第8.15节 Python重写自定义类的__repr__方法

本文深入探讨Python类中的__repr__方法,讲解其作用、语法与重写技巧,通过实例展示如何输出详尽的类信息,帮助开发者更好地进行程序调试。

一、 引言
前面两节分别介绍了Python类中的__str__和__repr__方法的作用和语法,所有新式类都支持这两个方法,因为object类实现了这两个方法,但实际上各位开发者在自定义类的过程中,应该重写相关方法,特别是__str__输出的应该是用户关注的容易理解的信息,因此对那些负责与客户交互的类,至少更应该重写__str__方法。
二、 重写__repr__方法
object类提供的__repr__方法,只是简单输出模块名、类名以及对象的ID信息,如果开发者要输出开发人员足够知悉的属性,就需要重写该方法。
重写__repr__方法注意:

  1. __repr__方法是实例方法,因此带一个参数self,也只能带这个参数;
  2. 输出的信息尽可能满足开发者的要求,信息必须详尽和准确;
  3. 可以调用父类的__repr__方法,也可以不调用,需要看开发时输出信息的安排,老猿建议最好调用,这样不用子类开发者去考虑父类的信息怎么输出。
    三、 例子
  4. 例子说明
    本节利用Car类来说明__repr__方法,在类中重写了__repr__方法,返回的信息包括:
  1. 证明调用了重写方法的“In repr:”提示;
  2. 调用repr(self.class),id(self)获取的对象的类名和ID基础信息;
  3. 使用__dict__特殊属性输出类的自定义属性及值的字典数据。
  1. 案例代码
    1> 类定义
class Car():  
    def __init__(self,brand,type,color,weight,oilcostper100km):
        self.brand = brand
        self.type = type
        self.oilcostper100km = oilcostper100km
        self.color = color
        self.weight = weight
      
    def __repr__(self):
        s="In __repr__:\n    <{} object at {:#016x}>\n".format(repr(self.__class__),id(self) )#使用类相关的信息输出类名和实例ID
        s+=super().__repr__() #直接调用object.__repr__方法输出类信息和实例ID
        #通过实例的self.__dict__获取实例的详细属性信息,注意使用了repr函数去获取字典信息字符串
        s+='\n'+repr(self.__dict__)
        return s

2> 实例定义及查看信息()

car = Car('雪佛兰','科帕奇','白色',1.8,10)
car #查看实例信息,此时会调用__repr__方法
print(car) #查看实例信息,此时会调用__str__方法,但没有重写__str__方法,效果会怎样?
  1. 案例截图在这里插入图片描述
  2. 案例总结
    通过《第8.13节 Python类中内置方法的__repr__详解》知道,没有重写__repr__方法时,调用的是object类的__repr__方法,仅输出简单的类信息和ID信息。通过上述案例说明:
    1> 重写__repr__方法后,查看对象会触发重写方法的调用,其他触发方式请见《第8.13节 Python类中内置方法__repr__详解》,在例子中没有再展开介绍;
    2>object类__repr__方法输出结果与repr(self.class),id(self)获取的基础信息相同,而直接调用object类__repr__方法更简单,因此老猿建议对基础信息的输出还是直接调用父类的方法输出;
    3> 使用print(car)应该输出__str__方法返回的内容,但实际输出的是重写__repr__方法输出的内容。这是因为自定义类没有重写__str__方法的情况下,Python会直接调用__ repr __方法输出,要输出单独的__str__信息需要重写__str__方法,这个在下节将进行验证。

本节结合案例介绍了 __repr__方法的重写,可以看到通过重写该方法如果输出数据得当,相关信息有利于程序调试和跟踪。
老猿Python系列文章用于逐步介绍老猿学习Python后总结的学习经验,这些经验有助于没有接触过Python的程序员可以很容易地进入Python的世界。
欢迎大家批评指正,谢谢大家关注!

代码出现问题:(style_tune) C:\Users\28996\Desktop\AI\persona_contrastive_finetuning>python Contrastive_Training_LM.py INFO:accelerate.utils.modeling:We will use 90% of the memory on device 0 for storing the model, and 10% for the buffer to avoid OOM. You can set `max_memory` in to a higher value to use more memory (at your own risk). trainable params: 1,572,864 || all params: 1,838,401,536 || trainable%: 0.0856 训练集样本示例: {'anchor_input_ids': [56568, 118919, 116122, 11319], 'positive_input_ids': [116122, 20412, 107340, 9370, 100357, 102323, 3837, 109202, 104078, 103975, 100675, 101940, 100912, 105054, 6313], 'negative_input_ids': [100323, 104307, 99245, 9370, 106059, 104060, 3837, 104530, 115604, 99329, 11319]} 验证集样本示例: {'anchor_input_ids': [56568, 118919, 116122, 11319], 'positive_input_ids': [116122, 20412, 107340, 9370, 100357, 102323, 3837, 109202, 104078, 103975, 100675, 101940, 100912, 105054, 6313], 'negative_input_ids': [100323, 104307, 99245, 9370, 106059, 104060, 3837, 104530, 115604, 99329, 11319]} Trainer.tokenizer is now deprecated. You should use `Trainer.processing_class = processing_class` instead. INFO:__main__:GPU内存使用: 已分配 2.93GB, 保留 4.13GB 可训练参数列表: - base_model.model.model.layers.0.self_attn.q_proj.lora_A.default.weight - base_model.model.model.layers.0.self_attn.q_proj.lora_B.default.weight - base_model.model.model.layers.0.self_attn.v_proj.lora_A.default.weight - base_model.model.model.layers.0.self_attn.v_proj.lora_B.default.weight - base_model.model.model.layers.1.self_attn.q_proj.lora_A.default.weight - base_model.model.model.layers.1.self_attn.q_proj.lora_B.default.weight - base_model.model.model.layers.1.self_attn.v_proj.lora_A.default.weight - base_model.model.model.layers.1.self_attn.v_proj.lora_B.default.weight - base_model.model.model.layers.2.self_attn.q_proj.lora_A.default.weight - base_model.model.model.layers.2.self_attn.q_proj.lora_B.default.weight - base_model.model.model.layers.2.self_attn.v_proj.lora_A.default.weight - base_model.model.model.layers.2.self_attn.v_proj.lora_B.default.weight - base_model.model.model.layers.3.self_attn.q_proj.lora_A.default.weight - base_model.model.model.layers.3.self_attn.q_proj.lora_B.default.weight - base_model.model.model.layers.3.self_attn.v_proj.lora_A.default.weight - base_model.model.model.layers.3.self_attn.v_proj.lora_B.default.weight - base_model.model.model.layers.4.self_attn.q_proj.lora_A.default.weight - base_model.model.model.layers.4.self_attn.q_proj.lora_B.default.weight - base_model.model.model.layers.4.self_attn.v_proj.lora_A.default.weight - base_model.model.model.layers.4.self_attn.v_proj.lora_B.default.weight - base_model.model.model.layers.5.self_attn.q_proj.lora_A.default.weight - base_model.model.model.layers.5.self_attn.q_proj.lora_B.default.weight - base_model.model.model.layers.5.self_attn.v_proj.lora_A.default.weight - base_model.model.model.layers.5.self_attn.v_proj.lora_B.default.weight - base_model.model.model.layers.6.self_attn.q_proj.lora_A.default.weight - base_model.model.model.layers.6.self_attn.q_proj.lora_B.default.weight - base_model.model.model.layers.6.self_attn.v_proj.lora_A.default.weight - base_model.model.model.layers.6.self_attn.v_proj.lora_B.default.weight - base_model.model.model.layers.7.self_attn.q_proj.lora_A.default.weight - base_model.model.model.layers.7.self_attn.q_proj.lora_B.default.weight - base_model.model.model.layers.7.self_attn.v_proj.lora_A.default.weight - base_model.model.model.layers.7.self_attn.v_proj.lora_B.default.weight - base_model.model.model.layers.8.self_attn.q_proj.lora_A.default.weight - base_model.model.model.layers.8.self_attn.q_proj.lora_B.default.weight - base_model.model.model.layers.8.self_attn.v_proj.lora_A.default.weight - base_model.model.model.layers.8.self_attn.v_proj.lora_B.default.weight - base_model.model.model.layers.9.self_attn.q_proj.lora_A.default.weight - base_model.model.model.layers.9.self_attn.q_proj.lora_B.default.weight - base_model.model.model.layers.9.self_attn.v_proj.lora_A.default.weight - base_model.model.model.layers.9.self_attn.v_proj.lora_B.default.weight - base_model.model.model.layers.10.self_attn.q_proj.lora_A.default.weight - base_model.model.model.layers.10.self_attn.q_proj.lora_B.default.weight - base_model.model.model.layers.10.self_attn.v_proj.lora_A.default.weight - base_model.model.model.layers.10.self_attn.v_proj.lora_B.default.weight - base_model.model.model.layers.11.self_attn.q_proj.lora_A.default.weight - base_model.model.model.layers.11.self_attn.q_proj.lora_B.default.weight - base_model.model.model.layers.11.self_attn.v_proj.lora_A.default.weight - base_model.model.model.layers.11.self_attn.v_proj.lora_B.default.weight - base_model.model.model.layers.12.self_attn.q_proj.lora_A.default.weight - base_model.model.model.layers.12.self_attn.q_proj.lora_B.default.weight - base_model.model.model.layers.12.self_attn.v_proj.lora_A.default.weight - base_model.model.model.layers.12.self_attn.v_proj.lora_B.default.weight - base_model.model.model.layers.13.self_attn.q_proj.lora_A.default.weight - base_model.model.model.layers.13.self_attn.q_proj.lora_B.default.weight - base_model.model.model.layers.13.self_attn.v_proj.lora_A.default.weight - base_model.model.model.layers.13.self_attn.v_proj.lora_B.default.weight - base_model.model.model.layers.14.self_attn.q_proj.lora_A.default.weight - base_model.model.model.layers.14.self_attn.q_proj.lora_B.default.weight - base_model.model.model.layers.14.self_attn.v_proj.lora_A.default.weight - base_model.model.model.layers.14.self_attn.v_proj.lora_B.default.weight - base_model.model.model.layers.15.self_attn.q_proj.lora_A.default.weight - base_model.model.model.layers.15.self_attn.q_proj.lora_B.default.weight - base_model.model.model.layers.15.self_attn.v_proj.lora_A.default.weight - base_model.model.model.layers.15.self_attn.v_proj.lora_B.default.weight - base_model.model.model.layers.16.self_attn.q_proj.lora_A.default.weight - base_model.model.model.layers.16.self_attn.q_proj.lora_B.default.weight - base_model.model.model.layers.16.self_attn.v_proj.lora_A.default.weight - base_model.model.model.layers.16.self_attn.v_proj.lora_B.default.weight - base_model.model.model.layers.17.self_attn.q_proj.lora_A.default.weight - base_model.model.model.layers.17.self_attn.q_proj.lora_B.default.weight - base_model.model.model.layers.17.self_attn.v_proj.lora_A.default.weight - base_model.model.model.layers.17.self_attn.v_proj.lora_B.default.weight - base_model.model.model.layers.18.self_attn.q_proj.lora_A.default.weight - base_model.model.model.layers.18.self_attn.q_proj.lora_B.default.weight - base_model.model.model.layers.18.self_attn.v_proj.lora_A.default.weight - base_model.model.model.layers.18.self_attn.v_proj.lora_B.default.weight - base_model.model.model.layers.19.self_attn.q_proj.lora_A.default.weight - base_model.model.model.layers.19.self_attn.q_proj.lora_B.default.weight - base_model.model.model.layers.19.self_attn.v_proj.lora_A.default.weight - base_model.model.model.layers.19.self_attn.v_proj.lora_B.default.weight - base_model.model.model.layers.20.self_attn.q_proj.lora_A.default.weight - base_model.model.model.layers.20.self_attn.q_proj.lora_B.default.weight - base_model.model.model.layers.20.self_attn.v_proj.lora_A.default.weight - base_model.model.model.layers.20.self_attn.v_proj.lora_B.default.weight - base_model.model.model.layers.21.self_attn.q_proj.lora_A.default.weight - base_model.model.model.layers.21.self_attn.q_proj.lora_B.default.weight - base_model.model.model.layers.21.self_attn.v_proj.lora_A.default.weight - base_model.model.model.layers.21.self_attn.v_proj.lora_B.default.weight - base_model.model.model.layers.22.self_attn.q_proj.lora_A.default.weight - base_model.model.model.layers.22.self_attn.q_proj.lora_B.default.weight - base_model.model.model.layers.22.self_attn.v_proj.lora_A.default.weight - base_model.model.model.layers.22.self_attn.v_proj.lora_B.default.weight - base_model.model.model.layers.23.self_attn.q_proj.lora_A.default.weight - base_model.model.model.layers.23.self_attn.q_proj.lora_B.default.weight - base_model.model.model.layers.23.self_attn.v_proj.lora_A.default.weight - base_model.model.model.layers.23.self_attn.v_proj.lora_B.default.weight 0%| | 0/3 [00:00<?, ?it/s]You're using a Qwen2TokenizerFast tokenizer. Please note that with a fast tokenizer, using the `__call__` method is faster than using a method to encode the text followed by a call to the `pad` method to get a padded encoding. Trainer.tokenizer is now deprecated. You should use Trainer.processing_class instead. Trainer.tokenizer is now deprecated. You should use Trainer.processing_class instead. INFO:__main__:GPU内存使用: 已分配 4.00GB, 保留 4.21GB Could not estimate the number of tokens of the input, floating-point operations will not be computed Trainer.tokenizer is now deprecated. You should use Trainer.processing_class instead. Trainer.tokenizer is now deprecated. You should use Trainer.processing_class instead. INFO:__main__:GPU内存使用: 已分配 4.02GB, 保留 4.22GB 33%|████████████████████████████ | 1/3 [00:03<00:06, 3.25s/it]Trainer.tokenizer is now deprecated. You should use Trainer.processing_class instead. Trainer.tokenizer is now deprecated. You should use Trainer.processing_class instead. INFO:__main__:GPU内存使用: 已分配 4.01GB, 保留 4.25GB Trainer.tokenizer is now deprecated. You should use Trainer.processing_class instead. Trainer.tokenizer is now deprecated. You should use Trainer.processing_class instead. INFO:__main__:GPU内存使用: 已分配 4.02GB, 保留 4.26GB 67%|████████████████████████████████████████████████████████ | 2/3 [00:06<00:02, 2.98s/it]Trainer.tokenizer is now deprecated. You should use Trainer.processing_class instead. Trainer.tokenizer is now deprecated. You should use Trainer.processing_class instead. INFO:__main__:GPU内存使用: 已分配 4.01GB, 保留 4.25GB Trainer.tokenizer is now deprecated. You should use Trainer.processing_class instead. Trainer.tokenizer is now deprecated. You should use Trainer.processing_class instead. INFO:__main__:GPU内存使用: 已分配 4.02GB, 保留 4.26GB {'train_runtime': 9.034, 'train_samples_per_second': 0.664, 'train_steps_per_second': 0.332, 'train_loss': 1.0772175788879395, 'epoch': 3.0} 100%|████████████████████████████████████████████████████████████████████████████████████| 3/3 [00:09<00:00, 3.01s/it] Traceback (most recent call last): File "C:\Users\28996\Desktop\AI\persona_contrastive_finetuning\Contrastive_Training_LM.py", line 356, in <module> eval_results = trainer.evaluate() File "C:\Users\28996\miniconda3\envs\style_tune\lib\site-packages\transformers\trainer.py", line 4076, in evaluate output = eval_loop( File "C:\Users\28996\miniconda3\envs\style_tune\lib\site-packages\transformers\trainer.py", line 4270, in evaluation_loop losses, logits, labels = self.prediction_step(model, inputs, prediction_loss_only, ignore_keys=ignore_keys) File "C:\Users\28996\miniconda3\envs\style_tune\lib\site-packages\transformers\trainer.py", line 4496, in prediction_step outputs = model(**inputs) File "C:\Users\28996\miniconda3\envs\style_tune\lib\site-packages\torch\nn\modules\module.py", line 1736, in _wrapped_call_impl return self._call_impl(*args, **kwargs) File "C:\Users\28996\miniconda3\envs\style_tune\lib\site-packages\torch\nn\modules\module.py", line 1747, in _call_impl return forward_call(*args, **kwargs) File "C:\Users\28996\miniconda3\envs\style_tune\lib\site-packages\accelerate\utils\operations.py", line 818, in forward return model_forward(*args, **kwargs) File "C:\Users\28996\miniconda3\envs\style_tune\lib\site-packages\accelerate\utils\operations.py", line 806, in __call__ return convert_to_fp32(self.model_forward(*args, **kwargs)) File "C:\Users\28996\miniconda3\envs\style_tune\lib\site-packages\torch\amp\autocast_mode.py", line 44, in decorate_autocast return func(*args, **kwargs) File "C:\Users\28996\miniconda3\envs\style_tune\lib\site-packages\peft\peft_model.py", line 1719, in forward return self.base_model( File "C:\Users\28996\miniconda3\envs\style_tune\lib\site-packages\torch\nn\modules\module.py", line 1736, in _wrapped_call_impl return self._call_impl(*args, **kwargs) File "C:\Users\28996\miniconda3\envs\style_tune\lib\site-packages\torch\nn\modules\module.py", line 1747, in _call_impl return forward_call(*args, **kwargs) File "C:\Users\28996\miniconda3\envs\style_tune\lib\site-packages\peft\tuners\tuners_utils.py", line 197, in forward return self.model.forward(*args, **kwargs) File "C:\Users\28996\miniconda3\envs\style_tune\lib\site-packages\transformers\models\qwen2\modeling_qwen2.py", line 816, in forward outputs = self.model( File "C:\Users\28996\miniconda3\envs\style_tune\lib\site-packages\torch\nn\modules\module.py", line 1736, in _wrapped_call_impl return self._call_impl(*args, **kwargs) File "C:\Users\28996\miniconda3\envs\style_tune\lib\site-packages\torch\nn\modules\module.py", line 1747, in _call_impl return forward_call(*args, **kwargs) File "C:\Users\28996\miniconda3\envs\style_tune\lib\site-packages\transformers\models\qwen2\modeling_qwen2.py", line 521, in forward raise ValueError("You must specify exactly one of input_ids or inputs_embeds") ValueError: You must specify exactly one of input_ids or inputs_embeds (style_tune) C:\Users\28996\Desktop\AI\persona_contrastive_finetuning>python Contrastive_Training_LM.py Traceback (most recent call last): File "C:\Users\28996\Desktop\AI\persona_contrastive_finetuning\Contrastive_Training_LM.py", line 57, in <module> class ContrastiveTrainer(Trainer): File "C:\Users\28996\Desktop\AI\persona_contrastive_finetuning\Contrastive_Training_LM.py", line 63, in ContrastiveTrainer eval_dataset: Optional[Dataset] = None, NameError: name 'Dataset' is not defined 原代码如下:import torch import torch.nn as nn import torch.nn.functional as F from transformers import ( AutoModelForCausalLM, AutoTokenizer, TrainingArguments, Trainer, PreTrainedTokenizerBase, BitsAndBytesConfig ) from transformers.tokenization_utils_base import PreTrainedTokenizerBase from transformers.utils import PaddingStrategy from datasets import load_dataset from typing import Any, Dict, List, Optional, Tuple, Union import logging from dataclasses import dataclass import os import gc from peft import LoraConfig, get_peft_model, prepare_model_for_kbit_training @dataclass class EvalDataCollator: """评估专用的数据收集器""" tokenizer: PreTrainedTokenizerBase padding: Union[bool, str, PaddingStrategy] = True max_length: Optional[int] = None pad_to_multiple_of: Optional[int] = None return_tensors: str = "pt" def __call__(self, features: List[Dict[str, Any]]) -> Dict[str, torch.Tensor]: # 评估时只使用正样本(用于语言建模评估) positive_features = [{"input_ids": f["positive_input_ids"]} for f in features] # 对正样本进行填充 batch_positive = self.tokenizer.pad( positive_features, padding=self.padding, max_length=self.max_length, pad_to_multiple_of=self.pad_to_multiple_of, return_tensors=self.return_tensors, ) # 创建注意力掩码 attention_mask = (batch_positive["input_ids"] != self.tokenizer.pad_token_id).int() # 创建标签(用于语言建模) labels = batch_positive["input_ids"].clone() labels[labels == self.tokenizer.pad_token_id] = -100 return { "input_ids": batch_positive["input_ids"], "attention_mask": attention_mask, "labels": labels } class ContrastiveTrainer(Trainer): """内存优化的训练器""" # ... [保持其他方法不变] ... def evaluate( self, eval_dataset: Optional[Dataset] = None, ignore_keys: Optional[List[str]] = None, metric_key_prefix: str = "eval", ) -> Dict[str, float]: """重写评估方法以使用专用的数据收集器""" # 创建评估专用的数据收集器 eval_data_collator = EvalDataCollator( tokenizer=self.tokenizer, max_length=256, padding="max_length" ) # 临时保存原始数据收集器 original_collator = self.data_collator try: # 使用评估专用的数据收集器 self.data_collator = eval_data_collator # 调用父的评估方法 return super().evaluate( eval_dataset=eval_dataset, ignore_keys=ignore_keys, metric_key_prefix=metric_key_prefix ) finally: # 恢复原始数据收集器 self.data_collator = original_collator # 设置日志 logging.basicConfig(level=logging.INFO) logger = logging.getLogger(__name__) # 内存优化工具函数 def clear_memory(): """清除Python和CUDA缓存""" gc.collect() if torch.cuda.is_available(): torch.cuda.empty_cache() torch.cuda.reset_peak_memory_stats() def print_memory_usage(): """打印当前内存使用情况""" if torch.cuda.is_available(): allocated = torch.cuda.memory_allocated() / (1024 ** 3) reserved = torch.cuda.memory_reserved() / (1024 ** 3) logger.info(f"GPU内存使用: 已分配 {allocated:.2f}GB, 保留 {reserved:.2f}GB") else: logger.info("未检测到GPU") def tokenize_function(examples, tokenizer, max_length=256): """将文本转换为token IDs""" tokenized = {} # 对每个字段进行分词 for key in ['anchor', 'positive', 'negative']: if key in examples: # 使用分词器处理文本 result = tokenizer( examples[key], max_length=max_length, truncation=True, padding=False, return_tensors=None ) tokenized[f"{key}_input_ids"] = result["input_ids"] return tokenized @dataclass class ContrastiveDataCollator: """内存优化的数据收集器""" tokenizer: PreTrainedTokenizerBase padding: Union[bool, str, PaddingStrategy] = True max_length: Optional[int] = None pad_to_multiple_of: Optional[int] = None return_tensors: str = "pt" def __call__(self, features: List[Dict[str, Any]]) -> Dict[str, torch.Tensor]: # 分离出三元组的各个部分 anchor_features = [{"input_ids": f["anchor_input_ids"]} for f in features] positive_features = [{"input_ids": f["positive_input_ids"]} for f in features] negative_features = [{"input_ids": f["negative_input_ids"]} for f in features] # 对每个部分分别进行填充 batch_anchor = self.tokenizer.pad( anchor_features, padding=self.padding, max_length=self.max_length, pad_to_multiple_of=self.pad_to_multiple_of, return_tensors=self.return_tensors, ) batch_positive = self.tokenizer.pad( positive_features, padding=self.padding, max_length=self.max_length, pad_to_multiple_of=self.pad_to_multiple_of, return_tensors=self.return_tensors, ) batch_negative = self.tokenizer.pad( negative_features, padding=self.padding, max_length=self.max_length, pad_to_multiple_of=self.pad_to_multiple_of, return_tensors=self.return_tensors, ) # 创建注意力掩码 def create_attention_mask(input_ids): return (input_ids != self.tokenizer.pad_token_id).int() # 释放中间变量内存 del anchor_features, positive_features, negative_features clear_memory() return { "anchor_input_ids": batch_anchor["input_ids"], "anchor_attention_mask": create_attention_mask(batch_anchor["input_ids"]), "positive_input_ids": batch_positive["input_ids"], "positive_attention_mask": create_attention_mask(batch_positive["input_ids"]), "negative_input_ids": batch_negative["input_ids"], "negative_attention_mask": create_attention_mask(batch_negative["input_ids"]), } class ContrastiveTrainer(Trainer): """内存优化的训练器""" def __init__(self, tokenizer=None, *args, contrastive_config=None, **kwargs): # 首先调用父初始化 super().__init__(*args, **kwargs) # 关键修复:设置tokenizer self.tokenizer = tokenizer if contrastive_config is None: contrastive_config = {} # 设置默认值 self.temperature = contrastive_config.get("temperature", 0.07) self.margin = contrastive_config.get("margin", 0.3) self.contrastive_weight = contrastive_config.get("weight", 0.8) self.repr_layer = contrastive_config.get("repr_layer", -1) # 验证必要参数 if not hasattr(self.model.config, "output_hidden_states") or not self.model.config.output_hidden_states: raise ValueError("模型必须设置output_hidden_states=True") self.cross_entropy = nn.CrossEntropyLoss() def compute_contrastive_loss(self, anchor_emb, pos_emb, neg_emb): """计算对比损失""" # 计算余弦相似度 pos_sim = F.cosine_similarity(anchor_emb, pos_emb) neg_sim = F.cosine_similarity(anchor_emb, neg_emb) # 计算InfoNCE损失 numerator = torch.exp(pos_sim / self.temperature) denominator = numerator + torch.exp(neg_sim / self.temperature) info_nce_loss = -torch.log(numerator / (denominator + 1e-8)).mean() # 计算三元组损失 triplet_loss = F.relu(neg_sim - pos_sim + self.margin).mean() return info_nce_loss + triplet_loss def get_sequence_representation(self, outputs, attention_mask): """获取序列表示(内存优化版)""" # 只获取需要的隐藏状态层 hidden_states = outputs.hidden_states[self.repr_layer] # 获取每个序列的最后一个非填充token seq_lengths = attention_mask.sum(dim=1) - 1 batch_indices = torch.arange(hidden_states.size(0)) # 返回对应位置的隐藏状态 return hidden_states[batch_indices, seq_lengths] def compute_loss(self, model, inputs, return_outputs=False): """内存优化的损失计算""" # 确保模型处于训练模式 model.train() # 提取输入 anchor_ids = inputs["anchor_input_ids"] anchor_mask = inputs["anchor_attention_mask"] positive_ids = inputs["positive_input_ids"] positive_mask = inputs["positive_attention_mask"] negative_ids = inputs["negative_input_ids"] negative_mask = inputs["negative_attention_mask"] # 前向传播获取隐藏状态 def get_embeddings(input_ids, attention_mask): outputs = model( input_ids=input_ids, attention_mask=attention_mask, output_hidden_states=True, return_dict=True ) return self.get_sequence_representation(outputs, attention_mask) # 获取三元组的嵌入表示 anchor_emb = get_embeddings(anchor_ids, anchor_mask) pos_emb = get_embeddings(positive_ids, positive_mask) neg_emb = get_embeddings(negative_ids, negative_mask) # 计算对比损失 cl_loss = self.compute_contrastive_loss(anchor_emb, pos_emb, neg_emb) cl_loss = cl_loss * self.contrastive_weight # 关键修复:确保tokenizer已设置 if self.tokenizer is None: raise ValueError("Tokenizer未设置!") # 计算语言建模损失 lm_labels = positive_ids.clone() # 关键修复:使用tokenizer的pad_token_id pad_token_id = self.tokenizer.pad_token_id lm_labels[lm_labels == pad_token_id] = -100 # 计算语言建模损失 lm_outputs = model( input_ids=positive_ids, attention_mask=positive_mask, labels=lm_labels ) lm_loss = lm_outputs.loss # 总损失 = LM损失 + 对比损失 total_loss = lm_loss + cl_loss # 记录内存使用 print_memory_usage() return (total_loss, lm_outputs) if return_outputs else total_loss # ================ 主程序 ================ # if __name__ == "__main__": # 配置量化以减少内存使用 bnb_config = BitsAndBytesConfig( load_in_4bit=True, # 使用4位量化 bnb_4bit_quant_type="nf4", # 使用NF4量化型 bnb_4bit_use_double_quant=True, # 双重量化 bnb_4bit_compute_dtype=torch.float16 # 计算使用FP16 ) # 加载模型和分词器(使用量化) model = AutoModelForCausalLM.from_pretrained( "model/Qwen/Qwen1.5-1.8B", quantization_config=bnb_config, # 应用量化配置 device_map="auto", # 自动选择设备 output_hidden_states=True, # 必须设置以获取隐藏状态 return_dict_in_generate=True, use_cache=False # 禁用缓存以省内存 ) tokenizer = AutoTokenizer.from_pretrained("model/Qwen/Qwen1.5-1.8B") tokenizer.pad_token = tokenizer.eos_token # 设置填充token # 为量化模型添加LoRA适配器 lora_config = LoraConfig( r=8, lora_alpha=32, target_modules=["q_proj", "v_proj"], # 针对Qwen1.5-1.8B模型 lora_dropout=0.05, bias="none", task_type="CAUSAL_LM" ) # 关键修复:准备模型用于k位训练 model = prepare_model_for_kbit_training(model, use_gradient_checkpointing=True) # 添加LoRA适配器 model = get_peft_model(model, lora_config) # 关键修复:显式启用LoRA参数的梯度 for param in model.parameters(): if param.requires_grad: param.requires_grad = True model.print_trainable_parameters() # 打印可训练参数数量 # 加载数据集 def load_and_tokenize_dataset(file_path, tokenizer): """加载数据集并进行分词处理""" # 加载原始数据集 dataset_dict = load_dataset('json', data_files=file_path) raw_dataset = dataset_dict['train'] # 应用分词函数 tokenized_dataset = raw_dataset.map( lambda ex: tokenize_function(ex, tokenizer, max_length=256), batched=True, batch_size=8, # 减小批处理大小 remove_columns=['anchor', 'positive', 'negative'] ) return tokenized_dataset train_dataset = load_and_tokenize_dataset('data/processed/train_style_triplets.json', tokenizer) val_dataset = load_and_tokenize_dataset('data/processed/val_style_triplets.json', tokenizer) # 验证数据集格式 print("训练集样本示例:", train_dataset[0]) print("验证集样本示例:", val_dataset[0]) # 训练参数配置(内存优化) training_args = TrainingArguments( output_dir="./model/lora_adapter", per_device_train_batch_size=1, # 减小批量大小 gradient_accumulation_steps=8, # 增加梯度累积步数 num_train_epochs=3, learning_rate=2e-4, logging_steps=10, # 更频繁的日志记录以监控内存 save_steps=500, fp16=True, report_to="none", remove_unused_columns=False, gradient_checkpointing=True, # 启用梯度检查点 optim="adafactor", # 使用内存更少的优化器 ) # 对比学习配置 contrastive_config = { "temperature": 0.07, "margin": 0.3, "weight": 0.8, "repr_layer": -1 } # 初始化数据收集器 data_collator = ContrastiveDataCollator( tokenizer=tokenizer, max_length=256, # 减少最大长度 padding="max_length" ) # 初始化训练器 - 关键修复:传递tokenizer trainer = ContrastiveTrainer( model=model, args=training_args, tokenizer=tokenizer, # 传递tokenizer data_collator=data_collator, train_dataset=train_dataset, eval_dataset=val_dataset, contrastive_config=contrastive_config ) # 开始训练前打印内存状态 print_memory_usage() # 关键修复:验证可训练参数 print("可训练参数列表:") for name, param in model.named_parameters(): if param.requires_grad: print(f"- {name}") # 开始训练 trainer.train() # 保存LoRA适配器 model.save_pretrained("./model/lora_adapter") # 评估模型 try: eval_results = trainer.evaluate() print("评估结果:", eval_results) except Exception as e: print(f"评估过程中发生错误: {e}") import traceback traceback.print_exc()
最新发布
07-21
将以下两份代码分别修改,使其可用完美配合:代码1:# 使用示例 from transformers import AutoModelForCausalLM, AutoTokenizer, TrainingArguments, Trainer from contrastive_trainer import ContrastiveTrainer, ContrastiveDataCollator from datasets import load_dataset class ContrastiveTrainer(Trainer): def __init__(self, *args, **kwargs): super().__init__(*args, **kwargs) # 确保父初始化 self._prepare_for_training() # 显式调用 def _prepare_for_training(self): # 自定义逻辑或直接复用父方法 super()._prepare_for_training() # 调用父实现 # 加载模型和分词器 model = AutoModelForCausalLM.from_pretrained("model/Qwen/Qwen1.5-1.8B") tokenizer = AutoTokenizer.from_pretrained("model/Qwen/Qwen1.5-1.8B") tokenizer.pad_token = tokenizer.eos_token train_dataset = load_dataset('json', data_files='data/processed/train_style_triplets.json')['train'] val_dataset = load_dataset('json', data_files='data/processed/val_style_triplets.json')['train'] # 验证集 # 训练参数配置 training_args = TrainingArguments( output_dir="./model/lora_adapter", per_device_train_batch_size=4, num_train_epochs=3, learning_rate=2e-4, logging_steps=50, save_steps=500, fp16=True, gradient_accumulation_steps=4, ) # 对比学习配置 contrastive_config = { "temperature": 0.07, "margin": 0.3, "weight": 0.8, "repr_layer": -1 # 使用最后一层隐藏状态 } # 初始化数据收集器 data_collator = ContrastiveDataCollator( tokenizer=tokenizer, max_length=512, padding="max_length" ) # 初始化训练器 (修正参数传递) trainer = ContrastiveTrainer( model=model, args=training_args, data_collator=data_collator, train_dataset=train_dataset, eval_dataset=val_dataset, contrastive_config=contrastive_config # 作为独立参数传递 ) # 开始训练 trainer.train() # 评估模型 eval_results = trainer.evaluate() 代码2:# trainer/contrastive_trainer.py import torch import torch.nn as nn import torch.nn.functional as F from torch.utils.data import Dataset, DataLoader from transformers import Trainer, TrainingArguments from transformers.tokenization_utils_base import PreTrainedTokenizerBase from transformers.utils import PaddingStrategy from typing import Any, Dict, List, Optional, Tuple, Union import logging import numpy as np import os from tqdm import tqdm from dataclasses import dataclass class ContrastiveTrainer(Trainer): def __init__(self, *args, contrastive_config=None, **kwargs): # 1. 显式声明并提取自定义参数 self.contrastive_config = contrastive_config or {} # 2. 验证配置完整性 required_keys = ["temperature", "margin", "weight", "repr_layer"] if any(key not in self.contrastive_config for key in required_keys): raise ValueError("Missing required keys in contrastive_config") # 3. 安全移除自定义参数后调用父初始化 super().__init__(*args, **kwargs) # 此时 kwargs 不含 contrastive_config # 4. 初始化依赖配置的参数 self.temperature = self.contrastive_config["temperature"] self.margin = self.contrastive_config["margin"] self.contrastive_weight = self.contrastive_config["weight"] self.repr_layer = self.contrastive_config["repr_layer"] # 5. 损失函数初始化 self.cross_entropy = nn.CrossEntropyLoss() # 设置日志 logging.basicConfig(level=logging.INFO) logger = logging.getLogger(__name__) @dataclass class ContrastiveDataCollator: """ 对比学习数据收集器,处理对比学习的正负样本对 """ tokenizer: PreTrainedTokenizerBase padding: Union[bool, str, PaddingStrategy] = True max_length: Optional[int] = None pad_to_multiple_of: Optional[int] = None return_tensors: str = "pt" def __call__(self, features: List[Dict[str, Any]]) -> Dict[str, torch.Tensor]: """ 处理一批数据,生成模型输入格式 """ # 分离出三元组的各个部分 anchor_features = [{"input_ids": f["anchor_input_ids"]} for f in features] positive_features = [{"input_ids": f["positive_input_ids"]} for f in features] negative_features = [{"input_ids": f["negative_input_ids"]} for f in features] # 对每个部分分别进行填充 batch_anchor = self.tokenizer.pad( anchor_features, padding=self.padding, max_length=self.max_length, pad_to_multiple_of=self.pad_to_multiple_of, return_tensors=self.return_tensors, ) batch_positive = self.tokenizer.pad( positive_features, padding=self.padding, max_length=self.max_length, pad_to_multiple_of=self.pad_to_multiple_of, return_tensors=self.return_tensors, ) batch_negative = self.tokenizer.pad( negative_features, padding=self.padding, max_length=self.max_length, pad_to_multiple_of=self.pad_to_multiple_of, return_tensors=self.return_tensors, ) # 创建注意力掩码 def create_attention_mask(input_ids): mask = torch.ones_like(input_ids) mask[input_ids == self.tokenizer.pad_token_id] = 0 return mask # 返回一个字典,包含所有部分 return { "anchor_input_ids": batch_anchor["input_ids"], "anchor_attention_mask": create_attention_mask(batch_anchor["input_ids"]), "positive_input_ids": batch_positive["input_ids"], "positive_attention_mask": create_attention_mask(batch_positive["input_ids"]), "negative_input_ids": batch_negative["input_ids"], "negative_attention_mask": create_attention_mask(batch_negative["input_ids"]), } class ContrastiveTrainer(Trainer): """ 对比学习训练器,实现对比学习训练逻辑 """ def __init__( self, model: nn.Module = None, args: TrainingArguments = None, data_collator: Optional[ContrastiveDataCollator] = None, train_dataset: Optional[Dataset] = None, eval_dataset: Optional[Dataset] = None, tokenizer: Optional[PreTrainedTokenizerBase] = None, model_init: Optional[callable] = None, compute_metrics: Optional[callable] = None, callbacks: Optional[List[Any]] = None, optimizers: Tuple[torch.optim.Optimizer, torch.optim.lr_scheduler.LambdaLR] = (None, None), preprocess_logits_for_metrics: Optional[callable] = None, contrastive_config: Optional[Dict] = None ): super().__init__( model=model, args=args, data_collator=data_collator, train_dataset=train_dataset, eval_dataset=eval_dataset, tokenizer=tokenizer, model_init=model_init, compute_metrics=compute_metrics, callbacks=callbacks, optimizers=optimizers, preprocess_logits_for_metrics=preprocess_logits_for_metrics ) # 对比学习配置 self.contrastive_config = contrastive_config or {} self.temperature = self.contrastive_config.get("temperature", 0.07) self.margin = self.contrastive_config.get("margin", 0.3) self.contrastive_weight = self.contrastive_config.get("weight", 0.8) self.repr_layer = self.contrastive_config.get("repr_layer", -1) # 默认最后一层隐藏状态 # 损失函数 self.cross_entropy = nn.CrossEntropyLoss() def compute_contrastive_loss(self, anchor_emb, pos_emb, neg_emb): """ 计算对比损失 (InfoNCE + Triplet Margin组合) """ # 计算余弦相似度 pos_sim = F.cosine_similarity(anchor_emb, pos_emb) neg_sim = F.cosine_similarity(anchor_emb, neg_emb) # InfoNCE损失 numerator = torch.exp(pos_sim / self.temperature) denominator = numerator + torch.exp(neg_sim / self.temperature) info_nce_loss = -torch.log(numerator / denominator).mean() # 三元组损失 triplet_loss = F.triplet_margin_loss( anchor_emb, pos_emb, neg_emb, margin=self.margin ) # 加权组合 return info_nce_loss + triplet_loss def get_sequence_representation(self, outputs, attention_mask): """ 获取序列表示(取最后一个token的隐藏状态) """ # 获取指定层的隐藏状态 hidden_states = outputs.hidden_states[self.repr_layer] # 获取每个序列的最后一个非填充token # 注意:attention_mask中1表示有效token,0表示填充 last_token_indices = attention_mask.sum(dim=1) - 1 # 收集每个序列的最后一个token的隐藏状态 batch_size = hidden_states.size(0) sequence_representations = hidden_states[ torch.arange(batch_size), last_token_indices ] return sequence_representations def compute_loss(self, model, inputs, return_outputs=False): """ 计算总损失(语言建模损失 + 对比损失) """ # 提取输入 anchor_input_ids = inputs.get("anchor_input_ids") anchor_attention_mask = inputs.get("anchor_attention_mask") positive_input_ids = inputs.get("positive_input_ids") positive_attention_mask = inputs.get("positive_attention_mask") negative_input_ids = inputs.get("negative_input_ids") negative_attention_mask = inputs.get("negative_attention_mask") # 前向传播获取隐藏状态 def get_embeddings(input_ids, attention_mask): outputs = model( input_ids=input_ids, attention_mask=attention_mask, output_hidden_states=True, return_dict=True ) return self.get_sequence_representation(outputs, attention_mask) # 获取三元组的嵌入表示 anchor_emb = get_embeddings(anchor_input_ids, anchor_attention_mask) pos_emb = get_embeddings(positive_input_ids, positive_attention_mask) neg_emb = get_embeddings(negative_input_ids, negative_attention_mask) # 计算对比损失 cl_loss = self.compute_contrastive_loss(anchor_emb, pos_emb, neg_emb) cl_loss = cl_loss * self.contrastive_weight # 计算语言建模损失(仅针对positive回复) lm_labels = positive_input_ids.clone() lm_labels[lm_labels == self.tokenizer.pad_token_id] = -100 # 忽略填充token lm_outputs = model( input_ids=positive_input_ids, attention_mask=positive_attention_mask, labels=lm_labels ) lm_loss = lm_outputs.loss # 总损失 = LM损失 + 对比损失 total_loss = lm_loss + cl_loss # 如果返回输出,则返回损失和输出 if return_outputs: outputs = { "lm_loss": lm_loss, "cl_loss": cl_loss, "total_loss": total_loss, "logits": lm_outputs.logits } return total_loss, outputs return total_loss def training_step(self, model, inputs): """ 自定义训练步骤 """ model.train() inputs = self._prepare_inputs(inputs) # 前向传播 with self.compute_loss_context_manager(): loss, outputs = self.compute_loss(model, inputs, return_outputs=True) # 如果使用梯度累积,需要除以累积步数 if self.args.gradient_accumulation_steps > 1: loss = loss / self.args.gradient_accumulation_steps # 反向传播 loss.backward() # 记录日志 self.log({ "train/lm_loss": outputs["lm_loss"].item(), "train/cl_loss": outputs["cl_loss"].item(), "train/loss": loss.item(), "train/lr": self.lr_scheduler.get_last_lr()[0] }) return loss.detach() def log(self, logs: Dict[str, float]): """ 自定义日志记录 """ if self.state.epoch is not None: logs["epoch"] = round(self.state.epoch, 2) # 每N步记录一次日志 if self.state.global_step % self.args.logging_steps == 0: logger.info(f"Step {self.state.global_step}: {logs}") def train(self, **kwargs): """ 自定义训练循环 """ # 初始化训练 self._prepare_for_training() # 训练循环 for epoch in range(int(self.args.num_train_epochs)): logger.info(f"Starting epoch {epoch + 1}/{self.args.num_train_epochs}") # 创建数据加载器 train_dataloader = self.get_train_dataloader() # 训练一个epoch for step, inputs in enumerate(tqdm(train_dataloader, desc=f"Epoch {epoch+1}")): # 训练步骤 loss = self.training_step(self.model, inputs) # 梯度更新 if (step + 1) % self.args.gradient_accumulation_steps == 0: self.optimizer.step() self.lr_scheduler.step() self.optimizer.zero_grad() self.state.global_step += 1 # 模型保存 if self.args.save_strategy == "steps" and self.state.global_step % self.args.save_steps == 0: self._save_model(self.args.output_dir) # 每个epoch结束时保存模型 if self.args.save_strategy == "epoch": self._save_model(self.args.output_dir) def _save_model(self, output_dir: str): """ 保存模型适配器 """ save_path = os.path.join(output_dir, f"checkpoint-{self.state.global_step}") os.makedirs(save_path, exist_ok=True) # 保存适配器权重 self.model.save_pretrained(save_path) self.tokenizer.save_pretrained(save_path) logger.info(f"Model saved to {save_path}") def evaluate(self, **kwargs): """ 自定义评估方法 """ self.model.eval() eval_dataloader = self.get_eval_dataloader() total_loss = 0.0 total_lm_loss = 0.0 total_cl_loss = 0.0 num_batches = 0 with torch.no_grad(): for inputs in tqdm(eval_dataloader, desc="Evaluating"): inputs = self._prepare_inputs(inputs) loss, outputs = self.compute_loss(self.model, inputs, return_outputs=True) total_loss += loss.item() total_lm_loss += outputs["lm_loss"].item() total_cl_loss += outputs["cl_loss"].item() num_batches += 1 avg_loss = total_loss / num_batches avg_lm_loss = total_lm_loss / num_batches avg_cl_loss = total_cl_loss / num_batches metrics = { "eval_loss": avg_loss, "eval_lm_loss": avg_lm_loss, "eval_cl_loss": avg_cl_loss } logger.info(f"Evaluation results: {metrics}") return metrics
07-20
评论 8
成就一亿技术人!
拼手气红包6.0元
还能输入1000个字符
 
红包 添加红包
表情包 插入表情
 条评论被折叠 查看
添加红包

请填写红包祝福语或标题

红包个数最小为10个

红包金额最低5元

当前余额3.43前往充值 >
需支付:10.00
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包

打赏作者

LaoYuanPython

你的鼓励将是我创作的最大动力

¥1 ¥2 ¥4 ¥6 ¥10 ¥20
扫码支付:¥1
获取中
扫码支付

您的余额不足,请更换扫码支付或充值

打赏作者

实付
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明:

1.余额是钱包充值的虚拟货币,按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载,可以购买VIP、付费专栏及课程。

余额充值