一,baseline代码结构
1、环境安装
!pip install simple-aesthetics-predictor
!pip install -v -e data-juicer
!pip uninstall pytorch-lightning -y
!pip install peft lightning pandas torchvision
!pip install -e DiffSynth-Studio
2、下载数据集
#下载数据集
from modelscope.msdatasets import MsDataset
ds = MsDataset.load(
'AI-ModelScope/lowres_anime', # 指定要下载的数据集的唯一标识符
subset_name='default', # 指定数据集的子集名称,这里使用默认子集
split='train', # 指定下载训练集部分的数据
cache_dir="/mnt/workspace/kolors/data" # 指定下载训练集部分的数据
)
import json, os
from data_juicer.utils.mm_utils import SpecialTokens
from tqdm import tqdm
os.makedirs("./data/lora_dataset/train", exist_ok=True) # 创建用于存储处理后的图像文件的目录,exist_ok=True参数意味着如果目录已经存在,则不会抛出错误
os.makedirs("./data/data-juicer/input", exist_ok=True) # 创建用于存储Data Juicer所需元数据文件的目录
with open("./data/data-juicer/input/metadata.jsonl", "w") as f:
# 使用enumerate函数遍历数据集ds,同时获取每个数据项的索引(即数据ID)和数据本身
for data_id, data in enumerate(tqdm(ds)):
image = data["image"].convert("RGB") # 使用convert("RGB")方法将其转换为RGB模式
image.save(f"/mnt/workspace/kolors/data/lora_dataset/train/{data_id}.jpg") # 将处理后的图像保存到指定的文件路径
# 构造一个包含文本描述和图像路径的字典metadata
metadata = {"text": "二次元", "image": [f"/mnt/workspace/kolors/data/lora_dataset/train/{data_id}.jpg"]}
f.write(json.dumps(metadata)) # 将metadata字典转换为JSON格式的字符串
f.write("\n")
3、处理数据集,保存数据处理结果
data_juicer_config = """
# global parameters
project_name: 'data-process' # 定义项目的名称
dataset_path: './data/data-juicer/input/metadata.jsonl' # path to your dataset directory or file
np: 4 # number of subprocess to process your dataset
text_keys: 'text' # 数据中文本的键名
image_key: 'image' # 数据中图像的键名
image_special_token: '<__dj__image>'
export_path: './data/data-juicer/output/result.jsonl' # 指定了处理后数据输出路径
# process schedule
# a list of several process operators with their arguments
# 包含了一系列数据处理的步骤和参数
process:
- image_shape_filter:
min_width: 1024
min_height: 1024
any_or_all: any
- image_aspect_ratio_filter:
min_ratio: 0.5
max_ratio: 2.0
any_or_all: any
"""
with open("data/data-juicer/data_juicer_config.yaml", "w") as file:
file.write(data_juicer_config.strip())
!dj-process --config data/data-juicer/data_juicer_config.yaml
import pandas as pd
import os, json
from PIL import Image
from tqdm import tqdm
texts, file_names = [], []
os.makedirs("./data/lora_dataset_processed/train", exist_ok=True)
with open("./data/data-juicer/output/result.jsonl", "r") as file:
for data_id, data in enumerate(tqdm(file.readlines())):
data = json.loads(data) # 将每行文本从 JSON 字符串转换为 Python 字典
text = data["text"]
texts.append(text)
image = Image.open(data["image"][0]) # 使用 PIL(Pillow)库的 Image.open 方法打开图像文件
image_path = f"./data/lora_dataset_processed/train/{data_id}.jpg"
image.save(image_path)
file_names.append(f"{data_id}.jpg")
data_frame = pd.DataFrame()
data_frame["file_name"] = file_names
data_frame["text"] = texts
data_frame.to_csv("./data/lora_dataset_processed/train/metadata.csv", index=False, encoding="utf-8-sig")
data_frame
4、lora微调
# 下载模型
from diffsynth import download_models
download_models(["Kolors", "SDXL-vae-fp16-fix"])
#模型训练
import os
cmd = """
python DiffSynth-Studio/examples/train/kolors/train_kolors_lora.py \
--pretrained_unet_path models/kolors/Kolors/unet/diffusion_pytorch_model.safetensors \
--pretrained_text_encoder_path models/kolors/Kolors/text_encoder \
--pretrained_fp16_vae_path models/sdxl-vae-fp16-fix/diffusion_pytorch_model.safetensors \
--lora_rank 16 \ # LoRA矩阵的秩
--lora_alpha 4.0 \ # alpha值
--dataset_path data/lora_dataset_processed \
--output_path ./models \
--max_epochs 1 \ # 指定了训练的最大轮次
--center_crop \ # 数据预处理选项,用于在训练前对图像进行中心裁剪
--use_gradient_checkpointing \ # 数据预处理选项,用于在训练前对图像进行中心裁剪
--precision "16-mixed" # 指定了训练时的精度设置,这里使用的是混合精度训练,其中某些操作使用16位浮点数来减少内存占用和加速计算
""".strip()
os.system(cmd)
5、加载微调好的模型
from diffsynth import ModelManager, SDXLImagePipeline
from peft import LoraConfig, inject_adapter_in_model
import torch
def load_lora(model, lora_rank, lora_alpha, lora_path):
lora_config = LoraConfig(
r=lora_rank,
lora_alpha=lora_alpha,
init_lora_weights="gaussian",
target_modules=["to_q", "to_k", "to_v", "to_out"],
)
model = inject_adapter_in_model(lora_config, model) # inject_adapter_in_model 函数将LoRA注入到模型中
state_dict = torch.load(lora_path, map_location="cpu") # 加载LoRA的权重到模型中,这里使用 torch.load 加载权重,并通过 map_location="cpu" 指定权重加载到CPU上
model.load_state_dict(state_dict, strict=False)
return model # 返回修改后的模型
# Load models
# 加载预训练的模型组件
model_manager = ModelManager(torch_dtype=torch.float16, device="cuda",
file_path_list=[
"models/kolors/Kolors/text_encoder",
"models/kolors/Kolors/unet/diffusion_pytorch_model.safetensors",
"models/kolors/Kolors/vae/diffusion_pytorch_model.safetensors"
])
# 通过 SDXLImagePipeline.from_model_manager 方法,使用 ModelManager 实例来构建一个图像处理管道 pipe
pipe = SDXLImagePipeline.from_model_manager(model_manager)
# Load LoRA
pipe.unet = load_lora(
pipe.unet,
lora_rank=16, # This parameter should be consistent with that in your training script.
lora_alpha=2.0, # lora_alpha can control the weight of LoRA.
lora_path="models/lightning_logs/version_0/checkpoints/epoch=0-step=500.ckpt"
)
6、图片生成
torch.manual_seed(0)
image = pipe(
prompt="二次元,一个紫色短发小女孩,在家中沙发上坐着,双手托着腮,很无聊,全身,粉色连衣裙",
negative_prompt="丑陋、变形、嘈杂、模糊、低对比度",
cfg_scale=4,
num_inference_steps=50, height=1024, width=1024,
)
image.save("1.jpg")
参考文章:Datawhale (linklearner.com)