首页 > 其他分享 >Transformers实战——文本相似度

Transformers实战——文本相似度

时间:2023-12-15 15:07:38浏览次数:36  
标签:实战 None False train eval Transformers model 文本 True



文章目录

  • 一、改写文本分类
  • 1.导入相关包
  • 2.加载数据集
  • 3.划分数据集
  • 4.数据集预处理
  • 5.创建模型
  • 6.创建评估函数
  • 7.创建 TrainingArguments
  • 8.创建 Trainer
  • 9.模型训练
  • 10.模型评估
  • 11.模型预测
  • 二、交互/单塔模式
  • 1.导入相关包
  • 2.加载数据集
  • 3.划分数据集
  • 4.数据集预处理
  • 5.创建模型(区别)
  • 6.创建评估函数(区别)
  • 7.创建 TrainingArguments
  • 8.创建 Trainer
  • 9.模型训练
  • 10.模型评估
  • 11.模型预测(区别)


!pip install transformers datasets evaluate accelerate

一、改写文本分类

1.导入相关包

from transformers import AutoTokenizer, AutoModelForSequenceClassification, Trainer, TrainingArguments
from datasets import load_dataset

2.加载数据集

dataset = load_dataset("json", data_files="./train_pair_1w.json", split="train")
dataset
'''
Dataset({
    features: ['sentence1', 'sentence2', 'label'],
    num_rows: 10000
})
'''
dataset[:3]
'''
{'sentence1': ['找一部小时候的动画片',
  '我不可能是一个有鉴赏能力的行家,小姐我把我的时间都花在书写上;象这样豪华的舞会,我还是头一次见到。',
  '胡子长得太快怎么办?'],
 'sentence2': ['求一部小时候的动画片。谢了', '蜡烛没熄就好了,夜黑得瘆人,情绪压抑。', '胡子长得快怎么办?'],
 'label': ['1', '0', '1']}
'''

3.划分数据集

datasets = dataset.train_test_split(test_size=0.2, seed=3407)
datasets['train']['sentence1'][0]
'''
王琦瑶说:你家阿大二十岁已经有儿有女了嘛
'''

4.数据集预处理

import torch

tokenizer = AutoTokenizer.from_pretrained("hfl/chinese-macbert-base")

def process_function(examples):
    tokenized_examples = tokenizer(examples["sentence1"], examples["sentence2"], max_length=128, truncation=True)
    tokenized_examples["labels"] = [int(label) for label in examples["label"]]
    return tokenized_examples

tokenized_datasets = datasets.map(process_function, batched=True, remove_columns=datasets["train"].column_names)
tokenized_datasets
'''
DatasetDict({
    train: Dataset({
        features: ['input_ids', 'token_type_ids', 'attention_mask', 'labels'],
        num_rows: 8000
    })
    test: Dataset({
        features: ['input_ids', 'token_type_ids', 'attention_mask', 'labels'],
        num_rows: 2000
    })
})
'''

5.创建模型

from transformers import BertForSequenceClassification 
model = AutoModelForSequenceClassification.from_pretrained("hfl/chinese-macbert-base", num_labels=2)

6.创建评估函数

import evaluate

acc_metric = evaluate.load("accuracy")
f1_metirc = evaluate.load("f1")
def eval_metric(eval_predict):
    predictions, labels = eval_predict
    labels = [int(l) for l in labels]
    predictions = predictions.argmax(axis=-1)
    acc = acc_metric.compute(predictions=predictions, references=labels)
    f1 = f1_metirc.compute(predictions=predictions, references=labels)
    acc.update(f1)
    return acc

7.创建 TrainingArguments

train_args = TrainingArguments(output_dir="./cross_model",      # 输出文件夹
                               per_device_train_batch_size=32,  # 训练时的batch_size
                               per_device_eval_batch_size=32,  # 验证时的batch_size
                               logging_steps=10,                # log 打印的频率
                               evaluation_strategy="epoch",     # 评估策略
                               save_strategy="epoch",           # 保存策略
                               save_total_limit=3,              # 最大保存数
                               learning_rate=2e-5,              # 学习率
                               weight_decay=0.01,               # weight_decay
                               metric_for_best_model="f1",      # 设定评估指标
                               load_best_model_at_end=True)     # 训练完成后加载最优模型
train_args
'''
TrainingArguments(
_n_gpu=1,
adafactor=False,
adam_beta1=0.9,
adam_beta2=0.999,
adam_epsilon=1e-08,
auto_find_batch_size=False,
bf16=False,
bf16_full_eval=False,
data_seed=None,
dataloader_drop_last=False,
dataloader_num_workers=0,
dataloader_pin_memory=True,
ddp_backend=None,
ddp_broadcast_buffers=None,
ddp_bucket_cap_mb=None,
ddp_find_unused_parameters=None,
ddp_timeout=1800,
debug=[],
deepspeed=None,
disable_tqdm=False,
dispatch_batches=None,
do_eval=True,
do_predict=False,
do_train=False,
eval_accumulation_steps=None,
eval_delay=0,
eval_steps=None,
evaluation_strategy=epoch,
fp16=False,
fp16_backend=auto,
fp16_full_eval=False,
fp16_opt_level=O1,
fsdp=[],
fsdp_config={'min_num_params': 0, 'xla': False, 'xla_fsdp_grad_ckpt': False},
fsdp_min_num_params=0,
fsdp_transformer_layer_cls_to_wrap=None,
full_determinism=False,
gradient_accumulation_steps=1,
gradient_checkpointing=False,
gradient_checkpointing_kwargs=None,
greater_is_better=True,
group_by_length=False,
half_precision_backend=auto,
hub_always_push=False,
hub_model_id=None,
hub_private_repo=False,
hub_strategy=every_save,
hub_token=<HUB_TOKEN>,
ignore_data_skip=False,
include_inputs_for_metrics=False,
include_tokens_per_second=False,
jit_mode_eval=False,
label_names=None,
label_smoothing_factor=0.0,
learning_rate=2e-05,
length_column_name=length,
load_best_model_at_end=True,
local_rank=0,
log_level=passive,
log_level_replica=warning,
log_on_each_node=True,
logging_dir=./cross_model/runs/Nov27_07-11-23_66feef283143,
logging_first_step=False,
logging_nan_inf_filter=True,
logging_steps=10,
logging_strategy=steps,
lr_scheduler_type=linear,
max_grad_norm=1.0,
max_steps=-1,
metric_for_best_model=f1,
mp_parameters=,
neftune_noise_alpha=None,
no_cuda=False,
num_train_epochs=3.0,
optim=adamw_torch,
optim_args=None,
output_dir=./cross_model,
overwrite_output_dir=False,
past_index=-1,
per_device_eval_batch_size=32,
per_device_train_batch_size=32,
prediction_loss_only=False,
push_to_hub=False,
push_to_hub_model_id=None,
push_to_hub_organization=None,
push_to_hub_token=<PUSH_TO_HUB_TOKEN>,
ray_scope=last,
remove_unused_columns=True,
report_to=['tensorboard'],
resume_from_checkpoint=None,
run_name=./cross_model,
save_on_each_node=False,
save_safetensors=True,
save_steps=500,
save_strategy=epoch,
save_total_limit=3,
seed=42,
skip_memory_metrics=True,
split_batches=False,
tf32=None,
torch_compile=False,
torch_compile_backend=None,
torch_compile_mode=None,
torchdynamo=None,
tpu_metrics_debug=False,
tpu_num_cores=None,
use_cpu=False,
use_ipex=False,
use_legacy_prediction_loop=False,
use_mps_device=False,
warmup_ratio=0.0,
warmup_steps=0,
weight_decay=0.01,
)
'''

8.创建 Trainer

from transformers import DataCollatorWithPadding
trainer = Trainer(model=model, 
                  args=train_args, 
                  train_dataset=tokenized_datasets["train"], 
                  eval_dataset=tokenized_datasets["test"], 
                  data_collator=DataCollatorWithPadding(tokenizer=tokenizer),
                  compute_metrics=eval_metric)

9.模型训练

trainer.train()

10.模型评估

trainer.evaluate(tokenized_datasets["test"])

11.模型预测

from transformers import pipeline, TextClassificationPipeline

model.config.id2label = {0: "不相似", 1: "相似"}
pipe = pipeline("text-classification", model=model, tokenizer=tokenizer, device=0)

result = pipe({"text": "我喜欢北京", "text_pair": "天气"})
result["label"] = "相似" if result["score"] > 0.5 else "不相似"
result
'''
{'label': '不相似', 'score': 0.8792306780815125}
'''

result = pipe({"text": "我喜欢北京", "text_pair": "我喜欢北京"})
result
'''
{'label': '相似', 'score': 0.9374899864196777}
'''

二、交互/单塔模式

  • label 设为 1,代表两个句子的相似度分数,通过设置阈值来判断类别
  • 对于同一个句子 A,存在若干候选句子,要找到与句子 A 最相似的某个候选句子(上述文本分类处理方式无法解决),此处将分类任务作为回归任务+阈值的方式进行处理,从而能够得到预测的得分,该得分可以用来判断 哪个候选句子与给定句子最相似

1.导入相关包

from transformers import AutoTokenizer, AutoModelForSequenceClassification, Trainer, TrainingArguments
from datasets import load_dataset

2.加载数据集

dataset = load_dataset("json", data_files="./train_pair_1w.json", split="train")
dataset
'''
Dataset({
    features: ['sentence1', 'sentence2', 'label'],
    num_rows: 10000
})
'''
dataset[:3]
'''
{'sentence1': ['找一部小时候的动画片',
  '我不可能是一个有鉴赏能力的行家,小姐我把我的时间都花在书写上;象这样豪华的舞会,我还是头一次见到。',
  '胡子长得太快怎么办?'],
 'sentence2': ['求一部小时候的动画片。谢了', '蜡烛没熄就好了,夜黑得瘆人,情绪压抑。', '胡子长得快怎么办?'],
 'label': ['1', '0', '1']}
'''

3.划分数据集

  • 注意这里有种子参数
datasets = dataset.train_test_split(test_size=0.2, seed=3407)
datasets['train']['sentence1'][0]
'''
王琦瑶说:你家阿大二十岁已经有儿有女了嘛
'''

4.数据集预处理

import torch

tokenizer = AutoTokenizer.from_pretrained("hfl/chinese-macbert-base")

def process_function(examples):
    tokenized_examples = tokenizer(examples["sentence1"], examples["sentence2"], max_length=128, truncation=True)
    
    # 这里float(label)的原因是要做MSE,需要float类型的数据
    tokenized_examples["labels"] = [float(label) for label in examples["label"]]
    
    return tokenized_examples

tokenized_datasets = datasets.map(process_function, batched=True, remove_columns=datasets["train"].column_names)

tokenized_datasets
'''
DatasetDict({
    train: Dataset({
        features: ['input_ids', 'token_type_ids', 'attention_mask', 'labels'],
        num_rows: 8000
    })
    test: Dataset({
        features: ['input_ids', 'token_type_ids', 'attention_mask', 'labels'],
        num_rows: 2000
    })
}
'''
print(tokenized_datasets["train"][0])
'''
{'input_ids': [101, 4374, 4425, 4457, 6432, 131, 872, 2157, 7350, 1920, 753, 1282, 2259, 2347, 5307, 3300, 1036, 3300, 1957, 749, 1658, 102, 7350, 1920, 1372, 3300, 1036, 2094, 3766, 3300, 1957, 1036, 102], 
'token_type_ids': [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], 
'attention_mask': [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], 
'labels': 0.0}
'''

5.创建模型(区别)

from transformers import BertForSequenceClassification 

model = AutoModelForSequenceClassification.from_pretrained("hfl/chinese-macbert-base", num_labels=1)

6.创建评估函数(区别)

import evaluate

acc_metric = evaluate.load("accuracy")
f1_metirc = evaluate.load("f1")
def eval_metric(eval_predict):
    predictions, labels = eval_predict
    predictions = [int(p > 0.5) for p in predictions]
    labels = [int(l) for l in labels]
    # predictions = predictions.argmax(axis=-1)
    acc = acc_metric.compute(predictions=predictions, references=labels)
    f1 = f1_metirc.compute(predictions=predictions, references=labels)
    acc.update(f1)
    return acc

7.创建 TrainingArguments

train_args = TrainingArguments(output_dir="./cross_model",      # 输出文件夹
                               per_device_train_batch_size=32,  # 训练时的batch_size
                               per_device_eval_batch_size=32,   # 验证时的batch_size
                               logging_steps=10,                # log 打印的频率
                               evaluation_strategy="epoch",     # 评估策略
                               save_strategy="epoch",           # 保存策略
                               save_total_limit=3,              # 最大保存数
                               learning_rate=2e-5,              # 学习率
                               weight_decay=0.01,               # weight_decay
                               metric_for_best_model="f1",      # 设定评估指标
                               load_best_model_at_end=True)     # 训练完成后加载最优模型

train_args
'''
TrainingArguments(
_n_gpu=1,
adafactor=False,
adam_beta1=0.9,
adam_beta2=0.999,
adam_epsilon=1e-08,
auto_find_batch_size=False,
bf16=False,
bf16_full_eval=False,
data_seed=None,
dataloader_drop_last=False,
dataloader_num_workers=0,
dataloader_pin_memory=True,
ddp_backend=None,
ddp_broadcast_buffers=None,
ddp_bucket_cap_mb=None,
ddp_find_unused_parameters=None,
ddp_timeout=1800,
debug=[],
deepspeed=None,
disable_tqdm=False,
dispatch_batches=None,
do_eval=True,
do_predict=False,
do_train=False,
eval_accumulation_steps=None,
eval_delay=0,
eval_steps=None,
evaluation_strategy=epoch,
fp16=False,
fp16_backend=auto,
fp16_full_eval=False,
fp16_opt_level=O1,
fsdp=[],
fsdp_config={'min_num_params': 0, 'xla': False, 'xla_fsdp_grad_ckpt': False},
fsdp_min_num_params=0,
fsdp_transformer_layer_cls_to_wrap=None,
full_determinism=False,
gradient_accumulation_steps=1,
gradient_checkpointing=False,
gradient_checkpointing_kwargs=None,
greater_is_better=True,
group_by_length=False,
half_precision_backend=auto,
hub_always_push=False,
hub_model_id=None,
hub_private_repo=False,
hub_strategy=every_save,
hub_token=<HUB_TOKEN>,
ignore_data_skip=False,
include_inputs_for_metrics=False,
include_tokens_per_second=False,
jit_mode_eval=False,
label_names=None,
label_smoothing_factor=0.0,
learning_rate=2e-05,
length_column_name=length,
load_best_model_at_end=True,
local_rank=0,
log_level=passive,
log_level_replica=warning,
log_on_each_node=True,
logging_dir=./cross_model/runs/Nov27_06-35-36_66feef283143,
logging_first_step=False,
logging_nan_inf_filter=True,
logging_steps=10,
logging_strategy=steps,
lr_scheduler_type=linear,
max_grad_norm=1.0,
max_steps=-1,
metric_for_best_model=f1,
mp_parameters=,
neftune_noise_alpha=None,
no_cuda=False,
num_train_epochs=3.0,
optim=adamw_torch,
optim_args=None,
output_dir=./cross_model,
overwrite_output_dir=False,
past_index=-1,
per_device_eval_batch_size=32,
per_device_train_batch_size=32,
prediction_loss_only=False,
push_to_hub=False,
push_to_hub_model_id=None,
push_to_hub_organization=None,
push_to_hub_token=<PUSH_TO_HUB_TOKEN>,
ray_scope=last,
remove_unused_columns=True,
report_to=['tensorboard'],
resume_from_checkpoint=None,
run_name=./cross_model,
save_on_each_node=False,
save_safetensors=True,
save_steps=500,
save_strategy=epoch,
save_total_limit=3,
seed=42,
skip_memory_metrics=True,
split_batches=False,
tf32=None,
torch_compile=False,
torch_compile_backend=None,
torch_compile_mode=None,
torchdynamo=None,
tpu_metrics_debug=False,
tpu_num_cores=None,
use_cpu=False,
use_ipex=False,
use_legacy_prediction_loop=False,
use_mps_device=False,
warmup_ratio=0.0,
warmup_steps=0,
weight_decay=0.01,
)
'''

8.创建 Trainer

from transformers import DataCollatorWithPadding
trainer = Trainer(model=model, 
                  args=train_args, 
                  train_dataset=tokenized_datasets["train"], 
                  eval_dataset=tokenized_datasets["test"], 
                  data_collator=DataCollatorWithPadding(tokenizer=tokenizer),
                  compute_metrics=eval_metric)

9.模型训练

trainer.train()
'''
TrainOutput(global_step=750, training_loss=0.09012470634778341, metrics={'train_runtime': 558.2367, 'train_samples_per_second': 42.993, 'train_steps_per_second': 1.344, 'total_flos': 1552456398705984.0, 'train_loss': 0.09012470634778341, 'epoch': 3.0})
'''

Transformers实战——文本相似度_深度学习


10.模型评估

trainer.evaluate(tokenized_datasets["test"])
'''
{'eval_loss': 0.06814368069171906,
 'eval_accuracy': 0.9095,
 'eval_f1': 0.8840486867392696,
 'eval_runtime': 14.6336,
 'eval_samples_per_second': 136.672,
 'eval_steps_per_second': 4.305,
 'epoch': 3.0}
'''

11.模型预测(区别)

from transformers import pipeline, TextClassificationPipeline

model.config.id2label = {0: "不相似", 1: "相似"}

pipe = pipeline("text-classification", model=model, tokenizer=tokenizer, device=0)

# function_to_apply="none" 应对softmax / sigmoid处理
result = pipe({"text": "我喜欢北京", "text_pair": "天气怎样"}, function_to_apply="none")
result["label"] = "相似" if result["score"] > 0.5 else "不相似"

result
'''
{'label': '不相似', 'score': -0.025799434632062912}
'''

Transformers实战——文本相似度_人工智能_02



标签:实战,None,False,train,eval,Transformers,model,文本,True
From: https://blog.51cto.com/u_14608932/8840803

相关文章

  • shell实战正则三贱客——awk
    shell实战正则三贱客——awk特点及应用场景awk一门语言类似于c语言过滤,统计,计算过滤,统计,日志执行过程awk-F,'BEGIN{print"name"}{print$2}END{print"endoffile"}'oldboy.txt执行过程说明awk读取文件之前命令复制或命令行参数......
  • shell实战正则三贱客—grep
    三剑客三剑客特点,及应用场景命令特点场景grep过滤grep过滤速度是最快的sed替换,修改文件内容,取行如果要进行替换/修改文件内容,取出某个范围的内容(从早上10:00到11:00)awk取列,统计结算取列对比,比较统计,计算(awk数组)shell实战正则三贱客—grep选项含......
  • shell实战正则三贱客——sed
    shell实战正则三贱客——sed特点及格式sedstreameditor:流编辑器,sed把处理的内容(文件),当做是水,源源不断的进行处理,一行一行读取文件,直到文件末尾。sed格式命令选项sed命令功能说明参数sed-r's#oldboy#oldgirl#g'单引号里面的s表示替换功能;g表示一......
  • shell实战正则表达式
    正则表达式匹配有规律的东西:手机号,身份证号,匹配日志什么正则表达式,regularexpression(RE)正则:使用一些符号表达重复出现,大小写,开头/结尾含义。应用场景哪些可以正则表达式Linux三剑客使用,一些开发语言应用场景过滤有规律的内容,尤其是日志正则符号......
  • 人工智能中的文本分类:技术突破与实战指导
    在本文中,我们全面探讨了文本分类技术的发展历程、基本原理、关键技术、深度学习的应用,以及从RNN到Transformer的技术演进。文章详细介绍了各种模型的原理和实战应用,旨在提供对文本分类技术深入理解的全面视角。关注TechLead,分享AI全维度知识。作者拥有10+年互联网服务架构、AI......
  • Chrome扩展开发实战:快速填充表单
    大家好,我是dom哥。我正在写关于Chrome扩展开发的系列文章,感兴趣的可以点个小星星。填表单是打工人经常面对的场景,作为一个前端,我经常开发一些PC端的页面,它们主要由表单和表格构成,而输入框又是表单里最常见的表单项。接下来就试着做一个简单的小扩展,用于快速给表单里的输入......
  • 八、文本处理工具
    文本提取工具(查看文本)文本分析工具文本操作工具1、文本提取工具cat、more、less查看文本内容cat:打印一个或多个文件到标准输出1#合并文件2[root@exampletmp]#catfile1.txtfile2.txt>file3.txt34#查看文件行号5[root@exampletmp]#cat-n/tmp......
  • LLMs在纯文本数据与图数据结合的应用优势
     目前的在线客户服务平台已经实现了通过大型语言模型(LLMs)自动生成客户服务响应的能力。这些平台通过深入理解和分析大量的用户查询和历史服务记录,能够提供准确且个性化的回复,极大地提高了用户满意度和操作效率。在这一场景中,LLMs的应用展现了其理解自然语言复杂性的能力,能够根据......
  • HarmonyOS 实战小项目开发(二)
    HarmonyOS实战小项目开发(二)日常逼逼叨在上期实战项目一中,已经对于练手项目的背景,后端搭建等做了一定的简述,那么本期将结合HarmonyOS页面搭建个人性格测试的移动端。如有一些错误,希望观众老爷们批评指正......
  • 听懂未来:AI语音识别技术的进步与实战
    本文全面探索了语音识别技术,从其历史起源、关键技术发展到广泛的实际应用案例,揭示了这一领域的快速进步和深远影响。文章深入分析了语音识别在日常生活及各行业中的变革作用,展望了其未来发展趋势。关注TechLead,分享AI全维度知识。作者拥有10+年互联网服务架构、AI产品研发经验......