P-Tuning提示词微调

P-Tuning： 在Prompt-Tuning的基础上，对Prompt部分进行进一步的编码计算，加速收敛。具体来说，PEFT中支持两种编码方式，一种是LSTM，一种是MLP。与Prompt-Tuning不同的是，Prompt的形式只有Soft Prompt。

思路： P-Tuning 引入了一个轻量级的编码器（如 LSTM 或 MLP）来动态生成提示嵌入。编码器的输入通常是任务相关的特征（例如输入序列的上下文信息），输出则是虚拟 token 的连续表示。

优点： Prompt-Tuning 是静态的，直接随机将一组向量与用户输入相加；而P-Tuning，引入了 LSTM、MLP，来动态嵌入这些提示，使得这些提示跟输入的上下文相关联。

from datasets import load_from_disk
from transformers import AutoTokenizer, AutoModelForCausalLM, DataCollatorForSeq2Seq
from transformers import TrainingArguments, Trainer
from peft import PromptEncoderConfig, TaskType, get_peft_model, PromptEncoderReparameterizationType
 
# 分词器
tokenizer = AutoTokenizer.from_pretrained("Langboat/bloom-1b4-zh")
 
# 函数内将instruction和response拆开分词的原因是：
# 为了便于mask掉不需要计算损失的labels, 即代码labels = [-100] * len(instruction["input_ids"]) + response["input_ids"]
def process_func(example):
    MAX_LENGTH = 256
    input_ids, attention_mask, labels = [], [], []
    instruction = tokenizer("\n".join(["Human: " + example["instruction"], example["input"]]).strip() + "\n\nAssistant: ")
    response = tokenizer(example["output"] + tokenizer.eos_token)
    input_ids = instruction["input_ids"] + response["input_ids"]
    attention_mask = instruction["attention_mask"] + response["attention_mask"]
    labels = [-100] * len(instruction["input_ids"]) + response["input_ids"]
    if len(input_ids) > MAX_LENGTH:
        input_ids = input_ids[:MAX_LENGTH]
        attention_mask = attention_mask[:MAX_LENGTH]
        labels = labels[:MAX_LENGTH]
    return {
        "input_ids": input_ids,
        "attention_mask": attention_mask,
        "labels": labels
    }
 
if __name__ == "__main__":
    # 加载数据集
    dataset = load_from_disk("/root/StudyLLM/prompt/03-PEFT/data/alpaca_data_zh")
    
    # 处理数据
    tokenized_ds = dataset.map(process_func, remove_columns = dataset.column_names)
    # print(tokenizer.decode(tokenized_ds[1]["input_ids"]))
    # print(tokenizer.decode(list(filter(lambda x: x != -100, tokenized_ds[1]["labels"]))))
    
    # 创建模型
    model = AutoModelForCausalLM.from_pretrained("Langboat/bloom-1b4-zh", low_cpu_mem_usage=True)
    
    # 设置 P-Tuning
    # 使用 MLP
    config = PromptEncoderConfig(task_type=TaskType.CAUSAL_LM, num_virtual_tokens=10,
                             encoder_reparameterization_type=PromptEncoderReparameterizationType.MLP,
                             encoder_hidden_size=1024)
    # 使用LSTM
    config = PromptEncoderConfig(task_type=TaskType.CAUSAL_LM, num_virtual_tokens=10,
                             encoder_reparameterization_type=PromptEncoderReparameterizationType.LSTM,
                             encoder_dropout=0.1, encoder_num_layers=1, encoder_hidden_size=1024)
 
    model = get_peft_model(model, config) # 生成P-Tuning对应的model
    print(model.print_trainable_parameters())
    
    # 训练参数
    args = TrainingArguments(
        output_dir = "/tmp_1203",
        per_device_train_batch_size = 1,
        gradient_accumulation_steps = 8,
        logging_steps = 10,
        num_train_epochs = 1
    )
    
    # trainer
    trainer = Trainer(
        model = model,
        args = args,
        train_dataset = tokenized_ds,
        data_collator = DataCollatorForSeq2Seq(tokenizer = tokenizer, padding = True)
    )
    
    # 训练模型
    trainer.train()
    
    # 模型推理
    model = model.cuda()
    ipt = tokenizer("Human: {}\n{}".format("考试有哪些技巧？", "").strip() + "\n\nAssistant: ", return_tensors="pt").to(model.device)
    print(tokenizer.decode(model.generate(**ipt, max_length=128, do_sample=True)[0], skip_special_tokens=True))