诗语灵伴——基于PaddleNLP构建的古诗对话数字人

Bash%cd ~!unzip PaddleNLP-develop.zip!pip install ./PaddleNLP-develop# !git clone https://github.com/PaddlePaddle/PaddleNLP.git
%cd ~!unzip PaddleSlim-develop.zip!cd ~/PaddleSlim-develop/csrc && python ./setup_cuda.py install && cd ~/PaddleSlim-develop && pip install .# !git clone https://github.com/PaddlePaddle/PaddleSlim.git

Python# 安装ERNIE Bot和tiktoken!pip install erniebot tiktoken

Plain Text本次使用的数据集都放在work文件夹里了，work文件夹目录如下：Poem ----> 放置了训练集和验证集。Poem_soul_Data.json ----> 源未处理格式数据集。lora_argument.json ----> Lora微调配置文件。poem_soul.json ----> 符合格式要求的数据集。

JSON{"conversation": [{"system": "你是一个专业的古诗歌专家，你知道很多古诗。用户报上关键词后，你可以把包含关键词的古诗告诉用户","input": "根据绿苔这个关键词写一首古诗","output": "生成的古诗为:\n崖悬百尺古,\n面削一屏开。\n晴日流丹草,\n春风长绿苔。"}]}

JSON{"src": ["根据绿苔这个关键词写一首古诗"], "tgt": ["生成的古诗为:\n崖悬百尺古,\n面削一屏开。\n晴日流丹草,\n春风长绿苔。"],"context": {"system": "你是一个专业的古诗歌专家，你知道很多古诗。用户报上关键词后，你可以把包含关键词的古诗告诉用户"}}

Python# 格式转换import jsonPoem_soul_Data = []with open('/home/aistudio/work/Poem_soul_Data.json', 'r') as f:data = json.load(f)
for data_item in data:src = []tgt = [] for j in data_item['conversation']:src.append(j['input'])tgt.append(j['output'])print(j['input'])print(j['output'])assert len(src) == len(tgt)Poem_soul_Data.append({'src':src, 'tgt':tgt, "context": {"system": f"{data_item['conversation'][0]['system']}"}})
with open('/home/aistudio/work/poem_soul.json', 'w', encoding='utf-8') as f:for item in Poem_soul_Data:json.dump(item, f, ensure_ascii=False)f.write('\n')

Python%cd ~
# 对古诗进行8:2数据划分import jsonimport randomimport os
# 定义文件路径file_path = 'work/Poem/'input_files = ['/home/aistudio/work/poem_soul.json']train_output_file = os.path.join(file_path,'train.json')dev_output_file = os.path.join(file_path,'dev.json')

directory = os.path.dirname(file_path)if not os.path.exists(directory):os.makedirs(directory)print(f'目录 {directory} 已创建')else:print(f'目录 {directory} 已存在')
train_data = []dev_data = []
for json_file in input_files:with open(json_file, 'r', encoding='utf-8') as f:lines = f.readlines()data = [json.loads(line) for line in lines]random.shuffle(data)split_index = int(0.8 * len(data))train_data += data[:split_index]dev_data += data[split_index:]
random.shuffle(train_data)random.shuffle(dev_data)

# 将数据保存到train.json和dev.jsonwith open(train_output_file, 'w', encoding='utf-8') as f:for item in train_data:f.write(json.dumps(item, ensure_ascii=False) + '\n')
with open(dev_output_file, 'w', encoding='utf-8') as f:for item in dev_data:f.write(json.dumps(item, ensure_ascii=False) + '\n')
print(f"训练数据保存在 {train_output_file}")print(f"测试数据保存在 {dev_output_file}")

Python#查看lora配置文件 记得修改配置文件中的"dataset_name_or_path"%cat ~/work/lora_argument.json

Bash# 修改配置文件中的"dataset_name_or_path"，运行下述代码直接将其相应配置写入。%%writefile ~/work/lora_argument.json {"model_name_or_path": "THUDM/chatglm2-6b","dataset_name_or_path": "/home/aistudio/work/Poem","output_dir": "./checkpoints/lora_ckpts","per_device_train_batch_size": 4,"gradient_accumulation_steps": 4,"per_device_eval_batch_size": 8,"eval_accumulation_steps":16,"num_train_epochs": 3,"learning_rate": 3e-04,"warmup_steps": 30,"logging_steps": 1,"evaluation_strategy": "epoch","save_strategy": "epoch","src_length": 1024,"max_length": 2048,"fp16": true,"fp16_opt_level": "O2","do_train": true,"do_eval": true,"disable_tqdm": false,"load_best_model_at_end": true,"eval_with_do_generation": false,"metric_for_best_model": "accuracy","recompute": true,"save_total_limit": 1,"tensor_parallel_degree": 1,"pipeline_parallel_degree": 1,"lora": true,"zero_padding": false,"unified_checkpoint": true,"use_flash_attention": false}

Python%cd /home/aistudio/PaddleNLP-develop/llm!python run_finetune.py /home/aistudio/work/lora_argument.json

Python%cd ~/PaddleNLP-develop/llm/tools
!python merge_lora_params.py \--model_name_or_path THUDM/chatglm2-6b \--lora_path /home/aistudio/PaddleNLP-develop/llm/checkpoints/lora_ckpts \--output_path ~/data/checkpoints/chatglm2_lora_merge \--device "gpu" \--safe_serialization True
# 脚本参数介绍# lora_path: Lora参数和配置路径，对Lora参数进行初始化，默认为None。# model_name_or_path: 必须，主干模型参数路径，默认为None。# merge_model_path: 必须，合并参数后保存路径，默认为None。# device: 运行环境，默认为gpu。# safe_serialization: 是否保存为safetensor格式，默认为True。

Python!pip uninstall paddlepaddle-gpu!python -m pip install paddlepaddle-gpu==0.0.0.post118 -f https://www.paddlepaddle.org.cn/whl/linux/gpu/develop.html

Python# 动态图推理%cd ~/PaddleNLP-develop/llm
!cp /home/aistudio/PaddleNLP-develop/llm/predict/predictor.py /home/aistudio/PaddleNLP-develop/llm
!python predictor.py \--model_name_or_path THUDM/chatglm2-6b \ # 此处放置lora_argument.json配置文件中的model_name_or_path--lora_path /home/aistudio/PaddleNLP-develop/llm/checkpoints/lora_ckpts/checkpoint-11250 \--data_file /home/aistudio/work/Poem/dev.json \--dtype float16

Python# 静态图导出%cd ~/PaddleNLP-develop/llm!python export_model.py \--model_name_or_path THUDM/chatglm2-6b \--output_path /home/aistudio/data/static_inference_model_chatglm2 \--lora_path /home/aistudio/PaddleNLP-develop/llm/checkpoints/lora_ckpts/checkpoint-11250 \--dtype float16

Python# 静态图推理%cd ~/PaddleNLP-develop/llm
!python predictor.py \--model_name_or_path /home/aistudio/data/static_inference_model_chatglm2 \--data_file ~/work/Poem/dev.json \--dtype float16 \--mode static

Python[2024-08-06 12:56:51,087] [INFO] - Start predict[2024-08-06 12:56:51,087] [ WARNING] - The last conversation is not a single-round, chat-template will skip the conversation: ('生成的古诗为:\n星榆叶叶昼离披,\n云粉千重凝不飞。\n昆玉楼台珠树密,\n夜来谁向月中归。',)[2024-08-06 12:56:56,471] [INFO] - End predict***********Source**********[('根据楼台这个关键词写一首古诗', '生成的古诗为:\n星榆叶叶昼离披,\n云粉千重凝不飞。\n昆玉楼台珠树密,\n夜来谁向月中归。')]***********Target*********************Output**********生成的古诗为:寂寂随缘止,年年归去来。山阴湖上雪,疑是旧楼台。
[2024-08-06 12:46:38,954] [INFO] - Start predict[2024-08-06 12:46:38,954] [ WARNING] - The last conversation is not a single-round, chat-template will skip the conversation: ('生成的古诗为:\n只说梅花似雪飞,\n朱颜谁信暗香随。\n不须添上徐熙画,\n付与西湖别赋诗。',)[2024-08-06 12:46:45,444] [INFO] - End predict***********Source**********[('根据朱颜这个关键词写一首古诗', '生成的古诗为:\n只说梅花似雪飞,\n朱颜谁信暗香随。\n不须添上徐熙画,\n付与西湖别赋诗。')]***********Target**********
***********Output**********生成的古诗为:一枝半落半还开,病酒消愁只自杯。莫怪朱颜消得尽,暗香疏影两三栽。

Python#-*- coding:utf-8 -*-from paddlenlp.transformers import AutoModelForCausalLM, AutoTokenizer
class Hutaos(object):def __init__(self):# load modelself.model_path = r"/home/aistudio/PaddleNLP/llm/checkpoints/llama_lora_merge"# "/root/gushi2/merged"self.model = AutoModelForCausalLM.from_pretrained(self.model_path)self.tokenizer = AutoTokenizer.from_pretrained(self.model_path)def answer(self,questions):# query = input() #"根据明月这个关键词给我生成一个古诗"input_features = self.tokenizer("{}".format(questions), return_tensors="pd")outputs = self.model.generate(**input_features, max_length=128)response = tokenizer.batch_decode(outputs[0])return response
llms = Hutaos()def get_answer(question):answer = llms.answer(question)return answer
if __name__ == "__main__":print(get_answer("根据孤影这个关键字给我生成一首古诗"))