本节我们简要介绍基于 transformers、peft 等框架,使用 Qwen2-VL-2B-Instruct 模型在COCO2014图像描述 上进行Lora微调训练,同时使用 SwanLab 监控训练过程与评估模型效果。
Lora 是一种高效微调方法,深入了解其原理可参见博客:知乎|深入浅出 Lora。
训练过程:Qwen2-VL-finetune
SwanLab 是一个开源的模型训练记录工具,常被称为"中国版 Weights&Biases + Tensorboard"。SwanLab面向AI研究者,提供了训练可视化、自动日志记录、超参数记录、实验对比、多人协同等功能。在SwanLab上,研究者能基于直观的可视化图表发现训练问题,对比多个实验找到研究灵感,并通过在线链接的分享与基于组织的多人协同训练,打破团队沟通的壁垒。
为什么要记录训练?
相较于软件开发,模型训练更像一个实验科学。一个品质优秀的模型背后,往往是成千上万次实验。研究者需要不断尝试、记录、对比,积累经验,才能找到最佳的模型结构、超参数与数据配比。在这之中,如何高效进行记录与对比,对于研究效率的提升至关重要。
可视化的价值在哪里?
机器学习模型训练往往伴随着大量的超参数、指标、日志等数据,很多关键信息往往存在于实验的中间而非结尾,如果不对连续的指标通过图表进行可视化,往往会错失发现问题的最佳时机,甚至错过关键信息。同时不进行可视化,也难以对比多个实验之间的差异。 可视化也为AI研究者提供了良好的交流基础,研究者们可以基于图表进行沟通、分析与优化,而非以往看着枯燥的终端打印。这打破了团队沟通的壁垒,提高了整体的研发效率。
环境配置分为三步:
-
确保你的电脑上至少有一张英伟达显卡,并已安装好了CUDA环境。
-
安装Python(版本>=3.8)以及能够调用CUDA加速的PyTorch。
-
安装Qwen2-VL微调相关的第三方库,可以使用以下命令:
python -m pip install --upgrade pip
# 更换 pypi 源加速库的安装
pip config set global.index-url https://pypi.tuna.tsinghua.edu.cn/simple
pip install modelscope==1.18.0
pip install transformers==4.46.2
pip install sentencepiece==0.2.0
pip install accelerate==1.1.1
pip install datasets==2.18.0
pip install peft==0.13.2
pip install swanlab==0.3.27
pip install qwen-vl-utils==0.0.8
本节使用的是 coco_2014_caption 数据集(中的500张图),该数据集主要用于多模态(Image-to-Text)任务。
数据集介绍:COCO 2014 Caption数据集是Microsoft Common Objects in Context (COCO)数据集的一部分,主要用于图像描述任务。该数据集包含了大约40万张图像,每张图像都有至少1个人工生成的英文描述语句。这些描述语句旨在帮助计算机理解图像内容,并为图像自动生成描述提供训练数据。
在本节的任务中,我们主要使用其中的前500张图像,并对它进行处理和格式调整,目标是组合成如下格式的json文件:
[
{
"id": "identity_1",
"conversations": [
{
"from": "user",
"value": "COCO Yes: <|vision_start|>图像文件路径<|vision_end|>"
},
{
"from": "assistant",
"value": "A snow skier assessing the mountain before starting to sky"
}
]
},
...
]
其中,"from"是角色(user代表人类,assistant代表模型),"value"是聊天的内容,其中<|vision_start|>
和<|vision_end|>
是Qwen2-VL模型识别图像的标记,中间可以放图像的文件路径,也可以是URL。
数据集下载与处理方式
-
我们需要做四件事情:
- 通过Modelscope下载coco_2014_caption数据集
- 加载数据集,将图像保存到本地
- 将图像路径和描述文本转换为一个csv文件
- 将csv文件转换为json文件
-
使用下面的代码完成从数据下载到生成csv的过程:
data2csv.py:
# 导入所需的库
from modelscope.msdatasets import MsDataset
import os
import pandas as pd
MAX_DATA_NUMBER = 500
# 检查目录是否已存在
if not os.path.exists('coco_2014_caption'):
# 从modelscope下载COCO 2014图像描述数据集
ds = MsDataset.load('modelscope/coco_2014_caption', subset_name='coco_2014_caption', split='train')
print(len(ds))
# 设置处理的图片数量上限
total = min(MAX_DATA_NUMBER, len(ds))
# 创建保存图片的目录
os.makedirs('coco_2014_caption', exist_ok=True)
# 初始化存储图片路径和描述的列表
image_paths = []
captions = []
for i in range(total):
# 获取每个样本的信息
item = ds[i]
image_id = item['image_id']
caption = item['caption']
image = item['image']
# 保存图片并记录路径
image_path = os.path.abspath(f'coco_2014_caption/{image_id}.jpg')
image.save(image_path)
# 将路径和描述添加到列表中
image_paths.append(image_path)
captions.append(caption)
# 每处理50张图片打印一次进度
if (i + 1) % 50 == 0:
print(f'Processing {i+1}/{total} images ({(i+1)/total*100:.1f}%)')
# 将图片路径和描述保存为CSV文件
df = pd.DataFrame({
'image_path': image_paths,
'caption': captions
})
# 将数据保存为CSV文件
df.to_csv('./coco-2024-dataset.csv', index=False)
print(f'数据处理完成,共处理了{total}张图片')
else:
print('coco_2014_caption目录已存在,跳过数据处理步骤')
3. 在同一目录下,用以下代码,将csv文件转换为json文件:
csv2json.py:
import pandas as pd
import json
# 载入CSV文件
df = pd.read_csv('./coco-2024-dataset.csv')
conversations = []
# 添加对话数据
for i in range(len(df)):
conversations.append({
"id": f"identity_{i+1}",
"conversations": [
{
"from": "user",
"value": f"COCO Yes: <|vision_start|>{df.iloc[i]['image_path']}<|vision_end|>"
},
{
"from": "assistant",
"value": df.iloc[i]['caption']
}
]
})
# 保存为Json
with open('data_vl.json', 'w', encoding='utf-8') as f:
json.dump(conversations, f, ensure_ascii=False, indent=2)
此时目录下会多出两个文件:
- coco-2024-dataset.csv
- data_vl.json
至此,我们完成了数据集的准备。
这里我们使用modelscope下载Qwen2-VL-2B-Instruct模型,然后把它加载到Transformers中进行训练:
from modelscope import snapshot_download, AutoTokenizer
from transformers import TrainingArguments, Trainer, DataCollatorForSeq2Seq, Qwen2VLForConditionalGeneration, AutoProcessor
import torch
# 在modelscope上下载Qwen2-VL模型到本地目录下
model_dir = snapshot_download("Qwen/Qwen2-VL-2B-Instruct", cache_dir="./", revision="master")
# 使用Transformers加载模型权重
tokenizer = AutoTokenizer.from_pretrained("./Qwen/Qwen2-VL-2B-Instruct/", use_fast=False, trust_remote_code=True)
# 特别的,Qwen2-VL-2B-Instruct模型需要使用Qwen2VLForConditionalGeneration来加载
model = Qwen2VLForConditionalGeneration.from_pretrained("./Qwen/Qwen2-VL-2B-Instruct/", device_map="auto", torch_dtype=torch.bfloat16, trust_remote_code=True,)
model.enable_input_require_grads() # 开启梯度检查点时,要执行该方法
模型大小为 4.5GB,下载模型大概需要 5 分钟。
SwanLab与Transformers已经做好了集成,用法是在Trainer的callbacks
参数中添加SwanLabCallback
实例,就可以自动记录超参数和训练指标,简化代码如下:
from swanlab.integration.transformers import SwanLabCallback
from transformers import Trainer
swanlab_callback = SwanLabCallback()
trainer = Trainer(
...
callbacks=[swanlab_callback],
)
首次使用SwanLab,需要先在官网注册一个账号,然后在用户设置页面复制你的API Key,然后在训练开始提示登录时粘贴即可,后续无需再次登录:
更多用法可参考快速开始、Transformers集成。
查看可视化训练过程:Qwen2-VL-finetune
本节代码做了以下几件事:
- 下载并加载Qwen2-VL-2B-Instruct模型
- 加载数据集,取前496条数据参与训练,4条数据进行主观评测
- 配置Lora,参数为r=64, lora_alpha=16, lora_dropout=0.05
- 使用SwanLab记录训练过程,包括超参数、指标和最终的模型输出结果
- 训练2个epoch
开始执行代码时的目录结构应该是:
|———— train.py
|———— coco_2014_caption
|———— coco-2024-dataset.csv
|———— data_vl.json
|———— data2csv.py
|———— csv2json.py
完整代码如下
train.py:
import torch
from datasets import Dataset
from modelscope import snapshot_download, AutoTokenizer
from swanlab.integration.transformers import SwanLabCallback
from qwen_vl_utils import process_vision_info
from peft import LoraConfig, TaskType, get_peft_model, PeftModel
from transformers import (
TrainingArguments,
Trainer,
DataCollatorForSeq2Seq,
Qwen2VLForConditionalGeneration,
AutoProcessor,
)
import swanlab
import json
def process_func(example):
"""
将数据集进行预处理
"""
MAX_LENGTH = 8192
input_ids, attention_mask, labels = [], [], []
conversation = example["conversations"]
input_content = conversation[0]["value"]
output_content = conversation[1]["value"]
file_path = input_content.split("<|vision_start|>")[1].split("<|vision_end|>")[0] # 获取图像路径
messages = [
{
"role": "user",
"content": [
{
"type": "image",
"image": f"{file_path}",
"resized_height": 280,
"resized_width": 280,
},
{"type": "text", "text": "COCO Yes:"},
],
}
]
text = processor.apply_chat_template(
messages, tokenize=False, add_generation_prompt=True
) # 获取文本
image_inputs, video_inputs = process_vision_info(messages) # 获取数据数据(预处理过)
inputs = processor(
text=[text],
images=image_inputs,
videos=video_inputs,
padding=True,
return_tensors="pt",
)
inputs = {key: value.tolist() for key, value in inputs.items()} #tensor -> list,为了方便拼接
instruction = inputs
response = tokenizer(f"{output_content}", add_special_tokens=False)
input_ids = (
instruction["input_ids"][0] + response["input_ids"] + [tokenizer.pad_token_id]
)
attention_mask = instruction["attention_mask"][0] + response["attention_mask"] + [1]
labels = (
[-100] * len(instruction["input_ids"][0])
+ response["input_ids"]
+ [tokenizer.pad_token_id]
)
if len(input_ids) > MAX_LENGTH: # 做一个截断
input_ids = input_ids[:MAX_LENGTH]
attention_mask = attention_mask[:MAX_LENGTH]
labels = labels[:MAX_LENGTH]
input_ids = torch.tensor(input_ids)
attention_mask = torch.tensor(attention_mask)
labels = torch.tensor(labels)
inputs['pixel_values'] = torch.tensor(inputs['pixel_values'])
inputs['image_grid_thw'] = torch.tensor(inputs['image_grid_thw']).squeeze(0) #由(1,h,w)变换为(h,w)
return {"input_ids": input_ids, "attention_mask": attention_mask, "labels": labels,
"pixel_values": inputs['pixel_values'], "image_grid_thw": inputs['image_grid_thw']}
def predict(messages, model):
# 准备推理
text = processor.apply_chat_template(
messages, tokenize=False, add_generation_prompt=True
)
image_inputs, video_inputs = process_vision_info(messages)
inputs = processor(
text=[text],
images=image_inputs,
videos=video_inputs,
padding=True,
return_tensors="pt",
)
inputs = inputs.to("cuda")
# 生成输出
generated_ids = model.generate(**inputs, max_new_tokens=128)
generated_ids_trimmed = [
out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
]
output_text = processor.batch_decode(
generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
)
return output_text[0]
# 在modelscope上下载Qwen2-VL模型到本地目录下
model_dir = snapshot_download("Qwen/Qwen2-VL-2B-Instruct", cache_dir="./", revision="master")
# 使用Transformers加载模型权重
tokenizer = AutoTokenizer.from_pretrained("./Qwen/Qwen2-VL-2B-Instruct/", use_fast=False, trust_remote_code=True)
processor = AutoProcessor.from_pretrained("./Qwen/Qwen2-VL-2B-Instruct")
model = Qwen2VLForConditionalGeneration.from_pretrained("./Qwen/Qwen2-VL-2B-Instruct/", device_map="auto", torch_dtype=torch.bfloat16, trust_remote_code=True,)
model.enable_input_require_grads() # 开启梯度检查点时,要执行该方法
# 处理数据集:读取json文件
# 拆分成训练集和测试集,保存为data_vl_train.json和data_vl_test.json
train_json_path = "data_vl.json"
with open(train_json_path, 'r') as f:
data = json.load(f)
train_data = data[:-4]
test_data = data[-4:]
with open("data_vl_train.json", "w") as f:
json.dump(train_data, f)
with open("data_vl_test.json", "w") as f:
json.dump(test_data, f)
train_ds = Dataset.from_json("data_vl_train.json")
train_dataset = train_ds.map(process_func)
# 配置LoRA
config = LoraConfig(
task_type=TaskType.CAUSAL_LM,
target_modules=["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"],
inference_mode=False, # 训练模式
r=64, # Lora 秩
lora_alpha=16, # Lora alaph,具体作用参见 Lora 原理
lora_dropout=0.05, # Dropout 比例
bias="none",
)
# 获取LoRA模型
peft_model = get_peft_model(model, config)
# 配置训练参数
args = TrainingArguments(
output_dir="./output/Qwen2-VL-2B",
per_device_train_batch_size=4,
gradient_accumulation_steps=4,
logging_steps=10,
logging_first_step=5,
num_train_epochs=2,
save_steps=100,
learning_rate=1e-4,
save_on_each_node=True,
gradient_checkpointing=True,
report_to="none",
)
# 设置SwanLab回调
swanlab_callback = SwanLabCallback(
project="Qwen2-VL-finetune",
experiment_name="qwen2-vl-coco2014",
config={
"model": "https://modelscope.cn/models/Qwen/Qwen2-VL-2B-Instruct",
"dataset": "https://modelscope.cn/datasets/modelscope/coco_2014_caption/quickstart",
"github": "https://github.com/datawhalechina/self-llm",
"prompt": "COCO Yes: ",
"train_data_number": len(train_data),
"lora_rank": 64,
"lora_alpha": 16,
"lora_dropout": 0.1,
},
)
# 配置Trainer
trainer = Trainer(
model=peft_model,
args=args,
train_dataset=train_dataset,
data_collator=DataCollatorForSeq2Seq(tokenizer=tokenizer, padding=True),
callbacks=[swanlab_callback],
)
# 开启模型训练
trainer.train()
# ====================测试模式===================
# 配置测试参数
val_config = LoraConfig(
task_type=TaskType.CAUSAL_LM,
target_modules=["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"],
inference_mode=True, # 训练模式
r=64, # Lora 秩
lora_alpha=16, # Lora alaph,具体作用参见 Lora 原理
lora_dropout=0.05, # Dropout 比例
bias="none",
)
# 获取测试模型
val_peft_model = PeftModel.from_pretrained(model, model_id="./output/Qwen2-VL-2B/checkpoint-62", config=val_config)
# 读取测试数据
with open("data_vl_test.json", "r") as f:
test_dataset = json.load(f)
test_image_list = []
for item in test_dataset:
input_image_prompt = item["conversations"][0]["value"]
# 去掉前后的<|vision_start|>和<|vision_end|>
origin_image_path = input_image_prompt.split("<|vision_start|>")[1].split("<|vision_end|>")[0]
messages = [{
"role": "user",
"content": [
{
"type": "image",
"image": origin_image_path
},
{
"type": "text",
"text": "COCO Yes:"
}
]}]
response = predict(messages, val_peft_model)
messages.append({"role": "assistant", "content": f"{response}"})
print(messages[-1])
test_image_list.append(swanlab.Image(origin_image_path, caption=response))
swanlab.log({"Prediction": test_image_list})
# 在Jupyter Notebook中运行时要停止SwanLab记录,需要调用swanlab.finish()
swanlab.finish()
看到下面的进度条即代表训练开始:
从SwanLab图表中我们可以看到,lr的下降策略是线性下降,loss随epoch呈现下降趋势,而grad_norm则在上升。这种形态往往反映了模型有过拟合的风险,训练不要超过2个epoch。
在Prediction
图表中记录着模型最终的输出结果,可以看到模型在回答的风格上是用的COCO数据集的简短英文风格进行的描述:
而同样的图像,没有被微调的模型输出结果如下:
1-没有微调:The image depicts a cozy living room with a rocking chair in the center, a bookshelf filled with books, and a table with a vase and a few other items. The walls are decorated with wallpaper, and there are curtains on the windows. The room appears to be well-lit, with sunlight streaming in from the windows.
1-微调后:A living room with a rocking chair, a bookshelf, and a table with a vase and a bowl.
2-没有微调:It looks like a family gathering or a party in a living room. There are several people sitting around a dining table, eating pizza. The room has a cozy and warm atmosphere.
2-微调后:A group of people sitting around a dining table eating pizza.
可以明显看到微调后风格的变化。
加载lora微调后的模型,并进行推理:
from transformers import Qwen2VLForConditionalGeneration, AutoProcessor
from qwen_vl_utils import process_vision_info
from peft import PeftModel, LoraConfig, TaskType
config = LoraConfig(
task_type=TaskType.CAUSAL_LM,
target_modules=["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"],
inference_mode=True,
r=64, # Lora 秩
lora_alpha=16, # Lora alaph,具体作用参见 Lora 原理
lora_dropout=0.05, # Dropout 比例
bias="none",
)
# default: Load the model on the available device(s)
model = Qwen2VLForConditionalGeneration.from_pretrained(
"./Qwen/Qwen2-VL-2B-Instruct", torch_dtype="auto", device_map="auto"
)
model = PeftModel.from_pretrained(model, model_id="./output/Qwen2-VL-2B/checkpoint-62", config=config)
processor = AutoProcessor.from_pretrained("./Qwen/Qwen2-VL-2B-Instruct")
messages = [
{
"role": "user",
"content": [
{
"type": "image",
"image": "测试图像路径",
},
{"type": "text", "text": "COCO Yes:"},
],
}
]
# Preparation for inference
text = processor.apply_chat_template(
messages, tokenize=False, add_generation_prompt=True
)
image_inputs, video_inputs = process_vision_info(messages)
inputs = processor(
text=[text],
images=image_inputs,
videos=video_inputs,
padding=True,
return_tensors="pt",
)
inputs = inputs.to("cuda")
# Inference: Generation of the output
generated_ids = model.generate(**inputs, max_new_tokens=128)
generated_ids_trimmed = [
out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
]
output_text = processor.batch_decode(
generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
)
print(output_text)
使用4张A100 40GB显卡,batch size为4,gradient accumulation steps为4,训练2个epoch的用时为1分钟57秒。
- 在微调脚本中,
val_peft_model
加载的是一共固定的checkpoint文件,如果你添加了数据或超参数,请根据实际情况修改checkpoint文件路径。