Datawhale X 魔搭 AI夏令营AIGC方向task2

第一步：安装 

安装 Data-Juicer 和 DiffSynth-Studio

!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

请在这里手动重启 Notebook kernel 

第二步：下载数据集 

from modelscope.msdatasets import MsDataset

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

保存数据集中的图片及元数据

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

os.makedirs("./data/lora_dataset/train", exist_ok=True)
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")

第三步：数据处理 

使用 data-juicer 处理数据

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

保存处理好的数据

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

第四步：训练模型 

下载模型

from diffsynth import download_models

download_models(["Kolors", "SDXL-vae-fp16-fix"])

查看训练脚本的输入参数

!python DiffSynth-Studio/examples/train/kolors/train_kolors_lora.py -h

开始训练

提示：

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)
    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"
)

生成图像

torch.manual_seed(0)
image = pipe(
    prompt="古风，水墨画，一个黑色长发少女，坐在教室里，盯着黑板，深思，上半身，红色长裙",
    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")

torch.manual_seed(1)
image = pipe(
    prompt="古风，水墨画，一个黑色长发少女，坐在教室里，趴在桌子上睡着了，上半身，红色长裙",
    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("2.jpg")

torch.manual_seed(2)
image = pipe(
    prompt="古风，水墨画，一个黑色长发少女，站在路边，上半身，红色长裙",
    negative_prompt="丑陋、变形、嘈杂、模糊、低对比度，色情擦边",
    cfg_scale=4,
    num_inference_steps=50, height=1024, width=1024,
)
image.save("3.jpg")

torch.manual_seed(5)
image = pipe(
    prompt="古风，水墨画，一个英俊少年，骑着白马，上半身，白色衬衫",
    negative_prompt="丑陋、变形、嘈杂、模糊、低对比度，扭曲的手指，多余的手指",
    cfg_scale=4,
    num_inference_steps=50, height=1024, width=1024,
)
image.save("4.jpg")

torch.manual_seed(0)
image = pipe(
    prompt="古风，水墨画，一个英俊少年，白色衬衫，一个黑色长发少女，红色长裙，两个人一起聊天，开心，上半身",
    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("5.jpg")

torch.manual_seed(1)
image = pipe(
    prompt="古风，水墨画，一个英俊少年，白色衬衫，一个黑色长发少女，红色长裙，两个人一起骑着马，全身",
    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("6.jpg")

torch.manual_seed(7)
image = pipe(
    prompt="古风，水墨画，一个黑色长发少女，坐在教室里，下课铃声响了，同学们开始走动，从睡梦中醒来，深思，上半身，红色长裙",
    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("7.jpg")

torch.manual_seed(0)
image = pipe(
    prompt="古风，水墨画，一个黑色长发少女，坐在教室里，盯着黑板，认真上课，上半身，红色长裙",
    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("8.jpg")

import numpy as np
from PIL import Image

images = [np.array(Image.open(f"{i}.jpg")) for i in range(1, 9)]
image = np.concatenate([
    np.concatenate(images[0:2], axis=1),
    np.concatenate(images[2:4], axis=1),
    np.concatenate(images[4:6], axis=1),
    np.concatenate(images[6:8], axis=1),
], axis=0)
image = Image.fromarray(image).resize((1024, 2048))
image

总结