0、IA-3基本原理

 IA3的基本思想就是抑制和放大内部激活，通过可学习的向量对激活值进行抑制或放大。具体来说，会对K、VFFN三部分的值进行调整，训练过程中同样冻结原始模型的权重，只更新可学习的部分向量部分。训练完成后，与Lora类似，也可以将学习部分的参数与原始权重合并，没有额外推理开销。

1、IA-3微调实战

1.1、导包

from datasets import Dataset
from transformers import AutoTokenizer, AutoModelForCausalLM, DataCollatorForSeq2Seq, TrainingArguments, Trainer

1.2、加载数据集

ds = Dataset.load_from_disk("../Data/alpaca_data_zh/")
ds

1.3、数据预处理

tokenizer = AutoTokenizer.from_pretrained("../Model/bloom-389m-zh")
tokenizer

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
    }

tokenized_ds = ds.map(process_func, remove_columns=ds.column_names)
tokenized_ds

1.4、创建模型

model = AutoModelForCausalLM.from_pretrained("../Model/bloom-389m-zh", low_cpu_mem_usage=True)

1.5、LORA微调

1.5.1、配置文件

from peft import IA3Config, TaskType, get_peft_model

config = IA3Config(task_type=TaskType.CAUSAL_LM)
config

1.5.2、创建模型

model = get_peft_model(model, config)

model.print_trainable_parameters()

1.6、配置训练参数

args = TrainingArguments(
    output_dir="./chatbot",
    per_device_train_batch_size=1,
    gradient_accumulation_steps=8,
    logging_steps=10,
    num_train_epochs=1,
    learning_rate=3e-3
)

1.7、创建训练器

trainer = Trainer(
    model=model,
    args=args,
    train_dataset=tokenized_ds,
    data_collator=DataCollatorForSeq2Seq(tokenizer=tokenizer, padding=True),
)

1.8、模型训练

trainer.train()

1.9、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)
tokenizer.decode(model.generate(**ipt, max_length=128, do_sample=True)[0], skip_special_tokens=True)