change_hair/train_lora_parallel.py

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""直接调 kohya train_network.py 真正并行训练 LoRA

绕过 photo_service 的串行阻塞，用 subprocess 直接起 train_network.py 进程，
通过 ThreadPoolExecutor 控制并发数。

用法:
  python /home/xsl/change_hair/train_lora_parallel.py

前提：step1 训练数据已准备好（data/train_material/<hid>/images/1_hairstyle/*.png）
"""
import os
import sys
import time
import subprocess
from concurrent.futures import ThreadPoolExecutor, as_completed

TRAIN_DIR = "/home/xsl/change_hair/data/train_material"
KOHYA_WORKDIR = "/home/xsl/change_hair/project/kohya_ss_home/kohya_ss"
BASE_MODEL = "/home/xsl/change_hair/project/onediff/stable-diffusion-webui/models/Stable-diffusion/v1-5-pruned-emaonly.safetensors"
KOHYA_ACCEL = "/home/xsl/miniconda3/envs/kohya/bin/accelerate"
LOG_DIR = "/home/xsl/change_hair/project/logs"

# 本次要训练的发型
HAIRSTYLES = ["chang_tuoyuan", "chang_zhixian", "chang_huaban", "chang_xinxing"]
# chang_bolang 已完成，跳过
PARALLEL = 2


def log(msg):
    print(f"[{time.strftime('%H:%M:%S')}] {msg}", flush=True)


def train_one(hair_id):
    """训练单个发型的 LoRA，返回 (hair_id, success, log_path)"""
    images_dir = os.path.join(TRAIN_DIR, hair_id, "images")
    model_dir = os.path.join(TRAIN_DIR, hair_id, "model")
    os.makedirs(model_dir, exist_ok=True)

    # 准备 sample prompt 文件
    sample_dir = os.path.join(model_dir, "sample")
    os.makedirs(sample_dir, exist_ok=True)
    sample_txt = os.path.join(sample_dir, "prompt.txt")
    with open(sample_txt, "w") as f:
        f.write('titor hairstyle, easyphoto, faceless, no human, white background, simple background, '
                ' --n low quality, worst quality, bad anatomy, bad composition, poor, low effort --h 768 '
                '--w 768 --s 30 --l 7')

    log_path = os.path.join(LOG_DIR, f"train_{hair_id}.log")

    cmd = [
        "bash", "-c",
        f'cd {KOHYA_WORKDIR} && '
        f'HF_HUB_OFFLINE=1 TRANSFORMERS_OFFLINE=1 CUDA_VISIBLE_DEVICES=0 '
        f'{KOHYA_ACCEL} launch --num_cpu_threads_per_process=2 "./train_network.py" --enable_bucket '
        f'--min_bucket_reso=256 --max_bucket_reso=2048 --pretrained_model_name_or_path="{BASE_MODEL}" '
        f'--train_data_dir={images_dir} --resolution="2000,2000" '
        f'--output_dir={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="AdamW" --max_data_loader_n_workers="0" --bucket_reso_steps=64 '
        '--sdpa --bucket_no_upscale --noise_offset=0.0'
    ]

    log(f"  ▶ {hair_id} 开始训练，日志: {log_path}")
    t0 = time.time()
    try:
        with open(log_path, "w") as lf:
            proc = subprocess.run(cmd, stdout=lf, stderr=subprocess.STDOUT, timeout=1800)
        ok = proc.returncode == 0
        lora = os.path.join(model_dir, "hairstyle_hd_lora.safetensors")
        ok = ok and os.path.exists(lora) and os.path.getsize(lora) > 100000000
        elapsed = time.time() - t0
        sz = os.path.getsize(lora) // 1024 // 1024 if os.path.exists(lora) else 0
        log(f"  {'✓' if ok else '✗'} {hair_id} {'训练完成' if ok else '训练失败'} ({sz}MB, {elapsed:.0f}s)")
        return (hair_id, ok, log_path)
    except subprocess.TimeoutExpired:
        log(f"  ✗ {hair_id} 训练超时")
        return (hair_id, False, log_path)
    except Exception as e:
        log(f"  ✗ {hair_id} 训练异常: {e}")
        return (hair_id, False, log_path)


def main():
    t0 = time.time()
    log(f"并行 LoRA 训练: {HAIRSTYLES} (并发={PARALLEL})")

    results = {}
    with ThreadPoolExecutor(max_workers=PARALLEL) as ex:
        futures = {ex.submit(train_one, hid): hid for hid in HAIRSTYLES}
        for fut in as_completed(futures):
            hid = futures[fut]
            try:
                hair_id, ok, log_path = fut.result()
                results[hair_id] = ok
            except Exception as e:
                log(f"  ✗ {hid} 异常: {e}")
                results[hid] = False

    log(f"\n{'='*60}")
    log(f"全部完成，耗时 {time.time()-t0:.0f}s")
    for hid, ok in results.items():
        lora = os.path.join(TRAIN_DIR, hid, "model", "hairstyle_hd_lora.safetensors")
        sz = os.path.getsize(lora) // 1024 // 1024 if os.path.exists(lora) else 0
        log(f"  {'✓' if ok else '✗'} {hid} ({sz}MB)")
    success = [h for h, ok in results.items() if ok]
    log(f"成功: {success}")

    # 写完成清单
    done_file = os.path.join(LOG_DIR, "train_lora_done.json")
    import json
    with open(done_file, "w") as f:
        json.dump({"completed": success, "all": HAIRSTYLES}, f)
    log(f"完成清单: {done_file}")


if __name__ == "__main__":
    main()