Datawhale X 魔搭 AI夏令营第四期 AIGC方向 task01笔记

赛事任务：

1.参赛者需在可图Kolors 模型的基础上训练LoRA 模型，生成无限风格，如水墨画风格、水彩风格、赛博朋克风格、日漫风格......

2.基于LoRA模型生成 8 张图片组成连贯故事，故事内容可自定义

本次使用lora微调后的SDXL-vae-fp16-fix预训练模型进行图像生成

下载数据集并将图片和元数据保存到本地

from modelscope.msdatasets import MsDataset

ds = MsDataset.load(
    'AI-ModelScope/lowres_anime',
    subset_name='default',
    split='train',
    cache_dir="/mnt/workspace/kolors/data"
)

import json, os
from data_juicer.utils.mm_utils import SpecialTokens
from tqdm import tqdm

os.makedirs("./data/lora_dataset/train", exist_ok=True)
os.makedirs("./data/data-juicer/input", exist_ok=True)
with open("./data/data-juicer/input/metadata.jsonl", "w") as f:
    for data_id, data in enumerate(tqdm(ds)):
        image = data["image"].convert("RGB")
        image.save(f"/mnt/workspace/kolors/data/lora_dataset/train/{data_id}.jpg")
        metadata = {"text": "二次元", "image": [f"/mnt/workspace/kolors/data/lora_dataset/train/{data_id}.jpg"]}
        f.write(json.dumps(metadata))
        f.write("\n")

使用 data-juicer （数据处理和转换工具，旨在简化数据的提取、转换和加载过程）处理数据（用pip下载 data-juicer后需要重启jupyter内核）然后保存

data_juicer_config = """
# global parameters
project_name: 'data-process'
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'
image_key: 'image'
image_special_token: '<__dj__image>'

export_path: './data/data-juicer/output/result.jsonl'

# process schedule
# a list of several process operators with their arguments
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)
        text = data["text"]
        texts.append(text)
        image = Image.open(data["image"][0])
        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

下载预训练模型用lora（Low-Rank Adaptation）微调，这里只训练一轮

# 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_alpha 4.0 \
  --dataset_path data/lora_dataset_processed \
  --output_path ./models \
  --max_epochs 1 \
  --center_crop \
  --use_gradient_checkpointing \
  --precision "16-mixed"
""".strip()

os.system(cmd)

加载训练好的模型，由于模型较大可能报cuda out of memory错误，nvidia-smi查看显存是否已满。可以通过更换GPU或者重启内核释放显存再直接加载模型的方法解决。

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)
    state_dict = torch.load(lora_path, map_location="cpu")
    model.load_state_dict(state_dict, strict=False)
    return model


# Load models
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"
                             ])
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"
)

 用如下方法指定提示词和图像大小来生成图片并保存

torch.manual_seed(0)
image = pipe(
    prompt="你想要生成的图片应该包含的内容",
    negative_prompt="你不希望生成的图片的内容",
    cfg_scale=4,
    num_inference_steps=50, height=1024, width=1024,
)
image.save("1.jpg")

 

单张图片质量较高，但图片间连贯性较差，地理背景和风格不统一