Text Classification

任务描述

文本分类：模型在基于文本对的微调后，可以在给定任意文本对与候选标签列表的情况下，完成对文本对关系的分类，文本对的两个文本之间以-分割。

相关论文 Jacob Devlin, Ming-Wei Chang, et al., BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding, 2019.

已支持数据集性能

model

type

datasets

Top1-accuracy

stage

example

bert

txtcls_bert_base_uncased

Mnli

30.9%

pretrain

–

txtcls

txtcls_bert_case_uncased_mnli

Mnli

84.8%

finetune
eval
predict

link
link
link

Mnli

  • 数据集大小：298M，共431992个样本，3个类别

    • 训练集：392702个样本

    • 匹配测试集：9796个样本

    • 非匹配测试集：9847个样本

    • 匹配开发集：9815个样本

    • 非匹配开发集：9832个样本

  • 数据格式：tsv文件

数据集目录格式
└─mnli
   ├─dev
   ├─test  
   └─train

  • 用户可以参考BERT代码仓中的run_classifier.py文件，进行Mnli数据集TFRecord格式文件的生成。

快速任务接口

脚本启动

需开发者提前clone工程。

# finetune
python run_mindformer.py --config ./configs/txtcls/run_txtcls_bert_base_uncased.yaml --run_mode finetune --load_checkpoint txtcls_bert_base_uncased

# evaluate
python run_mindformer.py --config ./configs/txtcls/run_txtcls_bert_base_uncased.yaml --run_mode eval --load_checkpoint txtcls_bert_base_uncased_mnli

# predict
python run_mindformer.py --config ./configs/txtcls/run_txtcls_bert_base_uncased.yaml --run_mode predict --load_checkpoint txtcls_bert_base_uncased_mnli --predict_data [TEXT]

调用API启动

  • Trainer接口开启训练/评估/推理：

import mindspore; mindspore.set_context(mode=0, device_id=0)
from mindformers import MindFormerBook
from mindformers.trainer import Trainer

# 显示Trainer的模型支持列表
MindFormerBook.show_trainer_support_model_list("text_classification")
# INFO - Trainer support model list for txt_classification task is:
# INFO -    ['txtcls_bert_base_uncased']
# INFO - -------------------------------------

# 初始化trainer
trainer = Trainer(task='text_classification',
    model='txtcls_bert_base_uncased',
    train_dataset='./mnli/train',
    eval_dataset='./mnli/eval')
# 测试数据，该input_data有两个测试案例，即两个文本对，单个文本对的两个文本之间用-分割
input_data = ["The new rights are nice enough-Everyone really likes the newest benefits ",
              "i don't know um do you do a lot of camping-I know exactly."]

#方式1：使用现有的预训练权重进行finetune， 并使用finetune获得的权重进行eval和推理
trainer.train(resume_or_finetune_from_checkpoint="txtcls_bert_base_uncased",
              do_finetune=True)
trainer.evaluate(eval_checkpoint=True)
trainer.predict(predict_checkpoint=True, input_data=input_data, top_k=1)

# 方式2： 从obs下载训练好的权重并进行eval和推理
trainer.evaluate()
# INFO - Top1 Accuracy=84.8%
trainer.predict(input_data=input_data, top_k=1)
# INFO - output result is [[{'label': 'neutral', 'score': 0.9714198708534241}],
#                         [{'label': 'contradiction', 'score': 0.9967639446258545}]]

  • pipeline接口开启快速推理

import mindspore; mindspore.set_context(mode=0, device_id=0)
from mindformers.pipeline import TextClassificationPipeline
from mindformers import AutoTokenizer, BertForMultipleChoice, AutoConfig

input_data = ["The new rights are nice enough-Everyone really likes the newest benefits ",
                "i don't know um do you do a lot of camping-I know exactly."]

tokenizer = AutoTokenizer.from_pretrained('txtcls_bert_base_uncased_mnli')
txtcls_mnli_config = AutoConfig.from_pretrained('txtcls_bert_base_uncased_mnli')

# Because batch_size parameter is required when bert model is created, and pipeline
# function deals with samples one by one, the batch_size parameter is seted one.
txtcls_mnli_config.batch_size = 1

model = BertForMultipleChoice(txtcls_mnli_config)
txtcls_pipeline = TextClassificationPipeline(task='text_classification',
                                             model=model,
                                             tokenizer=tokenizer,
                                             max_length=model.config.seq_length,
                                             padding="max_length")

results = txtcls_pipeline(input_data, top_k=1)
print(results)
# 输出
# [[{'label': 'neutral', 'score': 0.9714198708534241}], [{'label': 'contradiction', 'score': 0.9967639446258545}]]