学AI还能赢奖品?每天30分钟,25天打通AI任督二脉 (qq.com)
基于MindSpore的GPT2文本摘要
%%capture captured_output
# 实验环境已经预装了mindspore==2.2.14,如需更换mindspore版本,可更改下面mindspore的版本号
!pip uninstall mindspore -y
!pip install -i https://pypi.mirrors.ustc.edu.cn/simple mindspore==2.2.14
!pip install tokenizers==0.15.0 -i https://pypi.tuna.tsinghua.edu.cn/simple
# 该案例在 mindnlp 0.3.1 版本完成适配,如果发现案例跑不通,可以指定mindnlp版本,执行`!pip install mindnlp==0.3.1`
!pip install mindnlp
数据集加载与处理
-
数据集加载
本次实验使用的是nlpcc2017摘要数据,内容为新闻正文及其摘要,总计50000个样本。
from mindnlp.utils import http_get
# download dataset
url = 'https://download.mindspore.cn/toolkits/mindnlp/dataset/text_generation/nlpcc2017/train_with_summ.txt'
path = http_get(url, './')
from mindspore.dataset import TextFileDataset
# load dataset
dataset = TextFileDataset(str(path), shuffle=False)
dataset.get_dataset_size()
# split into training and testing dataset
train_dataset, test_dataset = dataset.split([0.9, 0.1], randomize=False)
使用http_get
函数从指定URL下载数据集。
使用TextFileDataset
加载数据集,并将其分割为训练集和测试集。
-
数据预处理
原始数据格式:
article: [CLS] article_context [SEP] summary: [CLS] summary_context [SEP]
预处理后的数据格式:
[CLS] article_context [SEP] summary_context [SEP]
import json
import numpy as np
# preprocess dataset
def process_dataset(dataset, tokenizer, batch_size=6, max_seq_len=1024, shuffle=False):
def read_map(text):
data = json.loads(text.tobytes())
return np.array(data['article']), np.array(data['summarization'])
def merge_and_pad(article, summary):
# tokenization
# pad to max_seq_length, only truncate the article
tokenized = tokenizer(text=article, text_pair=summary,
padding='max_length', truncation='only_first', max_length=max_seq_len)
return tokenized['input_ids'], tokenized['input_ids']
dataset = dataset.map(read_map, 'text', ['article', 'summary'])
# change column names to input_ids and labels for the following training
dataset = dataset.map(merge_and_pad, ['article', 'summary'], ['input_ids', 'labels'])
dataset = dataset.batch(batch_size)
if shuffle:
dataset = dataset.shuffle(batch_size)
return dataset
因GPT2无中文的tokenizer,我们使用BertTokenizer替代。
from mindnlp.transformers import BertTokenizer
# We use BertTokenizer for tokenizing chinese context.
tokenizer = BertTokenizer.from_pretrained('bert-base-chinese')
len(tokenizer)
train_dataset = process_dataset(train_dataset, tokenizer, batch_size=4)
next(train_dataset.create_tuple_iterator())
定义process_dataset
函数来预处理数据集,包括读取数据、进行分词和填充等操作。
采用BertTokenizer进行中文文本的预处理。
模型构建
- 构建GPT2ForSummarization模型,注意shift right的操作。
from mindspore import ops
from mindnlp.transformers import GPT2LMHeadModel
class GPT2ForSummarization(GPT2LMHeadModel):
def construct(
self,
input_ids = None,
attention_mask = None,
labels = None,
):
outputs = super().construct(input_ids=input_ids, attention_mask=attention_mask)
shift_logits = outputs.logits[..., :-1, :]
shift_labels = labels[..., 1:]
# Flatten the tokens
loss = ops.cross_entropy(shift_logits.view(-1, shift_logits.shape[-1]), shift_labels.view(-1), ignore_index=tokenizer.pad_token_id)
return loss
创建GPT2ForSummarization
类,继承自GPT2LMHeadModel
,用于文本摘要任务。
在construct
方法中,实现标签的右移操作,匹配序列到序列的需求。
- 动态学习率
from mindspore import ops
from mindspore.nn.learning_rate_schedule import LearningRateSchedule
class LinearWithWarmUp(LearningRateSchedule):
"""
Warmup-decay learning rate.
"""
def __init__(self, learning_rate, num_warmup_steps, num_training_steps):
super().__init__()
self.learning_rate = learning_rate
self.num_warmup_steps = num_warmup_steps
self.num_training_steps = num_training_steps
def construct(self, global_step):
if global_step < self.num_warmup_steps:
return global_step / float(max(1, self.num_warmup_steps)) * self.learning_rate
return ops.maximum(
0.0, (self.num_training_steps - global_step) / (max(1, self.num_training_steps - self.num_warmup_steps))
) * self.learning_rate
定义LinearWithWarmUp
类,实现线性预热衰减学习率策略,在训练初期逐步增加学习率以帮助模型快速收敛,随后在训练后期逐渐降低学习率以进行精细调整。
模型训练
num_epochs = 1
warmup_steps = 2000
learning_rate = 1.5e-4
num_training_steps = num_epochs * train_dataset.get_dataset_size()
from mindspore import nn
from mindnlp.transformers import GPT2Config, GPT2LMHeadModel
config = GPT2Config(vocab_size=len(tokenizer))
model = GPT2ForSummarization(config)
lr_scheduler = LinearWithWarmUp(learning_rate=learning_rate, num_warmup_steps=warmup_steps, num_training_steps=num_training_steps)
optimizer = nn.AdamWeightDecay(model.trainable_params(), learning_rate=lr_scheduler)
# 记录模型参数数量
print('number of model parameters: {}'.format(model.num_parameters()))
from mindnlp._legacy.engine import Trainer
from mindnlp._legacy.engine.callbacks import CheckpointCallback
ckpoint_cb = CheckpointCallback(save_path='checkpoint', ckpt_name='gpt2_summarization',
epochs=1, keep_checkpoint_max=2)
trainer = Trainer(network=model, train_dataset=train_dataset,
epochs=1, optimizer=optimizer, callbacks=ckpoint_cb)
trainer.set_amp(level='O1') # 开启混合精度
注:建议使用较高规格的算力,训练时间较长
trainer.run(tgt_columns="labels")
设置训练参数,如学习率、预热步数和总训练步数。
使用Trainer
进行模型训练,设置检查点回调保存模型。
原数据集50000样本,训练时间较长,实际用了10000样本减少训练时间。
没有尝试静态图mindnlp/mindnlp/transformers/models/gpt2 · mindnlp · GitHub
模型推理
数据处理,将向量数据变为中文数据
def process_test_dataset(dataset, tokenizer, batch_size=1, max_seq_len=1024, max_summary_len=100):
def read_map(text):
data = json.loads(text.tobytes())
return np.array(data['article']), np.array(data['summarization'])
def pad(article):
tokenized = tokenizer(text=article, truncation=True, max_length=max_seq_len-max_summary_len)
return tokenized['input_ids']
dataset = dataset.map(read_map, 'text', ['article', 'summary'])
dataset = dataset.map(pad, 'article', ['input_ids'])
dataset = dataset.batch(batch_size)
return dataset
test_dataset = process_test_dataset(test_dataset, tokenizer, batch_size=1)
print(next(test_dataset.create_tuple_iterator(output_numpy=True)))
model = GPT2LMHeadModel.from_pretrained('./checkpoint/gpt2_summarization_epoch_0.ckpt', config=config)
model.set_train(False)
model.config.eos_token_id = model.config.sep_token_id
i = 0
for (input_ids, raw_summary) in test_dataset.create_tuple_iterator():
output_ids = model.generate(input_ids, max_new_tokens=50, num_beams=5, no_repeat_ngram_size=2)
output_text = tokenizer.decode(output_ids[0].tolist())
print(output_text)
i += 1
if i == 1:
break
定义process_test_dataset
函数来处理测试数据集。
加载训练好的模型,使用generate
方法生成摘要。
代码基于MindSpore的GPT2文本摘要模型。首先,导入了所需的库和模块,下载并加载数据集。接着,对数据集进行预处理,包括分词、填充等操作。构建GPT2ForSummarization模型,模型继承自GPT2LMHeadModel,重写construct方法。模型训练部分,设置学习率调度器、优化器和检查点回调,使用Trainer进行训练。最后,对测试数据集进行处理,使用训练好的模型进行推理,输出摘要结果。