🚩🚩🚩Hugging Face 实战系列 总目录
有任何问题欢迎在下面留言
本篇文章的代码运行界面均在Jupyter Notebook中进行
本篇文章配套的代码资源已经上传
上篇内容:
Hugging Face实战-系列教程8:GLUE数据集/文本分类上(NLP实战/Transformer实战/预训练模型/分词器/模型微调/模型自动选择/PyTorch版本/代码逐行解析)
3 模型训练
咱玩一个东西,要带着问题去玩儿,有的人特别擅长做笔记,拿本拿笔记下来?能把所有参数都记下来,真没什么卵用。什么叫学习,多查,多练,遇到问题了,然后要去解决一个问题的一个过程,这才叫学习。
3.1模型参数
先打开这个API文档:
API文档,实际用的时候一定对应着来
API文档就是说明书,你得认真的看,有你想知道的一切答案
首先第一步,从Transformers中导进来训练参数
from transformers import TrainingArguments
training_args = TrainingArguments("test-trainer")
设置好后再打印出来看看:
print(training_args )
TrainingArguments(
_n_gpu=0, adafactor=False, adam_beta1=0.9, adam_beta2=0.999, adam_epsilon=1e-08, bf16=False, bf16_full_eval=False,
dataloader_drop_last=False, dataloader_num_workers=0,
dataloader_pin_memory=True, ddp_bucket_cap_mb=None,
ddp_find_unused_parameters=None, debug=[], deepspeed=None,
disable_tqdm=False, do_eval=False, do_predict=False, do_train=False,
eval_accumulation_steps=None, eval_steps=None,
evaluation_strategy=IntervalStrategy.NO, fp16=False,
fp16_backend=auto, fp16_full_eval=False, fp16_opt_level=O1,
gradient_accumulation_steps=1, gradient_checkpointing=False,
greater_is_better=None, group_by_length=False,
half_precision_backend=auto, hub_model_id=None,
hub_strategy=HubStrategy.EVERY_SAVE, hub_token=<HUB_TOKEN>,
ignore_data_skip=False, label_names=None, label_smoothing_factor=0.0,
learning_rate=5e-05, length_column_name=length,
load_best_model_at_end=False, local_rank=-1, log_level=-1,
log_level_replica=-1, log_on_each_node=True,
logging_dir=test-trainer\runs\May26_10-08-48_WIN-BM410VRSBIO,
logging_first_step=False, logging_nan_inf_filter=True,
logging_steps=500, logging_strategy=IntervalStrategy.STEPS,
lr_scheduler_type=SchedulerType.LINEAR, max_grad_norm=1.0,
max_steps=-1, metric_for_best_model=None, mp_parameters=,
no_cuda=False, num_train_epochs=3.0, optim=OptimizerNames.ADAMW_HF,
output_dir=test-trainer, overwrite_output_dir=False, past_index=-1,
per_device_eval_batch_size=8, per_device_train_batch_size=8,
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>, remove_unused_columns=True,
report_to=[‘tensorboard’, ‘wandb’], resume_from_checkpoint=None,
run_name=test-trainer, save_on_each_node=False, save_steps=500,
save_strategy=IntervalStrategy.STEPS, save_total_limit=None, seed=42,
sharded_ddp=[], skip_memory_metrics=True, tf32=None,
tpu_metrics_debug=False, tpu_num_cores=None,
use_legacy_prediction_loop=False, warmup_ratio=0.0, warmup_steps=0,
weight_decay=0.0, xpu_backend=None, )
我的天哪,这么多参数,这些参数都能改吗?
你都能改,要训练模型的时候,这些参数都要指定的
就算你背下来了,你还是要忘,就是要边查边用
比如说我要指定batch怎么指定呢?指定epochs怎么指定呢?
你打开API文档,看看人家API文档做的多漂亮。
鼠标停在第一个参数上:
第一个就是输出路径,自己读一遍,模型保存的位置对不对?后面的也是这样一个一个看的。
前面我们打印出来的都是默认的参数
3.2模型导入
接下来导一下模型:
from transformers import AutoModelForSequenceClassification
model = AutoModelForSequenceClassification.from_pretrained(checkpoint, num_labels=2)
模型有一些提示:
Some weights of the model checkpoint at bert-base-uncased were not
used when initializing BertForSequenceClassification:
[‘cls.predictions.bias’, ‘cls.predictions.transform.dense.bias’,
‘cls.predictions.transform.LayerNorm.weight’,
‘cls.predictions.transform.dense.weight’,
‘cls.predictions.decoder.weight’, ‘cls.seq_relationship.bias’,
‘cls.seq_relationship.weight’,
‘cls.predictions.transform.LayerNorm.bias’]
- This IS expected if you are initializing BertForSequenceClassification from the checkpoint of a model trained
on another task or with another architecture (e.g. initializing a
BertForSequenceClassification model from a BertForPreTraining model).- This IS NOT expected if you are initializing BertForSequenceClassification from the checkpoint of a model that you
expect to be exactly identical (initializing a
BertForSequenceClassification model from a
BertForSequenceClassification model). Some weights of
BertForSequenceClassification were not initialized from the model
checkpoint at bert-base-uncased and are newly initialized:
[‘classifier.weight’, ‘classifier.bias’] You should probably TRAIN
this model on a down-stream task to be able to use it for predictions
and inference.
首先确定你任务是什么,比如对序列进行分类,就导入AutoModelForSequenceClassification,选择模型checkpoint,num_labels=2是什么意思?我们要改输出层,输出层不用预训练模型了,输出层自己训练。
所以上面的提示告诉你,很多分类层的权重参数没有指定到,就是分类的输出层被自己初始化了,无法加载预训练模型了,当然了正合我们意。
3.3模型训练
模型咋训练?哎呀,太简单了,真的嗷嗷简单:
from transformers import Trainer
trainer = Trainer(
model,
training_args,
train_dataset=tokenized_datasets["train"],
eval_dataset=tokenized_datasets["validation"],
data_collator=data_collator,
tokenizer=tokenizer,
)
无论训练什么都把Trainer导进来,看看参数
- model:我们在上面已经定义了
- training_args:配置参数,前面打印过,现在全是默认的,但是可以改,后续再教怎么改
- train_dataset:训练集,自己指定,根据前面定义的字典
- eval_dataset:验证集,自己指定,根据前面定义的字典
- data_collator:这是前面提到的
- batch tokenizer:前面也定义了
不懂没关系,再次点开前面提到的API,搜一下Trainer,要等个几秒钟才会出现:
不懂就去API里面查:
看看人家这在线API做的,多招人稀罕啊,鼠标放上面就有解释了。
指定好参数,直接.train一下就开始训练了:
trainer.train()
训练过程中会给你打印出损失:
再看 training_args参数中,有一个叫logging_steps=500,就是说500次打印一次损失
还会告诉你一些已经指定的参数:
The following columns in the training set don’t have a corresponding
argument inBertForSequenceClassification.forward
and have been
ignored: sentence2, idx, sentence1.
***** Running training ***** Num examples = 3668 Num Epochs = 3 Instantaneous batch size per device = 8 Total train batch size (w.
parallel, distributed & accumulation) = 8 Gradient Accumulation
steps = 1 Total optimization steps = 1377
其实这个任务CPU也能跑,但是比较慢,但是最好还是有GPU这个东西哈。
跑完之后还有提示:
Saving model checkpoint to test-trainer\checkpoint-500 Configuration
saved in test-trainer\checkpoint-500\config.json Model weights saved
in test-trainer\checkpoint-500\pytorch_model.bin tokenizer config file
saved in test-trainer\checkpoint-500\tokenizer_config.json Special
tokens file saved in
test-trainer\checkpoint-500\special_tokens_map.json Saving model
checkpoint to test-trainer\checkpoint-1000 Configuration saved in
test-trainer\checkpoint-1000\config.json Model weights saved in
test-trainer\checkpoint-1000\pytorch_model.bin tokenizer config file
saved in test-trainer\checkpoint-1000\tokenizer_config.json Special
tokens file saved in
test-trainer\checkpoint-1000\special_tokens_map.jsonTraining completed. Do not forget to share your model on
huggingface.co/models =)
就是你的模型都保存在哪儿了,训练完成后,就可以得到模型了:
这分别是500打印一次损失的结果,1000打印一次损失的结果,点进去看,pytorch_model.bin这个文件,就是你训练的模型
这就是一个训练过程
4 模型测试
4.1模型测试
模型训练好了,用验证集进行一下验证:
predictions = trainer.predict(tokenized_datasets["validation"])
print(predictions.predictions.shape, predictions.label_ids.shape)
打印的结果:(408, 2) (408,),当然这是打印的维度
前面给到的都是损失值,能不能给出具体的评估呢?datasets 模块专门提供了评估子模块load_metric
from datasets import load_metric
metric = load_metric("glue", "mrpc")
metric.compute(predictions=preds, references=predictions.label_ids)
打印结果:
A Jupyter Widget {‘accuracy’: 0.8186274509803921, ‘f1’:
0.8754208754208753}
在评估的参数中,只需要传入两个值,一个是predictions,一个是references,预测和标签嘛
4.2训练评估函数
我们在训练过程中能不能指定评估参数呢,那就需要将它封装成一个函数了:
def compute_metrics(eval_preds):
metric = load_metric("glue", "mrpc")
logits, labels = eval_preds
predictions = np.argmax(logits, axis=-1)
return metric.compute(predictions=predictions, references=labels)
逐行解释:
首先函数名字无所谓
还是加载默认的方法
输入参数只有一个值,但是在这个函数中需要做一个解开操作logits, labels = eval_preds
labels 是真实的标签,logits是一个中间结果不是实际预测结果,将logits中最大的取出来(模型中预测的最大概率)
然后再把预测和标签传进去返回
最后在训练参数中将上面的函数指定进去:
training_args = TrainingArguments("test-trainer", evaluation_strategy="epoch")
model = AutoModelForSequenceClassification.from_pretrained(checkpoint, num_labels=2)
trainer = Trainer(
model,
training_args,
train_dataset=tokenized_datasets["train"],
eval_dataset=tokenized_datasets["validation"],
data_collator=data_collator,
tokenizer=tokenizer,
compute_metrics=compute_metrics,
)
compute_metrics=compute_metrics,这是一个固定的写法
再训练看一下:
trainer.train()
这回打印的指标就变多了:
这就完了,源码点我直达。
这就完了,这简直就是,简单TM给简单开门,简单到家了
上篇内容:
Hugging Face实战-系列教程8:GLUE数据集/文本分类上(NLP实战/Transformer实战/预训练模型/分词器/模型微调/模型自动选择/PyTorch版本/代码逐行解析)