xsl/change_hair_3090
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

初始化换发型项目：3个微服务代码 + 部署脚本

Browse Source 
包含：
- hair_service_sd: 换发型/换发色算法服务 (端口 8801)
- photo_service: LoRA 训练调度服务 (端口 32678)
- stable-diffusion-webui: SD WebUI 推理服务 (端口 57860)
- kohya_ss_home: 训练环境代码
- meidaojia: 监控测试脚本
- setup.sh: 一键部署脚本 (conda环境恢复 + 配置生成 + 完整性检查)
- start_all_services.sh: 启动3个服务
- configure.ini.template: 路径模板化 (BASE_DIR自动推导)
- conda_envs/py310.yml: py310 环境定义

大文件 (weights/, models/, data/, conda_envs/*.tar.gz 等) 通过 .gitignore 排除，
由网盘单独上传。
This commit is contained in:
colomi
2026-07-11 18:11:49 +08:00
commit 0eb61f3e60
628 changed files with 120882 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
photo_service/lora_train_service_1.py
+357
View File
@@ -0,0 +1,357 @@
# -*- coding: utf-8 -*-
import os
import sys
import uuid
from webui_im2im import ControlnetRequestImg2Img
import numpy as np
import base64
import cv2
import os, sys
from gevent import pywsgi, monkey
from multiprocessing import Process, Queue
import glob
import datetime
# from gpt4v_caption import caption_image
# monkey.patch_all()
# sys.setrecursionlimit(20000)
# 将当前工作目录切换到当前目录
project_dir = os.path.dirname(os.path.abspath(__file__))
os.chdir(project_dir)
sys.path.append(project_dir)
from flask import Flask, request, jsonify
import global_variable as global_var
import json
#from common.logger import config
BASE_DIR = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
version = "online"
if version == "local":
current_url = 'http://service.aicloud.fit:7393/'
kohya_ss_home_dir = '/home/chinatszrn/Documents/miaoya/kohya_ss_home'
webui_lora_dir = '/home/chinatszrn/Documents/miaoya/webui_home/stable-diffusion-webui/models/Lora'
inference_use_onediff = False
callback_url = 'http://service.aicloud.fit:7395/api/hair/trainCallBack'
else:
current_url = 'http://0.0.0.0:7393/'
kohya_ss_home_dir = os.path.join(BASE_DIR, 'kohya_ss_home')
webui_lora_dir = os.path.join(BASE_DIR, 'stable-diffusion-webui', 'models', 'Lora')
inference_use_onediff = False
callback_url = 'http://0.0.0.0:8801/api/hair/trainCallBack'
base_webui_port = '57860'
base_onediff_port = '9038'
base_webui_url = "http://127.0.0.1:57860/sdapi/v1/img2img"
app = Flask(__name__)
def send_request(url, state, task_dict):
import requests
import json
msg = '头发lora训练失败' if state == -1 else '头发lora训练成功'
payload = json.dumps({
"task_id": task_dict['task_id'] if task_dict is not None else '',
"hair_id": task_dict['hair_id'] if task_dict is not None else '',
"state": state,
"msg": msg,
"is_tj": task_dict['is_tj']
})
print("payload:", payload)
headers = {
'Content-Type': 'application/json'
}
requests.request("POST", url, headers=headers, data=payload)
@app.route('/api/hair/train', methods=['POST'])
def train_hair_lora():
import os
import json
task_id = ''
try:
# 获取请求参数
request_data = request.get_json()
assert 'task_id' in request_data, 'task_id is required'
task_id = request_data['task_id']
assert 'hair_id' in request_data and 'hair_material_dir' in request_data, 'hair_id and hair_material_dir is required'
hair_material_dir = request_data['hair_material_dir']
assert os.path.exists(hair_material_dir), f'{hair_material_dir} not exists'
img_dir = os.path.join(hair_material_dir, 'images')
assert os.path.exists(img_dir), f'{img_dir} not exists'
assert os.path.exists(
os.path.join(hair_material_dir, 'model')), f'{os.path.join(hair_material_dir, "model")} not exists'
#判断img_dir下面是否只有一个文件夹
img_dir_list = os.listdir(img_dir)
assert len(img_dir_list) == 1, f'{img_dir}下面只能有一个文件夹'
train_image_dir = os.path.join(img_dir, img_dir_list[0])
assert os.path.isdir(train_image_dir), f'{train_image_dir}不是文件夹'
# 检查该文件夹是否以数字加下划线开头
assert img_dir_list[0].split('_')[0].isdigit(), f'{img_dir_list[0]}不是数字开头'
#判断文件夹下面是否有图片
img_list = glob.glob(train_image_dir + '/*.png')
assert len(img_list) > 10, f'{os.path.join(img_dir, img_dir_list[0])}下面图片数量小于10张'
request_data['train_image_dir'] = train_image_dir
# is tianjin company
try:
request_data['is_tj'] = request_data['is_tj']
request_data['device_id'] = request_data['device_id']
except Exception as e:
print(e)
request_data['is_tj'] = '0'
request_data['device_id'] = '1'
# 将请求数据放入队列
global_var.hair_style_lora_train_task_sq.put(request_data)
# 返回结果
ret_dict = dict(state=0, msg='头发lora训练开始', task_id=task_id)
return jsonify(ret_dict)
except Exception as e:
ret_dict = dict(state=-1, msg=str(e), task_id=task_id)
return jsonify(ret_dict)
@app.route('/api/hair/inference', methods=['POST'])
def inference_webui():
import os
import requests
import time
try:
t1 = time.time()
# 获取请求参数
request_data = request.get_json()
# with open('request_data.json', 'w') as f:
# json.dump(request_data, f)
request_json = request_data.get('request_json')
# hd_version_flag = request_data.get('hd_version_flag', False)
hair_id = request_data['hair_id']
inference_port = request_data['inference_port']
if inference_use_onediff and 'refiner_checkpoint' not in request_json:
selected_url = base_webui_url.replace(base_webui_port, base_onediff_port)
request_json['script_name'] = 'onediff_diffusion_model'
print('user onediff')
else:
selected_url = base_webui_url.replace(base_webui_port, inference_port)
# if hd_version_flag:
# print('user hd version!!!!!')
# lora_file_name = 'hairstyle_hd_lora.safetensors'
# else:
# lora_file_name = 'hairstyle_lora.safetensors'
hair_material_dir = request_data.get('hair_material_dir')
print(hair_material_dir)
assert os.path.exists(hair_material_dir), f'{hair_material_dir} not exists'
hd_version_flag = True
lora_file_name = 'hairstyle_hd_lora.safetensors'
request_lora_path = os.path.join(hair_material_dir, 'model', lora_file_name)
if not os.path.exists(request_lora_path):
print('use low resolution lora')
hd_version_flag = False
lora_file_name = 'hairstyle_lora.safetensors'
request_lora_path = os.path.join(hair_material_dir, 'model', lora_file_name)
assert os.path.exists(request_lora_path), f'{request_lora_path} not exists'
if hd_version_flag:
tmp_lora_name = f'{hair_id}_hd'
else:
tmp_lora_name = f'{hair_id}'
lora_dst_path = os.path.join(global_var.webui_lora_dir, f'{tmp_lora_name}.safetensors')
if not os.path.exists(lora_dst_path):
os.system('cp {} {}'.format(request_lora_path, lora_dst_path))
request_json['prompt'] = f'<lora:{tmp_lora_name}:1.0>, titor hairstyle, ' + request_json['prompt']
request_json['negative_prompt'] = request_json['negative_prompt'] + ', faceless, no human'
print('pre_stage cost:', time.time() - t1)
start_time = time.time()
current_time = datetime.datetime.now()
print("*********cxt log1, infer***************, {}:{}:{}, {}".format(current_time.hour, current_time.minute,
current_time.second, selected_url))
response = requests.post(url=selected_url, json=request_json)
print('inference time:', time.time() - start_time)
ret_json = response.json()
# if os.path.exists(lora_dst_path):
# os.remove(lora_dst_path)
# 返回结果
return jsonify(ret_json)
except Exception as e:
ret_dict = dict(state=-1, msg=str(e))
return jsonify(ret_dict)
@app.route('/api/hair/inference_diy', methods=['POST'])
def inference_diy_webui():
import os
import requests
import time
try:
t1 = time.time()
# 获取请求参数
request_data = request.get_json()
# with open('request_data.json', 'w') as f:
# json.dump(request_data, f)
request_json = request_data.get('request_json')
inference_port = request_data['inference_port']
if inference_use_onediff and 'refiner_checkpoint' not in request_json:
selected_url = base_webui_url.replace(base_webui_port, base_onediff_port)
request_json['script_name'] = 'onediff_diffusion_model'
print('user onediff')
else:
selected_url = base_webui_url.replace(base_webui_port, inference_port)
request_json['prompt'] = f'titor hairstyle, ' + request_json['prompt']
request_json['negative_prompt'] = request_json['negative_prompt'] + ', faceless, no human'
print('pre_stage cost:', time.time() - t1)
start_time = time.time()
current_time = datetime.datetime.now()
print("*********cxt log1, diy***************, {}:{}:{}, {}".format(current_time.hour, current_time.minute,
current_time.second, selected_url))
response = requests.post(url=selected_url, json=request_json)
print('inference time:', time.time() - start_time)
ret_json = response.json()
# if os.path.exists(lora_dst_path):
# os.remove(lora_dst_path)
# 返回结果
return jsonify(ret_json)
except Exception as e:
ret_dict = dict(state=-1, msg=str(e))
return jsonify(ret_dict)
def train_thread(sq, gpu_id):
while True:
url = f'{current_url}api/hair/trainCallBack'
task_dict = None
try:
task_dict = sq.get()
hair_material_dir = task_dict['hair_material_dir']
images_dir = os.path.join(hair_material_dir, 'images')
model_dir = os.path.join(hair_material_dir, 'model')
train_image_dir = task_dict['train_image_dir']
# tag = task_dict['tag']
tag = ""
is_tj = task_dict['is_tj']
device_id = task_dict['device_id']
sample_dir = os.path.join(model_dir, 'sample')
if not os.path.exists(sample_dir):
os.makedirs(sample_dir)
sample_txt = os.path.join(sample_dir, 'prompt.txt')
with open(sample_txt, 'w') as f_s:
f_s.write('titor hairstyle, easyphoto, faceless, no human, white background, simple background, '
+ tag
+ ' --n low quality, worst quality, bad anatomy, bad composition, poor, low effort --h 768 '
'--w 768 --s 30 --l 7')
# 给训练图片打标签
# cmd_caption = ('docker run --rm --gpus all -v /home/chinatszrn/Documents/miaoya/kohya_ss_home:/home/chinatszrn '
# '-v /mnt:/mnt -e PATH=/home/chinatszrn/.local/bin -w /home/chinatszrn/kohya_ss '
# '--net=host chinatszrn/ubuntu:kohya_ss accelerate '
# 'launch "./finetune/tag_images_by_wd14_tagger.py" --batch_size=2 '
# '--general_threshold=0.5 --character_threshold=0.5 --caption_extension=".txt" '
# '--model="SmilingWolf/wd-v1-4-convnextv2-tagger-v2" --max_data_loader_n_workers="2" '
# '--debug --remove_underscore --frequency_tags --undesired_tags="1girl, 1boy" '
# f'"{train_image_dir}"')
# os.system(cmd_caption)
img_path_list = glob.glob(train_image_dir + '/*.png')
# 给训练图片打标签, gpt
for img_path in img_path_list:
# print(img_path)
txt_path = img_path[:-4] + '.txt'
print("tag img_path: ", img_path)
# tags = caption_image(img_path)
tags = tag
with open(txt_path, 'w') as f:
f.write('titor hairstyle, easyphoto, faceless, no human, white background, simple background, ' + tags)
#判断是否每个训练图片都有标签文件
# for img_path in img_path_list:
# assert os.path.exists(img_path[:-4]+'.txt'), f'{img_path}没有对应的标签文件'
# with open(img_path[:-4]+'.txt', 'r') as f:
# tags = f.readline()
# with open(img_path[:-4]+'.txt', 'w') as f:
# f.write('titor hairstyle, faceless, no human, gray background, simple background, ' + tags)
os.system(f'chmod 777 {model_dir}')
#训练头发lora
sd_models_dir = os.path.join(BASE_DIR, 'stable-diffusion-webui', 'models', 'Stable-diffusion')
container_images_dir = images_dir.replace(kohya_ss_home_dir, '/home/chinatszrn')
container_model_dir = model_dir.replace(kohya_ss_home_dir, '/home/chinatszrn')
cmd_train = (
f'sudo docker run --rm --privileged=true --gpus "device=0" '
f'-v {kohya_ss_home_dir}:/home/chinatszrn '
f'-v {sd_models_dir}:/mnt/sd_models '
'-e PATH=/home/chinatszrn/.local/bin -w /home/chinatszrn/kohya_ss '
'--net=host chinatszrn/ubuntu:kohya_ss accelerate launch --num_cpu_threads_per_process=2 "./train_network.py" --enable_bucket '
'--min_bucket_reso=256 --max_bucket_reso=2048 --pretrained_model_name_or_path="/mnt/sd_models/majicmixRealistic_v7.safetensors" '
f'--train_data_dir={container_images_dir} --resolution="2000,2000" '
f'--output_dir={container_model_dir} '
'--network_alpha="64" --save_model_as=safetensors --network_module=networks.lora --text_encoder_lr=5e-05 '
'--unet_lr=0.0001 --network_dim=128 --output_name="hairstyle_hd_lora" --lr_scheduler_num_cycles="20" '
'--no_half_vae --learning_rate="0.0001" --lr_scheduler="cosine" --lr_warmup_steps="650" --train_batch_size="1" '
'--max_train_steps="1500" --save_every_n_epochs="100" --mixed_precision="fp16" --save_precision="fp16" '
'--caption_extension=".txt" --sample_sampler=ddim '
f'--sample_prompts={sample_txt} --sample_every_n_epochs="1000" '
'--seed="1234" --cache_latents --optimizer_type="AdamW8bit" --max_data_loader_n_workers="0" --bucket_reso_steps=64 '
'--xformers --bucket_no_upscale --noise_offset=0.0 --tokenizer_cache_dir="/home/chinatszrn/.cache/clip"')
print("cmd_train:", cmd_train)
os.system(cmd_train)
lora_path = os.path.join(model_dir, 'hairstyle_hd_lora.safetensors')
# tianjin callback
if is_tj == "1":
url = f'{callback_url}'
if not os.path.exists(lora_path):
send_request(url, -1, task_dict)
else:
send_request(url, 0, task_dict)
except Exception as e:
send_request(url, -1, None)
continue
if __name__ == '__main__':
global_var.webui_lora_dir = webui_lora_dir
# avatar训练的任务队列
global_var.hair_style_lora_train_task_sq = Queue()
p = Process(target=train_thread, args=(
global_var.hair_style_lora_train_task_sq, 0))
p.start()
# # 启动服务
# app.run(debug=True, port=32678, host='0.0.0.0')
server = pywsgi.WSGIServer(('0.0.0.0', 32678), app) # test port
server.serve_forever()
p.join()