文章目录
python
复制代码
!pip install transformers datasets evaluate accelerate 一、改写文本分类
1.导入相关包
python
复制代码
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
})
'''
python
复制代码
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
python
复制代码
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
python
复制代码
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
python
复制代码
from transformers import BertForSequenceClassification
model = AutoModelForSequenceClassification.from_pretrained("hfl/chinese-macbert-base", num_labels=2) 6.创建评估函数
python
复制代码
import evaluate
acc_metric = evaluate.load("accuracy")
f1_metirc = evaluate.load("f1")
python
复制代码
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
python
复制代码
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
python
复制代码
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
python
复制代码
trainer.train() 10.模型评估
python
复制代码
trainer.evaluate(tokenized_datasets["test"]) 11.模型预测
python
复制代码
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}
'''
python
复制代码
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
python
复制代码
from transformers import AutoTokenizer, AutoModelForSequenceClassification, Trainer, TrainingArguments
from datasets import load_dataset 2.加载数据集
python
复制代码
dataset = load_dataset("json", data_files="./train_pair_1w.json", split="train")
dataset
'''
Dataset({
features: ['sentence1', 'sentence2', 'label'],
num_rows: 10000
})
'''
python
复制代码
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
python
复制代码
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
python
复制代码
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
})
}
'''
python
复制代码
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）
python
复制代码
from transformers import BertForSequenceClassification
model = AutoModelForSequenceClassification.from_pretrained("hfl/chinese-macbert-base", num_labels=1) 6.创建评估函数（区别）
python
复制代码
import evaluate
acc_metric = evaluate.load("accuracy")
f1_metirc = evaluate.load("f1")
python
复制代码
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
python
复制代码
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
python
复制代码
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
python
复制代码
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})
''' 10.模型评估
python
复制代码
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）
python
复制代码
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}
''' 
