Datawhale AI 夏令营---第四期(魔搭-从零入门AI生图)(小白文:一)
可图Kolors-LoRA风格故事挑战赛
任务一:参赛者需在可图Kolors 模型的基础上训练LoRA 模型,
任务二:基于LoRA模型生成 8 张图片组成连贯故事
例如:
向训练好的模型输入prompt,然后生成对应的图片。 如下:
1.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”,
2. prompt=“二次元,日系动漫,演唱会的观众席,人山人海,一个紫色短发小女孩穿着粉色吊带漏肩连衣裙坐在演唱会的观众席,舞台上衣着华丽的歌星们在唱歌”,
3. prompt=“二次元,一个紫色短发小女孩穿着粉色吊带漏肩连衣裙坐在演唱会的观众席,露出憧憬的神情”,
4. prompt=“二次元,一个紫色短发小女孩穿着粉色吊带漏肩连衣裙,对着流星许愿,闭着眼睛,十指交叉,侧面”,
5. prompt=“二次元,一个紫色中等长度头发小女孩穿着粉色吊带漏肩连衣裙,在练习室练习唱歌”,
6.prompt=“二次元,一个紫色长发小女孩穿着粉色吊带漏肩连衣裙,在练习室练习唱歌,手持话筒”
7.prompt=“二次元,紫色长发少女,穿着黑色连衣裙,试衣间,心情忐忑”,
8. prompt=“二次元,紫色长发少女,穿着黑色礼服,连衣裙,在台上唱歌”,
简单了解Kolors 模型
简单来讲 也是文本生图的模型,优点:对语言理解较深刻,可渲染中文
膜拜佬佬的文章【论文+中文文生图】
小白鼠的碎碎念:7月开源,好新;llm相关,萌新感到好难。
简单了解LoRA 模型
LoRA(Low-Rank Adaptation)是一种用于微调Stable Diffusion模型的训练技术。
LORA详解(史上最全)
LoRA的核心思想是通过将模型中的权重矩阵分解为低秩矩阵的乘积,从而减少参数的数量。在大型语言模型(如BERT、GPT)中,通常存在许多高维度的全连接层,LoRA通过对这些层的权重矩阵进行低秩分解,大幅降低了微调时的参数量。
LoRA的提出为大型预训练模型的广泛应用提供了一种更为经济有效的微调方式,使得在资源受限的条件下进行模型的特定任务优化成为可能
baseline
数据集构建:下载数据集kolors,处理数据集
#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"
)
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")
# 处理数据集 并保存
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
模型微调:模型微调训练
# 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)
图片生成:调用训练好的模型生成图片
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="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")
torch.manual_seed(1)
image = pipe(
prompt="二次元,日系动漫,演唱会的观众席,人山人海,一个紫色短发小女孩穿着粉色吊带漏肩连衣裙坐在演唱会的观众席,舞台上衣着华丽的歌星们在唱歌",
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")
# ......
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
最后,生成图片质量很高。