You can not select more than 25 topics
Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
904 lines
41 KiB
904 lines
41 KiB
#! -*- coding: utf-8 -*-
|
|
|
|
import os
|
|
os.environ["TF_KERAS"] = "1"
|
|
os.environ["CUDA_VISIBLE_DEVICES"] = "1"
|
|
import glob
|
|
from numpy import random
|
|
random.seed(1001)
|
|
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 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
|
|
|
|
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
|
|
)
|
|
|
|
# output = CrossEntropy(2)(model.inputs + model.outputs)
|
|
#
|
|
# model = Model(model.inputs, output)
|
|
model = Model(model.inputs, model.outputs)
|
|
path_model = './output_quan/best_model_20wan_1.weights'
|
|
model.load_weights(path_model)
|
|
|
|
return model, 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):
|
|
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)
|
|
return batch_token_ids, batch_segment_ids
|
|
|
|
|
|
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]
|
|
probas_b = probas[0, :]
|
|
probas_r = probas[:-beam_n, :]
|
|
output_ids_r = output_ids[:-beam_n, :]
|
|
output_ids_b = output_ids[-beam_n:, :]
|
|
else:
|
|
probas_b = probas[-beam_n:, :]
|
|
if whether_end_b == False:
|
|
inputs_r = [i[:-beam_n, :] for i in inputs]
|
|
inputs_b = [i[-beam_n:, :] for i in inputs]
|
|
probas_r = probas[:-beam_n, :]
|
|
output_ids_r = output_ids[:-beam_n, :]
|
|
output_ids_b = output_ids[-beam_n:, :]
|
|
else:
|
|
inputs_r = inputs
|
|
probas_r = probas
|
|
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
|
|
output_ids_b = np.concatenate([output_ids_b[indices_1], indices_2],
|
|
1) # 更新输出
|
|
|
|
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
|
|
|
|
sample_ids = np.apply_along_axis(sample_func, 1, probas_r) # 执行采样
|
|
|
|
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
|
|
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:
|
|
is_end_r = output_ids_r[:, -1] == self.end_id
|
|
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
|
|
index_r = index_r[flag]
|
|
|
|
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 random_sample_and_beam_search_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。
|
|
"""
|
|
whether_end_b = False
|
|
results_r = []
|
|
results_b = []
|
|
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]
|
|
probas_b = probas[0, :]
|
|
probas_r = probas[:-beam_n, :]
|
|
output_ids_r = output_ids[:-beam_n, :]
|
|
output_ids_b = output_ids[-beam_n:, :]
|
|
else:
|
|
probas_b = probas[-beam_n:, :]
|
|
if whether_end_b == False:
|
|
inputs_r = [i[:-beam_n, :] for i in inputs]
|
|
inputs_b = [i[-beam_n:, :] for i in inputs]
|
|
probas_r = probas[:-beam_n, :]
|
|
output_ids_r = output_ids[:-beam_n, :]
|
|
output_ids_b = output_ids[-beam_n:, :]
|
|
else:
|
|
inputs_r = inputs
|
|
probas_r = probas
|
|
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
|
|
output_ids_b = np.concatenate([output_ids_b[indices_1], indices_2],
|
|
1) # 更新输出
|
|
|
|
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
|
|
|
|
sample_ids = np.apply_along_axis(sample_func, 1, probas_r) # 执行采样
|
|
|
|
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
|
|
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:
|
|
is_end_r = output_ids_r[:, -1] == self.end_id
|
|
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
|
|
index_r = index_r[flag]
|
|
|
|
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 random_sample_seed(
|
|
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 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
|
|
random.seed(1001)
|
|
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 beam_search_batch(
|
|
# self,
|
|
# inputs_str,
|
|
# topk = 1
|
|
# 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 = self.first_output_ids_batch = 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)
|
|
# 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
|
|
|
|
def random_sample_topp_gentle(
|
|
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 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
|
|
# TODO
|
|
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
|
|
# me = np.mean(n)
|
|
# c = n + (2 * me)
|
|
# c /= c.sum()
|
|
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 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 = self.first_output_ids_batch = 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)
|
|
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):
|
|
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=5):
|
|
text = text[0]
|
|
token_ids, segment_ids = self.tokenizer.encode(text, maxlen=120)
|
|
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, topp=0.9):
|
|
token_ids, segment_ids = self.tokenizer.encode(text, maxlen=120)
|
|
# batch_token_ids = sequence_padding(batch_token_ids)
|
|
# batch_segment_ids = sequence_padding(batch_segment_ids)
|
|
# token_ids, segment_ids = self.data_generator(text)
|
|
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_random_sample_topp_gentle(self, text, n=20, topp=0.9):
|
|
token_ids, segment_ids = self.tokenizer.encode(text, maxlen=120)
|
|
# batch_token_ids = sequence_padding(batch_token_ids)
|
|
# batch_segment_ids = sequence_padding(batch_segment_ids)
|
|
# token_ids, segment_ids = self.data_generator(text)
|
|
output_ids = self.random_sample_topp_gentle([token_ids, segment_ids], n, topp=topp) # 基于随机采样
|
|
return [tokenizer.decode(ids) for ids in output_ids]
|
|
|
|
def generate_random_shortest(self, text, n=20, topk=5):
|
|
token_ids, segment_ids = self.tokenizer.encode(text, maxlen=120)
|
|
# batch_token_ids = sequence_padding(batch_token_ids)
|
|
# batch_segment_ids = sequence_padding(batch_segment_ids)
|
|
# token_ids, segment_ids = self.data_generator(text)
|
|
output_ids = self.random_sample_seed([token_ids, segment_ids], n, topk) # 基于随机采样
|
|
return_str = [tokenizer.decode(ids) for ids in output_ids][0]
|
|
return return_str
|
|
|
|
def generate_top(self, text):
|
|
output_ids = self.batch(text) # 基于随机采样
|
|
return [tokenizer.decode(ids) for ids in output_ids]
|
|
|
|
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
|
|
|
|
|
|
generatemodel = GenerateModel()
|
|
model, tokenizer = generatemodel.device_setup()
|
|
autotitle = AutoTitle(model, tokenizer, start_id=None, end_id=tokenizer._token_end_id, maxlen=60)
|
|
|
|
|
|
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):
|
|
"""
|
|
@param file:list
|
|
"""
|
|
# file = file[0]
|
|
# for i in range(100):
|
|
# pre = autotitle.generate_random_sample_topp_gentle(file)
|
|
# print(pre)
|
|
|
|
|
|
pre = autotitle.generate(file)
|
|
print(pre)
|
|
# print(pre)
|
|
# if isinstance(pre,list):
|
|
# for i in pre:
|
|
# print(i, len(i))
|
|
#
|
|
#
|
|
# if isinstance(pre,str):
|
|
# print(pre)
|
|
|
|
def just_show_csv(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_generate_random.csv")
|
|
# return pre
|
|
|
|
if __name__ == '__main__':
|
|
# file = "train_2842.txt"
|
|
# just_show(file)
|
|
text = ["迈向新时代,当代青年要立鸿鹄之志,做马克思主义的坚定信仰者。"]
|
|
just_show_sentence(text)
|
|
# "简言之，她不好过，李四也别想好过！"
|
|
# s = "张三的对话"
|
|
# print(autotitle.generate(s))
|
|
# file = "data/###第3章 非常尴尬.csv"
|
|
# just_show_csv(file)
|