexes(sequence_tensor, positions):
"""Gathers the vectors at the specific positions over a minibatch."""
sequence_shape = modeling.get_shape_list(sequence_tensor,
expected_rank=3) #list
batch_size = sequence_shape[
0] #'tensorflow.python.framework.ops.Tensor', Tensor("strided_slice_1:0", shape=(), dtype=int32)
seq_length = sequence_shape[1] #256
width = sequence_shape[2] #64
flat_offsets = tf.reshape(
tf.range(0, batch_size, dtype=tf.int32) * seq_length, [-1, 1]) #(?, 1)
flat_positions = tf.reshape(positions + flat_offsets, [-1]) #(?,)
flat_sequence_tensor = tf.reshape(sequence_tensor,
[batch_size * seq_length, width])
output_tensor = tf.gather(
flat_sequence_tensor, flat_positions
) #Gather slices from params axis axis according to indices.
##!for seq in tf.range(0, batch_size, dtype=tf.int32):
## output_tensor = tf.gather(sequence_tensor[seq], positions)
return output_tensor
def gather_indexes(
sequence_tensor, positions
): ###sequence_tensor 是transorformer的输出[batch_size,length,hidden_size]
###position是每个case中被mask的位置index，[batch,max_mask_length]，默认max_mask_length长度为20 每个实例：[7,10,15,20,24,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
"""Gathers the vectors at the specific positions over a minibatch."""
sequence_shape = modeling.get_shape_list(sequence_tensor, expected_rank=3)
batch_size = sequence_shape[0]
seq_length = sequence_shape[1]
width = sequence_shape[2]
flat_offsets = tf.reshape(
tf.range(0, batch_size, dtype=tf.int32) * seq_length,
[-1, 1]) ###[batch_size,1]
flat_positions = tf.reshape(positions + flat_offsets,
[-1]) #获得所有mask的position，并flat成一维
flat_sequence_tensor = tf.reshape(sequence_tensor,
[batch_size * seq_length, width])
output_tensor = tf.gather(flat_sequence_tensor,
flat_positions) ###获取被mask位置对应transformer的输出
return output_tensor
def get_matrix_mask_indices(matrix, num_rows=None):
if num_rows is None:
num_rows = modeling.get_shape_list(matrix)[0]
indices = tf.where(matrix)
num_indices = tf.shape(indices)[0]
elem_per_row = tf.bincount(tf.cast(indices[:, 0], tf.int32),
minlength=num_rows)
max_elem_per_row = tf.reduce_max(elem_per_row)
row_start = tf.concat([[0], tf.cumsum(elem_per_row[:-1])], axis=0)
r = tf.range(max_elem_per_row)
idx = tf.expand_dims(row_start, 1) + r
idx = tf.minimum(idx, num_indices - 1)
result = tf.gather(indices[:, 1], idx)
# replace invalid elements with -1
result = tf.where(
tf.expand_dims(elem_per_row, 1) > r, result, -tf.ones_like(result))
max_index_per_row = tf.reduce_max(result, axis=1, keepdims=True)
max_index_per_row = tf.tile(max_index_per_row, [1, max_elem_per_row])
result = tf.where(result >= 0, result, max_index_per_row)
return result
def gather_indexes(sequence_tensor, positions):
"""Gathers the vectors at the specific positions over a minibatch."""
#这个函数主要是从指定位置的position中（即被mask的位置），从输入batch_size *
#seqlength * embedding_size的tensor中，找出来被mask的字符向量。将设batch_size=8,
#,sequength_length的长度为128，在这个128的字符中，有20个位置被mask，则本函数就是要找出
#8 * 20 = 160个mask调的字符，每个字符用768的向量表示，则最终输出的结果为160*768
import ipdb
ipdb.set_trace()
sequence_shape = modeling.get_shape_list(sequence_tensor, expected_rank=3)
batch_size = sequence_shape[0]
seq_length = sequence_shape[1]
width = sequence_shape[2]
#由于后文要将sequence_tensor进行flat操作，即batch_size*sequence_length操作，所以
#这里要对position进行一个位置的offset（偏置），依次向后偏置sequence_length操作。
flat_offsets = tf.reshape(
tf.range(0, batch_size, dtype=tf.int32) * seq_length, [-1, 1])
flat_positions = tf.reshape(positions + flat_offsets, [-1])
flat_sequence_tensor = tf.reshape(sequence_tensor,
[batch_size * seq_length, width])
output_tensor = tf.gather(flat_sequence_tensor, flat_positions)
return output_tensor
def knn(labels, embeddings, k, embed_normed=True):
# make sure embedding should be l2-normalized
if not embed_normed:
embeddings = tf.nn.l2_normalize(embeddings, axis=1)
embed_shape = modeling.get_shape_list(embeddings)
batch_size = embed_shape[0]
sim_mat = tf.matmul(embeddings, embeddings, transpose_b=True)
sim_mat = sim_mat - tf.eye(batch_size) * 2.0
_, top_k_idx = tf.nn.top_k(sim_mat, k)
top_k_labels = tf.squeeze(tf.gather(labels, top_k_idx))
def knn_vote(v):
nearest_k_y, idx, votes = tf.unique_with_counts(v)
majority_idx = tf.argmax(votes)
predict_res = tf.gather(nearest_k_y, majority_idx)
return predict_res
majority = tf.map_fn(knn_vote, top_k_labels)
return majority
def replace_elements_by_indices(old, new, indices):
old_shape = modeling.get_shape_list(old)
print(old_shape)
batch_size = old_shape[0]
seq_length = old_shape[1]
flat_offsets = tf.reshape(
tf.range(0, batch_size, dtype=tf.int32) * seq_length, [-1, 1])
print(flat_offsets)
flat_positions = tf.reshape(indices + flat_offsets, [-1])
print(flat_positions)
zeros = tf.zeros(tf.shape(input=flat_positions)[0], dtype=tf.int32)
print(zeros)
flat_old = tf.reshape(old, [-1])
print(flat_old)
masked_lm_mask = tf.compat.v1.sparse_to_dense(flat_positions,
tf.shape(input=flat_old),
zeros,
default_value=1,
validate_indices=True,
name="masked_lm_mask")
print(masked_lm_mask)
flat_old_temp = tf.multiply(flat_old, masked_lm_mask)
print(flat_old_temp)
new_temp = tf.compat.v1.sparse_to_dense(flat_positions,
tf.shape(input=flat_old),
new,
default_value=0,
validate_indices=True,
name=None)
print(new_temp)
updated_old = tf.reshape(flat_old_temp + new_temp, old_shape)
print(updated_old)
return updated_old
def create_model(bert_config, is_training, input_ids, input_len, input_mask, segment_ids,
labels, num_labels, use_one_hot_embeddings):
"""Creates a sequence model."""
model = modeling.BertModel(
config=bert_config,
is_training=is_training,
input_ids=input_ids,
input_mask=input_mask,
token_type_ids=segment_ids,
use_one_hot_embeddings=use_one_hot_embeddings)
final_hidden = model.get_sequence_output()
final_hidden_shape = modeling.get_shape_list(final_hidden, expected_rank=3)
batch_size = final_hidden_shape[0]
seq_length = final_hidden_shape[1]
hidden_size = final_hidden_shape[2]
output_size = num_labels
with tf.variable_scope("bert_finetuning"):
output_weights = tf.get_variable(
"token_output_weights", [output_size, hidden_size],
initializer=tf.truncated_normal_initializer(stddev=0.02))
output_bias = tf.get_variable(
"token_output_bias", [output_size], initializer=tf.zeros_initializer())
final_hidden_matrix = tf.reshape(final_hidden,
[batch_size * seq_length, hidden_size])
logits = tf.matmul(final_hidden_matrix, output_weights, transpose_b=True)
logits = tf.nn.bias_add(logits, output_bias)
logits = tf.reshape(logits, [batch_size, seq_length, output_size])
one_hot_labels = tf.one_hot(labels, depth=num_labels, axis=-1, dtype=tf.float32)
entropy_loss = tf.nn.softmax_cross_entropy_with_logits(labels=one_hot_labels, logits=logits, dim=-1,
name="loss")
per_example_loss = tf.reduce_sum(tf.slice(entropy_loss,begin=[0,1],size=[-1,input_len[0]]), axis=-1)
loss = tf.reduce_mean(per_example_loss)
probs = tf.nn.softmax(logits, axis=-1)
return (loss, per_example_loss, probs, logits)
def create_model(bert_config, is_training, input_ids, input_mask, segment_ids,
use_one_hot_embeddings):
"""Creates a classification model."""
model = modeling.BertModel(
config=bert_config,
is_training=is_training,
input_ids=input_ids,
input_mask=input_mask,
token_type_ids=segment_ids,
use_one_hot_embeddings=use_one_hot_embeddings)
final_hidden = model.get_sequence_output()
final_hidden_shape = modeling.get_shape_list(final_hidden, expected_rank=3)
batch_size = final_hidden_shape[0]
seq_length = final_hidden_shape[1]
hidden_size = final_hidden_shape[2]
output_weights = tf.get_variable(
"cls/squad/output_weights", [2, hidden_size],
initializer=tf.truncated_normal_initializer(stddev=0.02))
output_bias = tf.get_variable(
"cls/squad/output_bias", [2], initializer=tf.zeros_initializer())
final_hidden_matrix = tf.reshape(final_hidden,
[batch_size * seq_length, hidden_size])
logits = tf.matmul(final_hidden_matrix, output_weights, transpose_b=True)
logits = tf.nn.bias_add(logits, output_bias)
logits = tf.reshape(logits, [batch_size, seq_length, 2])
logits = tf.transpose(logits, [2, 0, 1])
unstacked_logits = tf.unstack(logits, axis=0)
(start_logits, end_logits) = (unstacked_logits[0], unstacked_logits[1])
return (start_logits, end_logits)
def get_shuffle_loss(model_config, seq_output, label_ids, label_weights):
sequence_shape = modeling.get_shape_list(seq_output, expected_rank=[3])
seq_length = sequence_shape[1]
width = sequence_shape[2]
seq_output = tf.reshape(seq_output, [-1, width])
with tf.variable_scope("cls/shuffle"):
with tf.variable_scope("transform"):
seq_output = tf.layers.dense(
seq_output,
units=seq_length,
activation=modeling.get_activation(model_config.hidden_act),
kernel_initializer=modeling.create_initializer(
model_config.initializer_range))
seq_output = modeling.layer_norm(seq_output)
output_bias = tf.get_variable("output_bias",
shape=[seq_length],
initializer=tf.zeros_initializer())
logits = tf.nn.bias_add(seq_output, output_bias)
log_probs = tf.nn.log_softmax(logits, axis=-1)
label_ids = tf.reshape(label_ids, [-1])
label_weights = tf.reshape(tf.cast(label_weights, tf.float32), [-1])
one_hot_labels = tf.one_hot(label_ids,
depth=seq_length,
dtype=tf.float32)
per_example_loss = -tf.reduce_sum(log_probs * one_hot_labels,
axis=[-1])
numerator = tf.reduce_sum(label_weights * per_example_loss)
denominator = tf.reduce_sum(label_weights) + 1e-5
loss = numerator / denominator
return loss, per_example_loss, log_probs
def create_model(self, bert_config, is_training, input_ids, input_mask,
segment_ids, use_one_hot_embeddings):
"""Creates a classification model."""
model = modeling.BertModel(
config=bert_config,
is_training=is_training,
input_ids=input_ids,
input_mask=input_mask,
token_type_ids=segment_ids,
use_one_hot_embeddings=use_one_hot_embeddings,
compute_type=tf.float32)
final_hidden = model.get_sequence_output()
final_hidden_shape = modeling.get_shape_list(final_hidden,
expected_rank=3)
batch_size = final_hidden_shape[0]
seq_length = final_hidden_shape[1]
hidden_size = final_hidden_shape[2]
output_weights = tf.get_variable(
"cls/squad/output_weights", [2, hidden_size],
initializer=tf.truncated_normal_initializer(stddev=0.02))
output_bias = tf.get_variable("cls/squad/output_bias", [2],
initializer=tf.zeros_initializer())
final_hidden_matrix = tf.reshape(
final_hidden, [batch_size * seq_length, hidden_size])
logits = tf.matmul(final_hidden_matrix,
output_weights,
transpose_b=True)
logits = tf.nn.bias_add(logits, output_bias)
logits = tf.reshape(logits, [batch_size, seq_length, 2])
return logits
def get_masked_lm_output(self, bert_config, input_tensor, output_weights,
positions, label_ids, label_weights, trainable):
"""Get loss and log probs for the masked LM."""
input_tensor = self.gather_indexes(input_tensor, positions)
if self.is_negsample:
logits_2D = input_tensor
label_flat = tf.reshape(
label_ids, [-1, 1]) # 1 is the number of positive example
num_sampled = int(
0.2 * self.model_para['item_size']) # sample 20% as negatives
loss = tf.nn.sampled_softmax_loss(self.softmax_w, self.softmax_b,
label_flat, logits_2D,
num_sampled,
self.model_para['item_size'])
else:
sequence_shape = modeling.get_shape_list(positions)
batch_size = sequence_shape[0]
seq_length = sequence_shape[1]
residual_channels = input_tensor.get_shape().as_list()[-1]
input_tensor = tf.reshape(input_tensor,
[-1, seq_length, residual_channels])
logits = ops.conv1d(tf.nn.relu(input_tensor),
self.model_para['item_size'],
name='logits')
logits_2D = tf.reshape(logits, [-1, self.model_para['item_size']])
label_flat = tf.reshape(label_ids, [-1])
loss = tf.nn.sparse_softmax_cross_entropy_with_logits(
labels=label_flat, logits=logits_2D)
loss = tf.reduce_mean(loss)
#not sure the impact, 0.001 is empirical value
# regularization = 0.001 * tf.reduce_mean([tf.nn.l2_loss(v) for v in tf.trainable_variables()])
# loss=loss+regularization
return loss
def model_fn(features, labels, mode, params): # pylint: disable=unused-argument
"""The `model_fn` for TPUEstimator."""
tf.logging.info("*** Features ***")
for name in sorted(features.keys()):
tf.logging.info(" name = %s, shape = %s" % (name, features[name].shape))
unique_ids = features["unique_ids"]
input_ids = features["input_ids"]
input_mask = features["input_mask"]
segment_ids = features["segment_ids"]
is_training = (mode == tf.estimator.ModeKeys.TRAIN)
(start_logits, end_logits) = create_model(
bert_config=bert_config,
is_training=is_training,
input_ids=input_ids,
input_mask=input_mask,
segment_ids=segment_ids,
use_one_hot_embeddings=use_one_hot_embeddings)
tvars = tf.trainable_variables()
initialized_variable_names = {}
scaffold_fn = None
if init_checkpoint:
(assignment_map, initialized_variable_names
) = modeling.get_assignment_map_from_checkpoint(tvars, init_checkpoint)
if use_tpu:
def tpu_scaffold():
tf.train.init_from_checkpoint(init_checkpoint, assignment_map)
return tf.train.Scaffold()
scaffold_fn = tpu_scaffold
else:
tf.train.init_from_checkpoint(init_checkpoint, assignment_map)
tf.logging.info("**** Trainable Variables ****")
for var in tvars:
init_string = ""
if var.name in initialized_variable_names:
init_string = ", *INIT_FROM_CKPT*"
tf.logging.info(" name = %s, shape = %s%s", var.name, var.shape,
init_string)
output_spec = None
if mode == tf.estimator.ModeKeys.TRAIN:
seq_length = modeling.get_shape_list(input_ids)[1]
def compute_loss(logits, positions): #交叉熵损失函数
#把位置数转换成独热标签
one_hot_positions = tf.one_hot(
positions, depth=seq_length, dtype=tf.float32)
#对softmax结果取ln对数，计算熵的参数，logits部分的信息量
log_probs = tf.nn.log_softmax(logits, axis=-1)
#先累加（独热标签乘以每个位的概率），然后取均值
loss = -tf.reduce_mean(
tf.reduce_sum(one_hot_positions * log_probs, axis=-1)) #熵的计算公式正确值与（预测值概率对数）相乘的累加
return loss
start_positions = features["start_positions"]
end_positions = features["end_positions"]
#计算答案开始位置的损失
start_loss = compute_loss(start_logits, start_positions)
#计算答案结束位置的损失
end_loss = compute_loss(end_logits, end_positions)
#总的损失
total_loss = (start_loss + end_loss) / 2.0
#调用模型优化函数
train_op = optimization.create_optimizer(
total_loss, learning_rate, num_train_steps, num_warmup_steps, use_tpu)
output_spec = tf.contrib.tpu.TPUEstimatorSpec(
mode=mode,
loss=total_loss,
train_op=train_op,
scaffold_fn=scaffold_fn)
elif mode == tf.estimator.ModeKeys.PREDICT:
predictions = {
"unique_ids": unique_ids,
"start_logits": start_logits,
"end_logits": end_logits,
}
output_spec = tf.contrib.tpu.TPUEstimatorSpec(
mode=mode, predictions=predictions, scaffold_fn=scaffold_fn)
else:
raise ValueError(
"Only TRAIN and PREDICT modes are supported: %s" % (mode))
return output_spec
def model_fn(features, labels, mode, params): # pylint: disable=unused-argument
"""The `model_fn` for TPUEstimator."""
tf.logging.info("*** Features ***")
for name in sorted(features.keys()):
tf.logging.info(" name = %s, shape = %s" %
(name, features[name].shape))
unique_ids = features["unique_ids"]
input_ids = features["input_ids"]
input_mask = features["input_mask"]
segment_ids = features["segment_ids"]
is_training = (mode == tfes.estimator.ModeKeys.TRAIN)
(start_logits, end_logits) = create_model(
bert_config=bert_config,
is_training=is_training,
input_ids=input_ids,
input_mask=input_mask,
segment_ids=segment_ids,
use_one_hot_embeddings=use_one_hot_embeddings)
tvars = tf.trainable_variables()
initialized_variable_names = {}
scaffold_fn = None
if init_checkpoint:
(assignment_map, initialized_variable_names
) = modeling.get_assignment_map_from_checkpoint(
tvars, init_checkpoint)
if use_tpu:
def tpu_scaffold():
tf.train.init_from_checkpoint(init_checkpoint,
assignment_map)
return tf.train.Scaffold()
scaffold_fn = tpu_scaffold
else:
tf.train.init_from_checkpoint(init_checkpoint, assignment_map)
# tf.logging.info("**** Trainable Variables ****")
# for var in tvars:
# init_string = ""
# if var.name in initialized_variable_names:
# init_string = ", *INIT_FROM_CKPT*"
# tf.logging.info(" name = %s, shape = %s%s", var.name, var.shape,
# init_string)
output_spec = None
if mode == tfes.estimator.ModeKeys.TRAIN:
seq_length = modeling.get_shape_list(input_ids)[1]
def compute_loss(logits, positions):
one_hot_positions = tf.one_hot(positions,
depth=seq_length,
dtype=tf.float32)
log_probs = tf.nn.log_softmax(logits, axis=-1)
loss = -tf.reduce_mean(
tf.reduce_sum(one_hot_positions * log_probs, axis=-1))
return loss
start_positions = features["start_positions"]
end_positions = features["end_positions"]
start_loss = compute_loss(start_logits, start_positions)
end_loss = compute_loss(end_logits, end_positions)
total_loss = (start_loss + end_loss) / 2.0
train_op = optimization.create_optimizer(total_loss, learning_rate,
num_train_steps,
num_warmup_steps, use_tpu)
output_spec = tpu.TPUEstimatorSpec(mode=mode,
loss=total_loss,
train_op=train_op,
scaffold_fn=scaffold_fn)
elif mode == tfes.estimator.ModeKeys.PREDICT:
# outer = tf.matmul(tf.expand_dims(tf.nn.softmax(start_logits), axis=2),
# tf.expand_dims(tf.nn.softmax(end_logits), axis=1))
# outer = tf.matrix_band_part(outer, -1, 15) # 取上3角15条对角线，表示答案最大长度只能取到15+1个单词
# yp1 = tf.argmax(tf.reduce_max(outer, axis=2), axis=1) # 寻找最大值在L1轴的索引
# yp2 = tf.argmax(tf.reduce_max(outer, axis=1), axis=1)
predictions = {
"unique_ids": unique_ids,
"start_logits": start_logits,
"end_logits": end_logits,
# "yp1": yp1,
# "yp2": yp2,
}
output_spec = tpu.TPUEstimatorSpec(mode=mode,
predictions=predictions,
scaffold_fn=scaffold_fn)
else:
raise ValueError("Only TRAIN and PREDICT modes are supported: %s" %
(mode))
return output_spec
def model_fn(features, labels, mode, params): # pylint: disable=unused-argument
"""The `model_fn` for Estimator."""
if FLAGS.verbose_logging:
tf.compat.v1.logging.info("*** Features ***")
for name in sorted(features.keys()):
tf.compat.v1.logging.info(" name = %s, shape = %s" % (name, features[name].shape))
unique_ids = features["unique_ids"]
input_ids = features["input_ids"]
input_mask = features["input_mask"]
segment_ids = features["segment_ids"]
is_training = (mode == tf.estimator.ModeKeys.TRAIN)
if not is_training and FLAGS.use_trt:
trt_graph = get_frozen_tftrt_model(bert_config, input_ids.shape, use_one_hot_embeddings, init_checkpoint)
(start_logits, end_logits) = tf.import_graph_def(trt_graph,
input_map={'input_ids':input_ids, 'input_mask':input_mask, 'segment_ids':segment_ids},
return_elements=['unstack:0', 'unstack:1'],
name='')
predictions = {
"unique_ids": unique_ids,
"start_logits": start_logits,
"end_logits": end_logits,
}
output_spec = tf.estimator.EstimatorSpec(
mode=mode, predictions=predictions)
return output_spec
(start_logits, end_logits) = create_model(
bert_config=bert_config,
is_training=is_training,
input_ids=input_ids,
input_mask=input_mask,
segment_ids=segment_ids,
use_one_hot_embeddings=use_one_hot_embeddings)
tvars = tf.trainable_variables()
initialized_variable_names = {}
if init_checkpoint and (hvd is None or hvd.rank() == 0):
(assignment_map, initialized_variable_names
) = modeling.get_assignment_map_from_checkpoint(tvars, init_checkpoint)
tf.train.init_from_checkpoint(init_checkpoint, assignment_map)
if FLAGS.verbose_logging:
tf.compat.v1.logging.info("**** Trainable Variables ****")
for var in tvars:
init_string = ""
if var.name in initialized_variable_names:
init_string = ", *INIT_FROM_CKPT*"
tf.compat.v1.logging.info(" %d name = %s, shape = %s%s", 0 if hvd is None else hvd.rank(), var.name, var.shape,
init_string)
output_spec = None
if mode == tf.estimator.ModeKeys.TRAIN:
seq_length = modeling.get_shape_list(input_ids)[1]
def compute_loss(logits, positions):
one_hot_positions = tf.one_hot(
positions, depth=seq_length, dtype=tf.float32)
log_probs = tf.nn.log_softmax(logits, axis=-1)
loss = -tf.reduce_mean(
tf.reduce_sum(one_hot_positions * log_probs, axis=-1))
return loss
start_positions = features["start_positions"]
end_positions = features["end_positions"]
start_loss = compute_loss(start_logits, start_positions)
end_loss = compute_loss(end_logits, end_positions)
total_loss = (start_loss + end_loss) / 2.0
train_op = optimization.create_optimizer(
total_loss, learning_rate, num_train_steps, num_warmup_steps, hvd, False, amp, FLAGS.num_accumulation_steps)
output_spec = tf.estimator.EstimatorSpec(
mode=mode,
loss=total_loss,
train_op=train_op)
elif mode == tf.estimator.ModeKeys.PREDICT:
dummy_op = tf.no_op()
# Need to call mixed precision graph rewrite if fp16 to enable graph rewrite
if amp:
loss_scaler = tf.train.experimental.FixedLossScale(1)
dummy_op = tf.train.experimental.enable_mixed_precision_graph_rewrite(
optimization.LAMBOptimizer(learning_rate=0.0), loss_scaler)
predictions = {
"unique_ids": unique_ids,
"start_logits": start_logits,
"end_logits": end_logits,
}
output_spec = tf.estimator.EstimatorSpec(
mode=mode, predictions=predictions)
else:
raise ValueError(
"Only TRAIN and PREDICT modes are supported: %s" % (mode))
return output_spec
def create_model(bert_config, is_training, slot_list, features,
num_class_labels, use_one_hot_embeddings):
"""Creates a classification model."""
input_ids = features["input_ids"]
input_mask = features["input_mask"]
segment_ids = features["segment_ids"]
model = modeling.BertModel(config=bert_config,
is_training=is_training,
input_ids=input_ids,
input_mask=input_mask,
token_type_ids=segment_ids,
use_one_hot_embeddings=use_one_hot_embeddings)
# In the demo, we are doing a simple classification task on the entire
# segment.
#
# If you want to use the token-level output, use model.get_sequence_output()
# instead.
class_output_layer = model.get_pooled_output()
token_output_layer = model.get_sequence_output()
token_output_shape = modeling.get_shape_list(token_output_layer,
expected_rank=3)
batch_size = token_output_shape[0]
seq_length = token_output_shape[1]
hidden_size = token_output_shape[2]
# Define prediction variables
class_proj_layer_dim = [hidden_size]
for idx in range(FLAGS.num_class_hidden_layer):
class_proj_layer_dim.append(64)
class_proj_layer_dim.append(num_class_labels)
token_proj_layer_dim = [hidden_size]
for idx in range(FLAGS.num_token_hidden_layer):
token_proj_layer_dim.append(64)
token_proj_layer_dim.append(2)
if is_training:
# I.e., 0.1 dropout
class_output_layer = tf.nn.dropout(class_output_layer,
keep_prob=(1 - FLAGS.dropout_rate))
token_output_layer = tf.nn.dropout(token_output_layer,
keep_prob=(1 - FLAGS.dropout_rate))
total_loss = 0
per_slot_per_example_loss = {}
per_slot_class_logits = {}
per_slot_start_logits = {}
per_slot_end_logits = {}
for slot in slot_list:
start_pos = features["start_pos_%s" % slot]
end_pos = features["end_pos_%s" % slot]
class_label_id = features["class_label_id_%s" % slot]
slot_scope_name = "slot_%s" % slot
if slot == 'price range':
slot_scope_name = "slot_price"
with tf.variable_scope(slot_scope_name):
class_list_output_weights = []
class_list_output_bias = []
for l_idx in range(len(class_proj_layer_dim) - 1):
dim_in = class_proj_layer_dim[l_idx]
dim_out = class_proj_layer_dim[l_idx + 1]
class_list_output_weights.append(
tf.get_variable(
"class/output_weights_%d" % l_idx, [dim_in, dim_out],
initializer=tf.truncated_normal_initializer(
stddev=0.02)))
class_list_output_bias.append(
tf.get_variable("class/output_bias_%d" % l_idx, [dim_out],
initializer=tf.zeros_initializer()))
token_list_output_weights = []
token_list_output_bias = []
for l_idx in range(len(token_proj_layer_dim) - 1):
dim_in = token_proj_layer_dim[l_idx]
dim_out = token_proj_layer_dim[l_idx + 1]
token_list_output_weights.append(
tf.get_variable(
"token/output_weights_%d" % l_idx, [dim_in, dim_out],
initializer=tf.truncated_normal_initializer(
stddev=0.02)))
token_list_output_bias.append(
tf.get_variable("token/output_bias_%d" % l_idx, [dim_out],
initializer=tf.zeros_initializer()))
with tf.variable_scope("loss"):
class_logits = util.fully_connect_layers(
class_output_layer, class_list_output_weights,
class_list_output_bias)
one_hot_class_labels = tf.one_hot(class_label_id,
depth=num_class_labels,
dtype=tf.float32)
class_loss = tf.losses.softmax_cross_entropy(
one_hot_class_labels,
class_logits,
reduction=tf.losses.Reduction.NONE)
token_is_pointable = tf.cast(tf.equal(class_label_id, 2),
dtype=tf.float32)
token_output_layer = tf.reshape(
token_output_layer, [batch_size * seq_length, hidden_size])
token_logits = util.fully_connect_layers(
token_output_layer, token_list_output_weights,
token_list_output_bias)
token_logits = tf.reshape(token_logits,
[batch_size, seq_length, 2])
token_logits = tf.transpose(token_logits, [2, 0, 1])
unstacked_token_logits = tf.unstack(token_logits, axis=0)
(start_logits, end_logits) = (unstacked_token_logits[0],
unstacked_token_logits[1])
def compute_loss(logits, positions):
one_hot_positions = tf.one_hot(positions,
depth=seq_length,
dtype=tf.float32)
log_probs = tf.nn.log_softmax(logits, axis=1)
loss = -tf.reduce_sum(one_hot_positions * log_probs,
axis=1)
return loss
token_loss = (
compute_loss(start_logits, start_pos) +
compute_loss(end_logits, end_pos)) / 2.0 # per example
if not FLAGS.location_loss_for_nonpointable:
token_loss *= token_is_pointable
per_example_loss = FLAGS.class_loss_ratio * class_loss + (
1 - FLAGS.class_loss_ratio) * token_loss
total_loss += tf.reduce_sum(per_example_loss)
per_slot_per_example_loss[slot] = per_example_loss
per_slot_class_logits[slot] = class_logits
per_slot_start_logits[slot] = start_logits
per_slot_end_logits[slot] = end_logits
return (total_loss, per_slot_per_example_loss, per_slot_class_logits,
per_slot_start_logits, per_slot_end_logits)
use_one_hot_embeddings=False
)
(start_logits, end_logits) = cqa_model(bert_representation)
tvars = tf.trainable_variables()
initialized_variable_names = {}
if FLAGS.init_checkpoint:
(assignment_map, initialized_variable_names) = modeling.get_assigment_map_from_checkpoint(tvars,
FLAGS.init_checkpoint)
tf.train.init_from_checkpoint(FLAGS.init_checkpoint, assignment_map)
# compute loss
seq_length = modeling.get_shape_list(input_ids)[1]
def compute_loss(logits, positions):
one_hot_positions = tf.one_hot(
positions, depth=seq_length, dtype=tf.float32)
log_probs = tf.nn.log_softmax(logits, axis=-1)
loss = -tf.reduce_mean(tf.reduce_sum(one_hot_positions * log_probs, axis=-1))
return loss
# get the max prob for the predicted start/end position
start_probs = tf.nn.softmax(start_logits, axis=-1)
start_prob = tf.reduce_max(start_probs, axis=-1)
end_probs = tf.nn.softmax(end_logits, axis=-1)
end_prob = tf.reduce_max(end_probs, axis=-1)
start_loss = compute_loss(start_logits, start_positions)
end_loss = compute_loss(end_logits, end_positions)
def gec_create_model(bert_config, is_training, input_sequence,
input_mask, segment_ids, edit_sequence,
use_one_hot_embeddings, mode,
copy_weight,
use_bert_more,
insert_ids,
multitoken_insert_ids,
subtract_replaced_from_replacement):
"""Creates a classification model."""
# insert_ids: word ids of unigram inserts (list)
# multitoken_insert_ids: word_ids of bigram inserts (list of tuples of length 2)
# Defining the space of all possible edits:
# unk, sos and eos are dummy edits mapped to 0, 1 and 2 respectively
# copy is mapped to 3
# del is mapped to 4
num_appends = len(insert_ids) + len(multitoken_insert_ids)
num_replaces = num_appends # appends and replacements come from the same set (inserts and multitoken_inserts)
append_begin = 5 # First append edit (mapped to 5)
append_end = append_begin + num_appends - 1 #Last append edit
rep_begin = append_end + 1 # First replace edit
rep_end = rep_begin + num_replaces - 1 #Last replace edit
num_suffix_transforms = 58 #num of transformation edits
num_labels = 5 + num_appends + num_replaces + num_suffix_transforms # total number of edits
print("************ num of labels : {} ***************".format(num_labels))
config = bert_config
input_sequence_shape = modeling.get_shape_list(input_sequence,2)
batch_size = input_sequence_shape[0]
seq_len = input_sequence_shape[1]
if not use_bert_more: #default use of bert (without logit factorisation)
model = modeling.BertModel(
config=bert_config,
is_training=is_training,
input_ids=input_sequence,
input_mask=input_mask,
token_type_ids=segment_ids,
use_one_hot_embeddings=use_one_hot_embeddings)
output_layer = model.get_sequence_output()
else: # LOGIT FACTORISATION is On!
model = modified_modeling.BertModel(
config=bert_config,
is_training=is_training,
input_ids=input_sequence,
input_mask=input_mask,
token_type_ids=segment_ids,
use_one_hot_embeddings=use_one_hot_embeddings)
output_layer = model.get_sequence_output()
replace_layer = output_layer[:,seq_len:2*seq_len,:] #representation of replacement slots as described in paper
append_layer = output_layer[:,2*seq_len:3*seq_len,:] #representation of append slots as described in paper
output_layer = output_layer[:,0:seq_len,:]
output_layer_shape = modeling.get_shape_list(output_layer,3)
hidden_size = output_layer_shape[-1]
flattened_output_layer = tf.reshape(output_layer,[-1, hidden_size])
h_edit = flattened_output_layer
if use_bert_more:
h_word = flattened_output_layer
flattened_replace_layer = tf.reshape(replace_layer,[-1, hidden_size])
flattened_append_layer = tf.reshape(append_layer, [-1, hidden_size])
m_replace = flattened_replace_layer
m_append = flattened_append_layer
with tf.variable_scope("cls/predictions"):
with tf.variable_scope("transform"):
h_word = tf.layers.dense(
h_word,
units=bert_config.hidden_size,
activation=modeling.get_activation(bert_config.hidden_act),
kernel_initializer=modeling.create_initializer(
bert_config.initializer_range))
h_word = modeling.layer_norm(h_word)
with tf.variable_scope("cls/predictions",reuse=True):
with tf.variable_scope("transform",reuse=True):
m_replace = tf.layers.dense(
m_replace,
units=bert_config.hidden_size,
activation=modeling.get_activation(bert_config.hidden_act),
kernel_initializer=modeling.create_initializer(
bert_config.initializer_range))
m_replace = modeling.layer_norm(m_replace)
with tf.variable_scope("cls/predictions",reuse=True):
with tf.variable_scope("transform",reuse=True):
m_append = tf.layers.dense(
m_append,
units=bert_config.hidden_size,
activation=modeling.get_activation(bert_config.hidden_act),
kernel_initializer=modeling.create_initializer(
bert_config.initializer_range))
m_append = modeling.layer_norm(m_append)
word_embedded_input = model.word_embedded_input
flattened_word_embedded_input = tf.reshape(word_embedded_input, [-1, hidden_size])
labels = edit_sequence
edit_weights = tf.get_variable(
"edit_weights", [num_labels, hidden_size],
initializer=tf.truncated_normal_initializer(stddev=0.02))
if is_training:
h_edit = tf.nn.dropout(h_edit, keep_prob=0.9)
if use_bert_more:
# append/replace weight vector for a given append or replace operation
# correspond to word embedding for its token argument
# for multitoken append/replace (e.g. has been)
# weight vector is sum of word embeddings of token arguments
append_weights = edit_word_embedding_lookup(model.embedding_table, insert_ids,
use_one_hot_embeddings, config.vocab_size, config.hidden_size)
replace_weights = append_weights #tokens in replace and append vocab are same
#(i.e. inserts and multitoken_inserts)
multitoken_append_weights = wem_utils.edit_embedding_loopkup(model.embedding_table, multitoken_insert_ids,
use_one_hot_embeddings, config.vocab_size, config.hidden_size)
multitoken_replace_weights = multitoken_append_weights #tokens in replace and append vocab are same
#(i.e. inserts and multitoken_inserts)
append_weights = tf.concat([append_weights, multitoken_append_weights],0)
replace_weights = tf.concat([replace_weights, multitoken_replace_weights],0)
with tf.variable_scope("loss"):
edit_logits = tf.matmul(h_edit, edit_weights, transpose_b=True) #first term in eq3 in paper
logits = edit_logits
if use_bert_more:
#=============== inplace_word_logits==============# #2nd term in eq3 in paper
inplace_logit = tf.reduce_sum(h_word * flattened_word_embedded_input, axis=1, keepdims=True) #copy
#inplace_logit = tf.reduce_sum(m_replace * flattened_word_embedded_input, axis=1, keepdims=True) #copy
inplace_logit_appends = tf.tile(inplace_logit,[1,num_appends])
inplace_logit_transforms = tf.tile(inplace_logit,[1,num_suffix_transforms])
zero_3_logits = tf.zeros([batch_size*seq_len,3]) #unk sos eos
zero_1_logits = tf.zeros([batch_size*seq_len,1]) # del
zero_replace_logits = tf.zeros([batch_size*seq_len,num_replaces])
concat_list = [zero_3_logits, inplace_logit, zero_1_logits]\
+ [inplace_logit_appends]\
+ [zero_replace_logits]\
+ [inplace_logit_transforms]
inplace_word_logits = tf.concat(concat_list,1)
#======additional (insert,replace) logits ====# #3rd term in eqn3 in paper
zero_5_logits = tf.zeros([batch_size*seq_len,5])
append_logits = tf.matmul(m_append, append_weights, transpose_b=True)
if subtract_replaced_from_replacement:
replace_logits = replacement_minus_replaced_logits(m_replace,
flattened_word_embedded_input, replace_weights)
else:
replace_logits = tf.matmul(m_replace, replace_weights, transpose_b=True)
suffix_logits = tf.zeros([batch_size*seq_len,num_suffix_transforms])
concat_list = [zero_5_logits, append_logits, replace_logits, suffix_logits]
additional_logits = tf.concat(concat_list,1)
#====================================================#
logits = edit_logits + inplace_word_logits + additional_logits
logits_bias = tf.get_variable("output_bias", shape=[num_labels], initializer=tf.zeros_initializer())
logits += logits_bias
logits = tf.reshape(logits, [output_layer_shape[0], output_layer_shape[1], num_labels])
log_probs = tf.nn.log_softmax(logits, axis=-1)
probs = tf.nn.softmax(logits,axis=-1)
one_hot_labels = tf.one_hot(labels, depth=num_labels, dtype=tf.float32)
per_token_loss = -tf.reduce_sum(one_hot_labels * log_probs, axis=-1)
per_token_loss = per_token_loss * tf.to_float(input_mask)
mask = copy_weight*tf.to_float(tf.equal(labels,3)) + tf.to_float(tf.not_equal(labels,3))
masked_per_token_loss = per_token_loss * mask
per_example_loss = tf.reduce_sum(masked_per_token_loss, axis=-1)
loss = tf.reduce_mean(per_example_loss)
return (loss, per_example_loss, logits, probs)
def create_model(bert_config, is_training, input_ids, input_mask, segment_ids,
use_one_hot_embeddings):
"""Creates a classification model."""
model = modeling.BertModel(
config=bert_config,#たぶんこの設定にしたがってbertを呼び出すということ
is_training=is_training,
input_ids=input_ids,
input_mask=input_mask,
token_type_ids=segment_ids,
use_one_hot_embeddings=use_one_hot_embeddings)
final_hidden = model.get_sequence_output()#Bertの最終層
final_hidden_shape = modeling.get_shape_list(final_hidden, expected_rank=3)
batch_size = final_hidden_shape[0]
seq_length = final_hidden_shape[1]
hidden_size = final_hidden_shape[2]
final_hidden_matrix = tf.reshape(final_hidden,
[batch_size * seq_length, hidden_size])
#ここをTransformerにする
"""
output_weights = tf.get_variable(
#/はスコープの区切りを表す。だからcls/squad/output_weightsはclsスコープのsquadスコープのoutput_weightsという変数を表す
#変数がないときは定義し、ある時はそれを呼び出す
"cls/squad/output_weights", [2, hidden_size],
initializer=tf.truncated_normal_initializer(stddev=0.02))
output_bias = tf.get_variable(
#/はスコープの区切りを表す。だからcls/squad/output_weightsはclsスコープのsquadスコープのoutput_weightsという変数を表す
#変数がないときは定義し、ある時はそれを呼び出す
"cls/squad/output_bias", [2], initializer=tf.zeros_initializer())
logits = tf.matmul(final_hidden_matrix, output_weights, transpose_b=True)
logits = tf.nn.bias_add(logits, output_bias)
logits = tf.reshape(logits, [batch_size, seq_length, 2])
logits = tf.transpose(logits, [2, 0, 1])
unstacked_logits = tf.unstack(logits, axis=0)
(start_logits, end_logits) = (unstacked_logits[0], unstacked_logits[1])
return (start_logits, end_logits)
"""
#Transformer層
#bertの中のtransformerよりずっとスペック低くしている
transformer_outputs = modeling.transformer_model(input_tensor=final_hidden_matrix,
attention_mask=None,
hidden_size=5,
num_hidden_layers=2,
num_attention_heads=2,
intermediate_size=20,
intermediate_act_fn=modeling.gelu,
hidden_dropout_prob=0.1,
attention_probs_dropout_prob=0.1,
initializer_range=0.02,
do_return_all_layers=False)#現状Falseのみ
#線型層
output_weights = tf.get_variable(
#/はスコープの区切りを表す。だからcls/squad/output_weightsはclsスコープのsquadスコープのoutput_weightsという変数を表す
#変数がないときは定義し、ある時はそれを呼び出す
"cls/squad/output_weights", [30000, 5],
initializer=tf.truncated_normal_initializer(stddev=0.02))
output_bias = tf.get_variable(
#/はスコープの区切りを表す。だからcls/squad/output_weightsはclsスコープのsquadスコープのoutput_weightsという変数を表す
#変数がないときは定義し、ある時はそれを呼び出す
"cls/squad/output_bias", [30000], initializer=tf.zeros_initializer())
logits = tf.matmul(transformer_outputs, output_weights, transpose_b=True)
logits = tf.nn.bias_add(logits, output_bias)
#max
ids = tf.reduce_max(logits,axis=0)
#Transformerのテンソルとidを出力。損失を測るのに両方使うため
return (ids,transformer_outputs)
def model_fn(features, labels, mode, params): # pylint: disable=unused-argument
"""The `model_fn` for TPUEstimator."""
tf.logging.info("*** Features ***")
for name in sorted(features.keys()):
tf.logging.info(" name = %s, shape = %s" % (name, features[name].shape))
unique_ids = features["unique_ids"]
input_ids = features["input_ids"]
input_mask = features["input_mask"]
segment_ids = features["segment_ids"]
is_training = (mode == tf.estimator.ModeKeys.TRAIN)
(start_logits, end_logits) = create_model(
bert_config=bert_config,
is_training=is_training,
input_ids=input_ids,
input_mask=input_mask,
segment_ids=segment_ids,
use_one_hot_embeddings=use_one_hot_embeddings)
tvars = tf.trainable_variables()
initialized_variable_names = {}
scaffold_fn = None
if init_checkpoint:
(assignment_map, initialized_variable_names
) = modeling.get_assignment_map_from_checkpoint(tvars, init_checkpoint)
if use_tpu:
def tpu_scaffold():
tf.train.init_from_checkpoint(init_checkpoint, assignment_map)
return tf.train.Scaffold()
scaffold_fn = tpu_scaffold
else:
tf.train.init_from_checkpoint(init_checkpoint, assignment_map)
tf.logging.info("**** Trainable Variables ****")
for var in tvars:
init_string = ""
if var.name in initialized_variable_names:
init_string = ", *INIT_FROM_CKPT*"
tf.logging.info(" name = %s, shape = %s%s", var.name, var.shape,
init_string)
output_spec = None
if mode == tf.estimator.ModeKeys.TRAIN:
seq_length = modeling.get_shape_list(input_ids)[1]
def compute_loss(logits, positions):
one_hot_positions = tf.one_hot(
positions, depth=seq_length, dtype=tf.float32)
log_probs = tf.nn.log_softmax(logits, axis=-1)
loss = -tf.reduce_mean(
tf.reduce_sum(one_hot_positions * log_probs, axis=-1))
return loss
start_positions = features["start_positions"]
end_positions = features["end_positions"]
start_loss = compute_loss(start_logits, start_positions)
end_loss = compute_loss(end_logits, end_positions)
total_loss = (start_loss + end_loss) / 2.0
train_op = optimization.create_optimizer(
total_loss, learning_rate, num_train_steps, num_warmup_steps, use_tpu)
output_spec = tf.contrib.tpu.TPUEstimatorSpec(
mode=mode,
loss=total_loss,
train_op=train_op,
scaffold_fn=scaffold_fn)
elif mode == tf.estimator.ModeKeys.PREDICT:
predictions = {
"unique_ids": unique_ids,
"start_logits": start_logits,
"end_logits": end_logits,
}
output_spec = tf.contrib.tpu.TPUEstimatorSpec(
mode=mode, predictions=predictions, scaffold_fn=scaffold_fn)
else:
raise ValueError(
"Only TRAIN and PREDICT modes are supported: %s" % (mode))
return output_spec
