无需服务器，免费使用2个GPU T4 在云上训练自己的lora

%cd /kaggle/working

!git clone https://github.com/bmaltais/kohya_ss.git
    
%cd /kaggle/working/kohya_ss

!./setup.sh -n
print("Kohya_SS is done setting up!")

import os
import shutil

# Create the directories
os.makedirs("/kaggle/working/training-data/img", exist_ok=True)
os.makedirs("/kaggle/working/training-data/my-lora", exist_ok=True)
os.makedirs("/kaggle/working/training-data/log", exist_ok=True)
os.makedirs("/kaggle/working/training-data/config", exist_ok=True)

!rm -r /kaggle/working/training-data/*
# Copy the files
shutil.copytree("/kaggle/input/training/image", f"/kaggle/working/training-data/img")
shutil.copytree("/kaggle/input/training/config", f"/kaggle/working/training-data/config")

# 这行代码安装 pyngrok
!pip install pyngrok

#---------------Ngrok---------------------
#ngrok穿透token
#前往https://dashboard.ngrok.com/get-started/your-authtoken 获取
ngrok_token = '2bswgJj**********M5V4' #直接将Token粘贴到这里
# Ngrok近期更新了流量限制，Ngrok每个月有10G流量限制，如果用完可以用临时邮箱再注册一个

# 确保运行这个单元，以便函数被定义
def ngrok_start(port: int, address_name: str, ngrok_token: str, should_run: bool):
    if not should_run:
        print('Skipping ngrok start')
        return
    print('Attempting to start ngrok...')
    from pyngrok import conf, ngrok
    conf.get_default().auth_token = ngrok_token
    conf.get_default().monitor_thread = False
    ssh_tunnels = ngrok.get_tunnels(conf.get_default())
    if len(ssh_tunnels) == 0:
        ssh_tunnel = ngrok.connect(port, bind_tls=True)
        print(f'{address_name} URL: ' + ssh_tunnel.public_url)
    else:
        print(f'{address_name} URL: ' + ssh_tunnels[0].public_url)

# 使用实际的 ngrok token 和正确的端口号
ngrok_token = '2bswgJj**********M5V4'
ngrok_start(7860, 'Gradio', ngrok_token, True)

!bash gui.sh --share --headless