#! -*- coding: utf-8 -*-
import os
# os.environ["TF_KERAS"] = "1"
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
import glob
import random
from tqdm import tqdm
import numpy as np
import pandas as pd
from bert4keras.backend import keras, K
from bert4keras.layers import Loss
from bert4keras.models import build_transformer_model
from bert4keras.tokenizers import Tokenizer, load_vocab
from bert4keras.optimizers import Adam
from bert4keras.snippets import sequence_padding, open
from bert4keras.snippets import DataGenerator, AutoRegressiveDecoder
from keras.models import Model
import tensorflow as tf
from keras.backend import set_session
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
set_session(tf.Session(config=config)) # 此处不同
global graph
graph = tf.get_default_graph()
sess = tf.Session(graph=graph)
set_session(sess)
# global graph,model
# graph = tf.get_default_graph()
# sess = tf.Session(graph=graph)
# K.set_session(sess)
# 基本参数
class TotalLoss(Loss):
"""loss分两部分，一是seq2seq的交叉熵，二是相似度的交叉熵。
"""
def compute_loss(self, inputs, mask=None):
loss1 = self.compute_loss_of_seq2seq(inputs, mask)
loss2 = self.compute_loss_of_similarity(inputs, mask)
self.add_metric(loss1, name='seq2seq_loss')
self.add_metric(loss2, name='similarity_loss')
return loss1 + loss2
def compute_loss_of_seq2seq(self, inputs, mask=None):
y_true, y_mask, _, y_pred = inputs
y_true = y_true[:, 1:] # 目标token_ids
y_mask = y_mask[:, 1:] # segment_ids，刚好指示了要预测的部分
y_pred = y_pred[:, :-1] # 预测序列，错开一位
loss = K.sparse_categorical_crossentropy(y_true, y_pred)
loss = K.sum(loss * y_mask) / K.sum(y_mask)
return loss
def compute_loss_of_similarity(self, inputs, mask=None):
_, _, y_pred, _ = inputs
y_true = self.get_labels_of_similarity(y_pred) # 构建标签
y_pred = K.l2_normalize(y_pred, axis=1) # 句向量归一化
similarities = K.dot(y_pred, K.transpose(y_pred)) # 相似度矩阵
similarities = similarities - K.eye(K.shape(y_pred)[0]) * 1e12 # 排除对角线
similarities = similarities * 30 # scale
loss = K.categorical_crossentropy(
y_true, similarities, from_logits=True
)
return loss
def get_labels_of_similarity(self, y_pred):
idxs = K.arange(0, K.shape(y_pred)[0])
idxs_1 = idxs[None, :]
idxs_2 = (idxs + 1 - idxs % 2 * 2)[:, None]
labels = K.equal(idxs_1, idxs_2)
labels = K.cast(labels, K.floatx())
return labels
class GenerateModel(object):
def __init__(self):
self.epoch_acc_vel = 0
self.config_path = r'./chinese_roberta_wwm_ext_L-12_H-768_A-12/bert_config.json'
self.checkpoint_path = r'./chinese_roberta_wwm_ext_L-12_H-768_A-12/bert_model.ckpt'
self.dict_path = r'./chinese_roberta_wwm_ext_L-12_H-768_A-12/vocab.txt'
self.maxlen = 120
self.novel_maxlen = 60
def device_setup(self):
token_dict, keep_tokens = load_vocab(
dict_path=self.dict_path,
simplified=True,
startswith=['[PAD]', '[UNK]', '[CLS]', '[SEP]'],
)
tokenizer = Tokenizer(token_dict, do_lower_case=True)
# model = build_transformer_model(
# self.config_path,
# self.checkpoint_path,
# application='unilm',
# keep_tokens=keep_tokens, # 只保留keep_tokens中的字，精简原字表
# )
bert = build_transformer_model(
self.config_path,
self.checkpoint_path,
with_pool='linear',
application='unilm',
keep_tokens=keep_tokens,
return_keras_model=False,
)
encoder = keras.models.Model(bert.model.inputs, bert.model.outputs[0])
seq2seq = keras.models.Model(bert.model.inputs, bert.model.outputs[1])
# output = CrossEntropy(2)(model.inputs + model.outputs)
#
# model = Model(model.inputs, output)
# model = Model(model.inputs, model.outputs)
outputs = TotalLoss([2, 3])(bert.model.inputs + bert.model.outputs)
model = keras.models.Model(bert.model.inputs, outputs)
path_model = './output_simbert_yy/best_simbertmodel_datasim.weights'
model.load_weights(path_model)
return encoder,seq2seq, tokenizer
class CrossEntropy(Loss):
"""交叉熵作为loss，并mask掉输入部分
"""
def compute_loss(self, inputs, mask=None):
y_true, y_mask, y_pred = inputs
y_true = y_true[:, 1:] # 目标token_ids
y_mask = y_mask[:, 1:] # segment_ids，刚好指示了要预测的部分
y_pred = y_pred[:, :-1] # 预测序列，错开一位
loss = K.sparse_categorical_crossentropy(y_true, y_pred)
loss = K.sum(loss * y_mask) / K.sum(y_mask)
return loss
class AutoTitle(AutoRegressiveDecoder):
"""seq2seq解码器
"""
def __init__(self, model, tokenizer, start_id, end_id, maxlen, minlen=1):
super(AutoTitle, self).__init__(start_id, end_id, maxlen, minlen)
self.model = model
self.tokenizer = tokenizer
self.start_id = start_id
self.end_id = end_id
self.minlen = minlen
self.models = {}
if start_id is None:
self.first_output_ids = np.empty((1, 0), dtype=int)
else:
self.first_output_ids = np.array([[self.start_id]])
def data_generator(self, inputs, output_ids):
try:
batch_token_ids, batch_segment_ids = [], []
if output_ids == []:
for txt in inputs:
token_ids, segment_ids = self.tokenizer.encode(txt, maxlen=120)
batch_token_ids.append(token_ids)
batch_segment_ids.append(segment_ids)
else:
for txt,output_id in zip(inputs, output_ids):
token_ids, segment_ids = self.tokenizer.encode(txt, output_id)
batch_token_ids.append(token_ids[:-1])
batch_segment_ids.append(segment_ids[:-1])
batch_token_ids = sequence_padding(batch_token_ids)
batch_segment_ids = sequence_padding(batch_segment_ids)
except:
print(inputs,output_ids)
return batch_token_ids, batch_segment_ids
def beam_search_batch_(self, inputs, topk, states=None, temperature=1, min_ends=1):
"""beam search解码
说明：这里的topk即beam size；
返回：最优解码序列。
"""
inputs = [np.array([i]) for i in inputs]
output_ids, output_scores = self.first_output_ids, np.zeros(1)
for step in range(self.maxlen):
scores, states = self.predict(
inputs, output_ids, states, temperature, 'logits'
) # 计算当前得分
if step == 0: # 第1步预测后将输入重复topk次
inputs = [np.repeat(i, topk, axis=0) for i in inputs]
scores = output_scores.reshape((-1, 1)) + scores # 综合累积得分
indices = scores.argpartition(-topk, axis=None)[-topk:] # 仅保留topk
indices_1 = indices // scores.shape[1] # 行索引
indices_2 = (indices % scores.shape[1]).reshape((-1, 1)) # 列索引
output_ids = np.concatenate([output_ids[indices_1], indices_2],
1) # 更新输出
output_scores = np.take_along_axis(
scores, indices, axis=None
) # 更新得分
is_end = output_ids[:, -1] == self.end_id # 标记是否以end标记结束
end_counts = (output_ids == self.end_id).sum(1) # 统计出现的end标记
if output_ids.shape[1] >= self.minlen: # 最短长度判断
best = output_scores.argmax() # 得分最大的那个
if is_end[best] and end_counts[best] >= min_ends: # 如果已经终止
return output_ids[best] # 直接输出
else: # 否则，只保留未完成部分
flag = ~is_end | (end_counts < min_ends) # 标记未完成序列
if not flag.all(): # 如果有已完成的
inputs = [i[flag] for i in inputs] # 扔掉已完成序列
output_ids = output_ids[flag] # 扔掉已完成序列
output_scores = output_scores[flag] # 扔掉已完成序列
end_counts = end_counts[flag] # 扔掉已完成end计数
topk = flag.sum() # topk相应变化
# 达到长度直接输出
return output_ids[output_scores.argmax()]
def random_sample_batch(
self,
inputs,
n,
topk=None,
topp=None,
states=None,
temperature=1,
min_ends=1
):
"""随机采样n个结果
说明：非None的topk表示每一步只从概率最高的topk个中采样；而非None的topp
表示每一步只从概率最高的且概率之和刚好达到topp的若干个token中采样。
返回：n个解码序列组成的list。
"""
inputs = [np.array([i for j in i]) for i in inputs]
output_ids = self.first_output_ids
results = []
for step in range(self.maxlen):
probas, states = self.predict(
inputs, output_ids, states, temperature, 'probas'
) # 计算当前概率
probas /= probas.sum(axis=1, keepdims=True) # 确保归一化
if step == 0: # 第1步预测后将结果重复n次
probas = np.repeat(probas, n, axis=0)
inputs = [np.repeat(i, n, axis=0) for i in inputs]
output_ids = np.repeat(output_ids, n, axis=0)
if topk is not None:
k_indices = probas.argpartition(-topk,
axis=1)[:, -topk:] # 仅保留topk
probas = np.take_along_axis(probas, k_indices, axis=1) # topk概率
probas /= probas.sum(axis=1, keepdims=True) # 重新归一化
if topp is not None:
p_indices = probas.argsort(axis=1)[:, ::-1] # 从高到低排序
probas = np.take_along_axis(probas, p_indices, axis=1) # 排序概率
cumsum_probas = np.cumsum(probas, axis=1) # 累积概率
flag = np.roll(cumsum_probas >= topp, 1, axis=1) # 标记超过topp的部分
flag[:, 0] = False # 结合上面的np.roll，实现平移一位的效果
probas[flag] = 0 # 后面的全部置零
probas /= probas.sum(axis=1, keepdims=True) # 重新归一化
sample_func = lambda p: np.random.choice(len(p), p=p) # 按概率采样函数
sample_ids = np.apply_along_axis(sample_func, 1, probas) # 执行采样
sample_ids = sample_ids.reshape((-1, 1)) # 对齐形状
if topp is not None:
sample_ids = np.take_along_axis(
p_indices, sample_ids, axis=1
) # 对齐原id
if topk is not None:
sample_ids = np.take_along_axis(
k_indices, sample_ids, axis=1
) # 对齐原id
output_ids = np.concatenate([output_ids, sample_ids], 1) # 更新输出
is_end = output_ids[:, -1] == self.end_id # 标记是否以end标记结束
end_counts = (output_ids == self.end_id).sum(1) # 统计出现的end标记
if output_ids.shape[1] >= self.minlen: # 最短长度判断
flag = is_end & (end_counts >= min_ends) # 标记已完成序列
if flag.any(): # 如果有已完成的
for ids in output_ids[flag]: # 存好已完成序列
results.append(ids)
flag = (flag == False) # 标记未完成序列
inputs = [i[flag] for i in inputs] # 只保留未完成部分输入
output_ids = output_ids[flag] # 只保留未完成部分候选集
end_counts = end_counts[flag] # 只保留未完成部分end计数
if len(output_ids) == 0:
break
# 如果还有未完成序列，直接放入结果
for ids in output_ids:
results.append(ids)
# 返回结果
return results
def random_sample_and_beam_search(
self,
inputs,
n,
topk=None,
topp=None,
states=None,
temperature=1,
min_ends=1
):
"""随机采样n个结果
说明：非None的topk表示每一步只从概率最高的topk个中采样；而非None的topp
表示每一步只从概率最高的且概率之和刚好达到topp的若干个token中采样。
返回：n个解码序列组成的list。
"""
whether_end_b = False
results_r = []
results_b = []
# index_r = [i for i in range(n)]
# index_b = [i for i in range(topk)]
index_r = np.arange(n)
index_b = np.arange(topk)
inputs = [np.array([i]) for i in inputs]
output_ids, output_scores = self.first_output_ids, np.zeros(1)
results = []
for step in range(self.maxlen):
beam_n = len(index_b)
probas, states = self.predict(
inputs, output_ids, states, temperature, 'probas'
) # 计算当前概率
probas = probas / probas.sum(axis=1, keepdims=True) # 确保归一化
if step == 0: # 第1步预测后将结果重复n次
probas = np.repeat(probas, n + topk, axis=0)
inputs_r = [np.repeat(i, n, axis=0) for i in inputs]
output_ids = np.repeat(output_ids, n + topk, axis=0)
inputs_b = [np.repeat(i, topk, axis=0) for i in inputs]
else:
if whether_end_b == False:
inputs_r = [i[:-beam_n, :] for i in inputs]
inputs_b = [i[-beam_n:, :] for i in inputs]
else:
inputs_r = inputs
if whether_end_b == False:
probas_r = probas[:-beam_n, :]
else:
probas_r = probas
if step == 0:
probas_b = probas[0,:]
else:
probas_b = probas[-beam_n:, :]
if whether_end_b == False:
output_ids_r = output_ids[:-beam_n, :]
output_ids_b = output_ids[-beam_n:, :]
else:
output_ids_r = output_ids
k_indices = probas_r.argpartition(-topk,
axis=1)[:, -topk:] # 仅保留topk
probas_r = np.take_along_axis(probas_r, k_indices, axis=1) # topk概率
probas_r /= probas_r.sum(axis=1, keepdims=True) # 重新归一化
if whether_end_b == False:
scores = output_scores.reshape((-1, 1)) + probas_b # 综合累积得分
indices = scores.argpartition(-topk, axis=None)[-topk:] # 仅保留topk
indices_1 = indices // scores.shape[1] # 行索引
indices_2 = (indices % scores.shape[1]).reshape((-1, 1)) # 列索引
try:
output_ids_b = np.concatenate([output_ids_b[indices_1], indices_2],
1) # 更新输出
except:
print(output_ids_b.shape)
print(indices_1)
print(indices_2)
exit()
output_scores = np.take_along_axis(
scores, indices, axis=None
) # 更新得分
sample_func = lambda p: np.random.choice(len(p), p=p) # 按概率采样函数
try:
sample_ids = np.apply_along_axis(sample_func, 1, probas_r) # 执行采样
except:
print(probas_r)
sample_ids = sample_ids.reshape((-1, 1)) # 对齐形状
if topk is not None:
sample_ids = np.take_along_axis(
k_indices, sample_ids, axis=1
) # 对齐原id
output_ids_r = np.concatenate([output_ids_r, sample_ids], 1) # 更新输出
# output_ids = np.concatenate([output_ids_r, output_ids_b], 0)
if whether_end_b == False:
is_end_r = output_ids_r[:, -1] == self.end_id # 标记是否以end标记结束
is_end_b = output_ids_b[:, -1] == self.end_id # 标记是否以end标记结束
else:
is_end_r = output_ids_r[:, -1] == self.end_id
if whether_end_b == False:
end_counts_r = (output_ids_r == self.end_id).sum(1) # 统计出现的end标记
end_counts_b = (output_ids_b == self.end_id).sum(1) # 统计出现的end标记
else:
end_counts_r = (output_ids_r == self.end_id).sum(1)
# random_serach
if output_ids_r.shape[1] >= self.minlen: # 最短长度判断
flag = is_end_r & (end_counts_r >= min_ends) # 标记已完成序列
if flag.any(): # 如果有已完成的
for ids in output_ids_r[flag]: # 存好已完成序列
results_r.append(ids)
flag = (flag == False) # 标记未完成序列
try:
index_r = index_r[flag]
except:
print("flag",flag)
print("index_r",index_r)
inputs_r = [i[flag] for i in inputs_r] # 只保留未完成部分输入
output_ids_r = output_ids_r[flag] # 只保留未完成部分候选集
end_counts_r = end_counts_r[flag] # 只保留未完成部分end计数
# beam_serach
if whether_end_b == False:
if output_ids_b.shape[1] >= self.minlen: # 最短长度判断
best = output_scores.argmax() # 得分最大的那个
if is_end_b[best] and end_counts_b[best] >= min_ends: # 如果已经终止
results_b.append(output_ids_b[best]) # 直接输出
whether_end_b = True
else: # 否则，只保留未完成部分
flag_b = ~is_end_b | (end_counts_b < min_ends) # 标记未完成序列
if not flag_b.all(): # 如果有已完成的
index_b = index_b[flag_b]
inputs_b = [i[flag_b] for i in inputs_b] # 扔掉已完成序列
output_ids_b = output_ids_b[flag_b] # 扔掉已完成序列
output_scores = output_scores[flag_b] # 扔掉已完成序列
end_counts_b = end_counts_b[flag_b] # 扔掉已完成end计数
topk = flag_b.sum() # topk相应变化
if whether_end_b == False and len(output_ids_r) != 0:
token_r = inputs_r[0]
sample_ids_r = inputs_r[1]
token_b = inputs_b[0]
sample_ids_b = inputs_b[1]
token = np.concatenate([token_r,token_b],0)
sample_ids = np.concatenate([sample_ids_r,sample_ids_b],0)
inputs = [token,sample_ids]
output_ids = np.concatenate([output_ids_r, output_ids_b], 0)
elif whether_end_b == True and len(output_ids_r) != 0:
inputs = inputs_r
output_ids = output_ids_r
elif whether_end_b == False and len(output_ids_r) == 0:
inputs = inputs_b
output_ids = output_ids_b
else:
break
# 如果还有未完成序列，直接放入结果
for ids in output_ids:
results.append(ids)
# 返回结果
return results_r, results_b
def beam_search(self, inputs, topk, states=None, temperature=1, min_ends=1):
"""beam search解码
说明：这里的topk即beam size；
返回：最优解码序列。
"""
inputs = [np.array([i]) for i in inputs]
output_ids, output_scores = self.first_output_ids, np.zeros(1)
for step in range(self.maxlen):
scores, states = self.predict(
inputs, output_ids, states, temperature, 'probas'
) # 计算当前得分
if step == 0: # 第1步预测后将输入重复topk次
inputs = [np.repeat(i, topk, axis=0) for i in inputs]
scores = output_scores.reshape((-1, 1)) + scores # 综合累积得分
indices = scores.argpartition(-topk, axis=None)[-topk:] # 仅保留topk
indices_1 = indices // scores.shape[1] # 行索引
indices_2 = (indices % scores.shape[1]).reshape((-1, 1)) # 列索引
output_ids = np.concatenate([output_ids[indices_1], indices_2],
1) # 更新输出
output_scores = np.take_along_axis(
scores, indices, axis=None
) # 更新得分
is_end = output_ids[:, -1] == self.end_id # 标记是否以end标记结束
end_counts = (output_ids == self.end_id).sum(1) # 统计出现的end标记
if output_ids.shape[1] >= self.minlen: # 最短长度判断
best = output_scores.argmax() # 得分最大的那个
if is_end[best] and end_counts[best] >= min_ends: # 如果已经终止
return output_ids[best] # 直接输出
else: # 否则，只保留未完成部分
flag = ~is_end | (end_counts < min_ends) # 标记未完成序列
if not flag.all(): # 如果有已完成的
inputs = [i[flag] for i in inputs] # 扔掉已完成序列
output_ids = output_ids[flag] # 扔掉已完成序列
output_scores = output_scores[flag] # 扔掉已完成序列
end_counts = end_counts[flag] # 扔掉已完成end计数
topk = flag.sum() # topk相应变化
# 达到长度直接输出
return output_ids[output_scores.argmax()]
# def beam_search_batch_ceshi(
# self,
# inputs_str,
# states=None,
# temperature=1,
# min_ends=1,
# num_beam=3
# ):
# """随机采样n个结果
# 说明：非None的topk表示每一步只从概率最高的topk个中采样；而非None的topp
# 表示每一步只从概率最高的且概率之和刚好达到topp的若干个token中采样。
# 返回：n个解码序列组成的list。
# """
# output_str = []
# # token_ids, segment_ids = self.data_generator(inputs, output_ids)
# batch_nums = len(inputs_str)
# # output_ids = np.empty((batch_nums, 0), dtype=int)
# output_ids = []
# generated = [Beamdataone(num_beam, inputs_str[i], self.end_id, self.minlen, min_ends, self.tokenizer, output_ids) for i in range(batch_nums)]
# # index_data = [i for i in range(batch_nums)]
# for step in range(self.maxlen):
# # if step == 0:
# # token_ids, segment_ids = self.data_generator(inputs_str, output_str)
# # else:
# inputs_str, output_str = [], []
# token_ids_batch, segment_ids_batch, output_ids_batch = [], [], []
# batch_input_num_beam_num = []
# for i in generated:
# token_ids, segment_ids = i.token_ids, i.segment_ids
# token_ids_batch.append(token_ids)
# segment_ids_batch.append(segment_ids)
# output_ids_batch.append(i.output_ids)
# if step != 0:
# batch_input_num_beam_num.append(token_ids.shape()[0])
#
# token_ids_batch = sequence_padding(token_ids_batch)
# segment_ids_batch = sequence_padding(segment_ids_batch)
# output_ids_batch = np.array(output_ids_batch)
# # if step == 0:
#
# inputs = [token_ids_batch, segment_ids_batch]
# if step == 0:
# probas = self.predict(
# inputs, 0, states, temperature, 'probas'
# ) # 计算当前概率
#
# else:
# probas = self.predict(
# inputs, output_ids_batch, states, temperature, 'logits'
# ) # 计算当前概率
#
# if step != 0:
# num = 0
# for index in range(len(batch_input_num_beam_num)-1):
# cc = num
# generated[index].add_data(step, probas[cc:num,:,:])
#
# i = 0
# while True:
# if i == len(generated) - 1:
# break
# bool = generated[i].over
# if bool == True:
# generated.pop(i)
# else:
# i += 1
# else:
# for index in range(len(generated)):
# generated[index].add_data(step, probas[0][index,:])
def beam_search_batch(
self,
inputs_str,
states=None,
temperature=1,
min_ends=1,
num_beam=3
):
"""随机采样n个结果
说明：非None的topk表示每一步只从概率最高的topk个中采样；而非None的topp
表示每一步只从概率最高的且概率之和刚好达到topp的若干个token中采样。
返回：n个解码序列组成的list。
"""
output_str = []
# token_ids, segment_ids = self.data_generator(inputs, output_ids)
batch_nums = len(inputs_str)
return_str_batch = [0] * batch_nums
# output_ids = np.empty((batch_nums, 0), dtype=int)
output_ids = np.empty((1, 0), dtype=int)
generated = [Beamdataone(num_beam, i, [inputs_str[i]], self.end_id, self.minlen, min_ends, self.tokenizer, output_ids) for i in range(batch_nums)]
# index_data = [i for i in range(batch_nums)]
for step in range(self.maxlen):
# if step == 0:
# token_ids, segment_ids = self.data_generator(inputs_str, output_str)
# else:
# inputs_str, output_str = [], []
text_batch, output_str_batch = [], []
batch_input_num_beam_num = []
for i in generated:
text = i.text
text_batch.extend(text)
if i.output_str != "":
output_str_batch.extend(i.output_str)
if step != 0:
batch_input_num_beam_num.append(i.num_beams)
token_ids, segment_ids = self.data_generator(text_batch, output_str_batch)
# token_ids_batch = sequence_padding(token_ids_batch)
# segment_ids_batch = sequence_padding(segment_ids_batch)
# output_ids_batch = np.array(output_ids_batch)
# if step == 0:
inputs = [token_ids, segment_ids]
probas = self.predict_batch(
inputs
) # 计算当前概率
probas_new = []
probas_bool = np.array(token_ids, dtype=bool)
# np.array(np.where(probas_bool == True))
for i, sentence in enumerate(probas_bool):
lie = np.array(np.where(sentence == True))[0]
probas_new.append(probas[i, lie[-1]])
probas = np.array(probas_new)
if step != 0:
num = 0
print()
if len(generated) > 1:
index = 0
for index in range(len(batch_input_num_beam_num)-1):
cc = num
num += batch_input_num_beam_num[index]
generated[index].add_data(step, probas[cc:num,:])
generated[index+1].add_data(step, probas[num:,:])
else:
generated[0].add_data(step, probas[:,:])
else:
for index in range(len(generated)):
generated[index].add_data(step, probas[index,:])
# i = 0
# while True:
# bool_ = generated[i].over
# if bool_ == True:
# one_sentence = generated.pop(i)
# return_str_batch[i] = one_sentence.return_str
# if i > len(generated) - 1:
# break
# else:
# i += 1
# if i > len(generated) - 1:
# break
generated_new = []
for i in range(len(generated)):
bool_ = generated[i].over
if bool_ == False:
generated_new.append(generated[i])
else:
return_str_batch[generated[i].batch_id] = generated[i].return_str
generated = generated_new
if generated == []:
return return_str_batch
return return_str_batch
def top_batch(
self,
inputs_str,
temperature=1,
min_ends=1
):
"""随机采样n个结果
说明：非None的topk表示每一步只从概率最高的topk个中采样；而非None的topp
表示每一步只从概率最高的且概率之和刚好达到topp的若干个token中采样。
返回：n个解码序列组成的list。
"""
output_str = []
# token_ids, segment_ids = self.data_generator(inputs, output_ids)
batch_nums = len(inputs_str)
output_ids = np.empty((batch_nums, 0), dtype=int)
results = [[] for i in range(batch_nums)]
index_data = [i for i in range(batch_nums)]
for step in range(self.maxlen):
token_ids, segment_ids = self.data_generator(inputs_str, output_str)
inputs = [token_ids, segment_ids]
probas = self.predict_batch(
inputs
) # 计算当前概率
# probas /= probas.sum(axis=1, keepdims=True) # 确保归一化
probas_new = []
probas_bool = np.array(token_ids, dtype=bool)
# np.array(np.where(probas_bool == True))
for i,sentence in enumerate(probas_bool):
lie = np.array(np.where(sentence == True))[0]
probas_new.append(probas[i,lie[-1]])
probas = np.array(probas_new)
k_indices = np.argmax(probas,axis=1) # 仅保留topk
k_indices = k_indices.reshape(-1,1)
sample_ids = k_indices
output_ids = np.concatenate([output_ids, sample_ids], 1) # 更新输出
is_end = output_ids[:, -1] == self.end_id # 标记是否以end标记结束
end_counts = (output_ids == self.end_id).sum(1) # 统计出现的end标记
if output_ids.shape[1] >= self.minlen: # 最短长度判断
flag = is_end & (end_counts >= min_ends) # 标记已完成序列
if flag.any(): # 如果有已完成的
index = np.array(np.where(flag == True))[0]
pop_index = []
for i in index:
results[index_data[i]] = output_ids[i]
pop_index.append(index_data[i])
for i in pop_index:
index_data.remove(i)
# for ids in output_ids[flag]: # 存好已完成序列
# results.append(ids)
flag = (flag == False) # 标记未完成序列
inputs_str = [inputs_str[i] for i in index_data] # 只保留未完成部分输入
output_ids = output_ids[flag] # 只保留未完成部分候选集
if len(output_ids) == 0:
break
else:
output_str = [tokenizer.decode(ids) for ids in output_ids]
else:
output_str = [tokenizer.decode(ids) for ids in output_ids]
# 如果还有未完成序列，直接放入结果
# for ids in output_ids:
# results.append(ids)
# 返回结果
return results
@AutoRegressiveDecoder.wraps(default_rtype='probas')
def predict(self, inputs, output_ids, states):
token_ids, segment_ids = inputs
token_ids = np.concatenate([token_ids, output_ids], 1)
segment_ids = np.concatenate([segment_ids, np.ones_like(output_ids)], 1)
with graph.as_default():
K.set_session(sess)
nodes = self.last_token(self.model).predict([token_ids, segment_ids])
return nodes
# return self.last_token(self.model).predict([token_ids, segment_ids])
def predict_batch(self, inputs):
# inputs, output_ids, states, temperature, 'probas'
token_ids, segment_ids = inputs
# token_ids = np.concatenate([token_ids, output_ids], 1)
# segment_ids = np.concatenate([segment_ids, np.ones_like(output_ids)], 1)
with graph.as_default():
K.set_session(sess)
nodes = self.model.predict([token_ids, segment_ids])
return nodes
# return self.last_token(self.model).predict([token_ids, segment_ids])
def generate(self, text, topk=1):
token_ids, segment_ids = self.tokenizer.encode(text, maxlen=256)
output_ids = self.beam_search([token_ids, segment_ids],
topk=topk) # 基于beam search
return self.tokenizer.decode(output_ids)
def generate_random(self, text, n=20, topk=5):
if isinstance(text, list):
text = text[0]
token_ids, segment_ids = self.tokenizer.encode(text, maxlen=120)
output_ids = self.random_sample([token_ids, segment_ids], n, topk) # 基于随机采样
return [tokenizer.decode(ids) for ids in output_ids]
def generate_random_topp(self, text, n=20, topp=0.98):
if isinstance(text, list):
text = text[0]
token_ids, segment_ids = self.tokenizer.encode(text, maxlen=120)
output_ids = self.random_sample([token_ids, segment_ids], n, topp=topp) # 基于随机采样
return [tokenizer.decode(ids) for ids in output_ids]
def generate_top(self, text):
output_ids = self.top_batch(text) # 基于随机采样
return [tokenizer.decode(ids) for ids in output_ids]
def generate_beam_search_batch(self, text):
output_str = self.beam_search_batch(text) # 基于随机采样
return output_str
def generate_random_sample_and_beam_search(self, text, n=20, topk=5):
text = text[0]
token_ids, segment_ids = self.tokenizer.encode(text, maxlen=120)
output_ids_r, output_ids_b = self.random_sample_and_beam_search([token_ids, segment_ids], n=n,
topk=topk) # 基于beam search
output_str_r = [self.tokenizer.decode(ids) for ids in output_ids_r]
output_str_b = [self.tokenizer.decode(ids) for ids in output_ids_b]
return output_str_r, output_str_b
def gen_synonyms(self, text, n=20):
""""含义： 产生sent的n个相似句，然后返回最相似的k个。
做法：用seq2seq生成，并用encoder算相似度并排序。
"""
r = self.generate_random_topp(text, n)
r = [i for i in set(r) if i != text]
r = [text] + r
X, S = [], []
for t in r:
x, s = tokenizer.encode(t)
X.append(x)
S.append(s)
X = sequence_padding(X)
S = sequence_padding(S)
Z = encoder.predict([X, S])
Z /= (Z ** 2).sum(axis=1, keepdims=True) ** 0.5
argsort = np.dot(Z[1:], -Z[0]).argsort()
return [r[i + 1] for i in argsort]
def gen_synonyms_short(self, text, n=20, len_s = 0.9):
""""含义： 产生sent的n个相似句，然后返回最相似的k个。
做法：用seq2seq生成，并用encoder算相似度并排序。
"""
if isinstance(text, list):
text = text[0]
new_text_len = int(len(text) * len_s)
r = self.generate_random(text, n)
r = [i for i in set(r) if i != text]
r = [text] + r
X, S = [], []
for t in r:
x, s = tokenizer.encode(t)
X.append(x)
S.append(s)
X = sequence_padding(X)
S = sequence_padding(S)
with graph.as_default():
K.set_session(sess)
Z = encoder.predict([X, S])
Z /= (Z ** 2).sum(axis=1, keepdims=True) ** 0.5
argsort = np.dot(Z[1:], -Z[0]).argsort()
sentence_list = [r[i + 1] for i in argsort]
return_list = []
for i in sentence_list:
if len(i) < new_text_len:
return_list.append(i)
break
for i in sentence_list:
if new_text_len <len(i) < len(text):
return_list.append(i)
break
if return_list != []:
return return_list[0]
else:
return sentence_list[0]
class BeamHypotheses(object):
def __init__(self, num_beams, max_length, length_penalty):
"""
Initialize n-best list of hypotheses.
"""
self.max_length = max_length - 1 # ignoring bos_token
self.num_beams = num_beams
self.beams = []
self.worst_score = 1e9
def __len__(self):
"""
Number of hypotheses in the list.
"""
return len(self.beams)
def add(self, hyp, sum_logprobs):
"""
Add a new hypothesis to the list.
"""
score = sum_logprobs / len(hyp) ** self.length_penalty
if len(self) < self.num_beams or score > self.worst_score:
# 可更新的情况：数量未饱和或超过最差得分
self.beams.append((score, hyp))
if len(self) > self.num_beams:
# 数量饱和需要删掉一个最差的
sorted_scores = sorted([(s, idx) for idx, (s, _) in enumerate(self.beams)])
del self.beams[sorted_scores[0][1]]
self.worst_score = sorted_scores[1][0]
else:
self.worst_score = min(score, self.worst_score)
def is_done(self, best_sum_logprobs, cur_len=None):
"""
相关样本是否已经完成生成。
best_sum_logprobs是新的候选序列中的最高得分。
"""
if len(self) < self.num_beams:
return False
else:
if cur_len is None:
cur_len = self.max_length
cur_score = best_sum_logprobs / cur_len ** self.length_penalty
# 是否最高分比当前保存的最低分还差
ret = self.worst_score >= cur_score
return ret
class Beamdataone(object):
def __init__(self, num_beams, batch_id, text, end_id, minlen, min_ends, tokenizer, output_ids):
"""
Initialize n-best list of hypotheses.
"""
self.num_beams = num_beams
self.batch_id = batch_id
self.beams = []
self.minlen = minlen
self.min_ends = min_ends
self.end_id = end_id
self.text = text
self.output_scores = np.zeros(1)
self.output_ids = output_ids
self.return_str = ""
self.over = False
self.tokenizer = tokenizer
# self.data()
self.output_str = ""
def add_data(self, step, output_probas):
'''
还存有的数据，直接可以被迭代，
@param text:
@return:
'''
# inputs = [np.array([i]) for i in inputs]
# output_ids, output_scores = self.first_output_ids, np.zeros(1)
#
# scores, states = self.predict(
# inputs, output_ids, states, temperature, 'logits'
# ) # 计算当前得分
# if step == 0: # 第1步预测后将输入重复topk次
# inputs = [np.repeat(i, self.num_beams, axis=0) for i in self.inputs]
# inputs = [self.token_ids, self.segment_ids]
# inputs = [np.array([i]) for i in inputs]
scores = output_probas
scores = self.output_scores.reshape((-1, 1)) + scores # 综合累积得分
indices = scores.argpartition(-self.num_beams, axis=None)[-self.num_beams:] # 仅保留topk
indices_1 = indices // scores.shape[1] # 行索引
indices_2 = (indices % scores.shape[1]).reshape((-1, 1)) # 列索引
self.output_ids = np.concatenate([self.output_ids[indices_1], indices_2],
1) # 更新输出
self.output_scores = np.take_along_axis(
scores, indices, axis=None
) # 更新得分
is_end = self.output_ids[:, -1] == self.end_id # 标记是否以end标记结束
self.end_counts = (self.output_ids == self.end_id).sum(1) # 统计出现的end标记
if self.output_ids.shape[1] >= self.minlen: # 最短长度判断
best = self.output_scores.argmax() # 得分最大的那个
if is_end[best] and self.end_counts[best] >= self.min_ends: # 如果已经终止
# return output_ids[best] # 直接输出
self.return_str_main(self.output_ids, best)
self.over = True
else: # 否则，只保留未完成部分
flag = ~is_end | (self.end_counts < self.min_ends) # 标记未完成序列
if not flag.all(): # 如果有已完成的
self.output_ids = self.output_ids[flag] # 扔掉已完成序列
self.output_scores = self.output_scores[flag] # 扔掉已完成序列
self.end_counts = self.end_counts[flag] # 扔掉已完成end计数
self.num_beams = flag.sum() # topk相应变化
self.output_str = [tokenizer.decode(ids) for ids in self.output_ids]
self.text = [self.text[0] for i in range(len(self.output_ids))]
# # 达到长度直接输出
# return output_ids[output_scores.argmax()]
# def data(self):
# token_ids, segment_ids = self.tokenizer.encode(self.text, maxlen=256)
# self.token_ids = token_ids
# self.segment_ids = segment_ids
# input_str = [text for i in range(self.num_beams)]
# output_str = self.output_str
# return input_str, output_str
def return_str_main(self, output_ids, best):
output_ids_best = output_ids[best]
self.return_str = self.tokenizer.decode(output_ids_best)
generatemodel = GenerateModel()
encoder,seq2seq, tokenizer = generatemodel.device_setup()
autotitle = AutoTitle(seq2seq, tokenizer, start_id=None, end_id=tokenizer._token_end_id, maxlen=120)
def just_show(file):
data = []
try:
with open(file, 'r', encoding="utf-8") as f:
lines = [x.strip() for x in f if x.strip() != '']
except:
with open(file, 'r', encoding="gbk") as f:
lines = [x.strip() for x in f if x.strip() != '']
# s2 = u'她只能应下来。'
# lines = pd.read_csv(file,encoding="gbk").values.tolist()
# random.shuffle(lines)
# lines = lines[:20]
for s in tqdm(lines[:2]):
print(s)
pre = autotitle.generate_random(s)
print(s)
print(pre)
# data.append([s, pre])
# pd.DataFrame(data,columns=["原始文本","生成文本"]).to_csv("data/text_测试一万字_unilm_修正数据_小说预训练_全部数据_epoch72_反向训练.csv")
def just_show_sentence(file: list) -> object:
"""
@param file:list
"""
text = file[0]
pre = autotitle.generate(text)
return pre
def just_show_sentence_batch(file: list) -> object:
text = file
pre = autotitle.generate_beam_search_batch(text)
return pre
def just_show_sentence_top(file: list) -> object:
text = file
pre = autotitle.generate_top(text)
return pre
def just_show_csv_random(file):
data_new = []
data = pd.read_csv(file).values.tolist()
for sentence in tqdm(data):
sentence = sentence[1]
print(sentence)
data_new_dan = []
data_new_dan.extend([sentence, len(sentence)])
pre = autotitle.generate_random(sentence)
for i in pre:
data_new_dan.extend([i, len(i)])
data_new.append(data_new_dan)
pd.DataFrame(data_new).to_csv("data/###第3章 非常尴尬_sim_topK_5.csv")
# return pre
def just_show_chachong_random(file):
text = file[0]
pre = autotitle.gen_synonyms(text)
return pre
def just_show_csv_beam(file):
data_new = []
data = pd.read_csv(file).values.tolist()
for sentence in tqdm(data):
sentence = sentence[1]
print(sentence)
data_new_dan = []
data_new_dan.extend([sentence, len(sentence)])
pre = autotitle.generate([sentence])
print(pre)
data_new_dan.extend([pre, len(pre)])
data_new.append(data_new_dan)
pd.DataFrame(data_new).to_csv("data/###第3章 非常尴尬_sim_topK_1.csv")
def chulichangju_1(text, chulipangban_return_list):
fuhao = [",","，","？","！","…"]
text_1 = text[:60]
text_2 = text[60:]
text_1_new = ""
for i in range(len(text_1)-1, -1, -1):
if text_1[i] in fuhao:
text_1_new = text_1[:i]
text_1_new += text_1[i]
if len(text_1_new) > 10:
text_1_new_pre = autotitle.generate(text_1_new)
else:
text_1_new_pre = text_1_new
chulipangban_return_list.append(text_1_new_pre)
if text_2 != "":
if i+1 != 60:
text_2 = text_1[i+1:] + text_2
break
# else:
# chulipangban_return_list.append(text_1)
if text_1_new == "":
if len(text_1) > 10:
text_1_new_pre = autotitle.gen_synonyms_short(text_1)
else:
text_1_new_pre = text_1
chulipangban_return_list.append(text_1_new_pre)
if text_2 != "":
chulipangban_return_list = chulichangju_1(text_2, chulipangban_return_list)
return chulipangban_return_list
def chulipangban_test_1(text):
sentence_list = text.split("。")
sentence_list_new = []
for i in sentence_list:
if i != "":
sentence_list_new.append(i)
sentence_list = sentence_list_new
return_list = []
for sentence in sentence_list:
if len(sentence) < 60:
if len(sentence) > 10:
sentence_pre = autotitle.generate(sentence)
else:
sentence_pre = sentence
return_list.append(sentence_pre)
else:
sentence_split_list = chulichangju_1(sentence,[])
sentence_split_text = "".join(sentence_split_list)
return_list.append(sentence_split_text)
return return_list
def paragraph_test(text, text_new):
text = chulipangban_test_1(text)
text = "。".join(text)
text_new.append(text)
# text_new_str = "".join(text_new)
return text_new
if __name__ == '__main__':
# text = ["历史和当下都证明,创新是民族生存、“发展的不竭源泉”,是是自身发展的必然选择",
# "自身发展的必然选择",
# "强调轻资产经营, 更加重视经营风险的规避",
# "历史和当下都证明,创新是民族生存、发展的不竭源泉,是是自身发展的必然选择",
# "是时代对于青年们的深切呼唤"]
text = ["基本消除“热桥”影响。"]
print(just_show_sentence(text))
# print(just_show_sentence_top(text))
# print(just_show_chachong_random(text))
# print(tokenizer.encode("\"", maxlen=120))
# print(just_show_sentence_batch(text))
# path = "./data/700条论文测试.xlsx"
# df_list = pd.read_excel(path).values.tolist()
#
# df_list_new = []
# print(len(df_list))
# for i in tqdm(df_list):
# try:
# pre = just_show_chachong_random([i[0]])
# df_list_new.append([i[0], i[1]] + pre)
# except:
# print(i[0])
# continue
# df = pd.DataFrame(df_list_new)
# df.to_excel("./data/700条论文测试_15.xlsx", index=None)