VLAC/train.py

import os
import sys
import json
import subprocess
from pathlib import Path

def create_dataset_meta(train_json: str, val_json: str, output_meta: str):
    """创建ms-swift需要的dataset meta文件"""
    meta = {
        "root": "/scratch1/home/zhicao/ATM/data/atm_libero/libero_goal",
        "annotation": {
            "train": train_json,
            "val": val_json
        },
        "data_augment": False,
        "max_dynamic_patch": 1,
        "image_aspect_ratio": "square"
    }
    
    os.makedirs(os.path.dirname(output_meta), exist_ok=True)
    with open(output_meta, 'w', encoding='utf-8') as f:
        json.dump(meta, f, indent=2, ensure_ascii=False)
    
    print(f"Dataset meta已保存到: {output_meta}")
    return output_meta

def main():
    import argparse
    
    parser = argparse.ArgumentParser(description="VLAC模型训练")
    parser.add_argument("--pretrained_model", type=str,
                       default="/scratch1/home/zhicao/VLAC/models/VLAC",
                       help="预训练模型路径")
    parser.add_argument("--checkpoint_dir", type=str,
                       default="/scratch1/home/zhicao/VLAC/models/1",
                       help="checkpoint保存目录")
    parser.add_argument("--train_json", type=str,
                       default="/scratch1/home/zhicao/VLAC/data/train_dataset_train.json",
                       help="训练数据集JSON文件（messages格式）")
    parser.add_argument("--val_json", type=str,
                       default="/scratch1/home/zhicao/VLAC/data/train_dataset_val.json",
                       help="验证数据集JSON文件（messages格式）")
    parser.add_argument("--num_epochs", type=int, default=3,
                       help="训练轮数")
    parser.add_argument("--batch_size", type=int, default=2,
                       help="每设备batch size")
    parser.add_argument("--gradient_accumulation_steps", type=int, default=8,
                       help="梯度累积步数")
    parser.add_argument("--learning_rate", type=float, default=2e-5,
                       help="学习率")
    parser.add_argument("--use_lora", action="store_true",
                       help="使用LoRA微调（节省显存）")
    parser.add_argument("--freeze_vit", action="store_true", default=True,
                       help="冻结视觉编码器")
    
    args = parser.parse_args()
    
    # 检查数据集是否存在
    if not os.path.exists(args.train_json):
        print(f"错误: 训练数据集不存在: {args.train_json}")
        print("请先运行 prepare_dataset.py 准备数据集")
        print("示例: python prepare_dataset.py")
        sys.exit(1)
    
    if not os.path.exists(args.val_json):
        print(f"错误: 验证数据集不存在: {args.val_json}")
        print("请先运行 prepare_dataset.py 准备数据集")
        sys.exit(1)
    
    # 创建dataset meta
    data_dir = os.path.dirname(args.train_json)
    dataset_meta = os.path.join(data_dir, "dataset_meta.json")
    create_dataset_meta(args.train_json, args.val_json, dataset_meta)
    
    # 检查预训练模型
    if not os.path.exists(args.pretrained_model):
        print(f"警告: 预训练模型不存在: {args.pretrained_model}")
        print("将使用HuggingFace上的模型: OpenGVLab/InternVL2-2B")
        model_id = "OpenGVLab/InternVL2-2B"
        ckpt_dir = None
    else:
        model_id = "OpenGVLab/InternVL2-2B"  # 模型类型标识（用于推断模板等）
        ckpt_dir = args.pretrained_model  # 本地模型路径
        print(f"使用本地预训练模型: {ckpt_dir}")
        # 验证模型文件是否存在
        model_file = os.path.join(ckpt_dir, "model.safetensors")
        config_file = os.path.join(ckpt_dir, "config.json")
        if not os.path.exists(model_file) and not os.path.exists(os.path.join(ckpt_dir, "pytorch_model.bin")):
            print(f"警告: 模型权重文件不存在: {model_file}")
        if not os.path.exists(config_file):
            print(f"警告: 配置文件不存在: {config_file}")
    
    # 构建swift sft命令
    cmd_parts = [
        "swift", "sft",
        "--model", model_id,
        "--model_type", "internvl2",
    ]
    
    # 如果使用本地checkpoint，使用 --ckpt_dir 指定路径
    if ckpt_dir:
        cmd_parts.extend(["--ckpt_dir", ckpt_dir])
    
    cmd_parts.extend([
        "--dataset", dataset_meta,
        "--val_dataset", dataset_meta,
        "--system", "You are a visual-language assistant designed to interpret spatial and task-related information from images and text. Provide precise, context-aware responses and actionable guidance to assist in achieving task objectives.",
        "--max_length", "10240",
        "--output_dir", args.checkpoint_dir,
        "--overwrite_output_dir",
        "--per_device_train_batch_size", str(args.batch_size),
        "--per_device_eval_batch_size", "1",
        "--gradient_accumulation_steps", str(args.gradient_accumulation_steps),
        "--learning_rate", str(args.learning_rate),
        "--weight_decay", "0.1",
        "--num_train_epochs", str(args.num_epochs),
        "--lr_scheduler_type", "cosine",
        "--warmup_steps", "100",
        "--save_steps", "500",
        "--eval_steps", "100",
        "--logging_steps", "10",
        "--save_total_limit", "3",
        "--dataloader_num_workers", "4",
        "--seed", "42",
        "--bf16",
        "--gradient_checkpointing",
    ])
    
    if args.freeze_vit:
        cmd_parts.append("--freeze_vit")
        cmd_parts.append("--freeze_aligner")
    
    if args.use_lora:
        cmd_parts.extend([
            "--lora",
            "--lora_rank", "8",
            "--lora_alpha", "32",
            "--lora_dropout", "0.05",
            "--target_modules", "all-linear",
        ])
    
    cmd = " ".join(cmd_parts)
    
    print("\n" + "="*80)
    print("训练配置:")
    print("="*80)
    if ckpt_dir:
        print(f"预训练模型: {ckpt_dir}")
    else:
        print(f"预训练模型: {model_id} (从HuggingFace下载)")
    print(f"训练数据集: {args.train_json}")
    print(f"验证数据集: {args.val_json}")
    print(f"Checkpoint目录: {args.checkpoint_dir}")
    print(f"训练轮数: {args.num_epochs}")
    print(f"Batch size: {args.batch_size}")
    print(f"梯度累积: {args.gradient_accumulation_steps}")
    print(f"学习率: {args.learning_rate}")
    print(f"使用LoRA: {args.use_lora}")
    print(f"冻结ViT: {args.freeze_vit}")
    print("="*80)
    print("\n训练命令:")
    print("="*80)
    print(cmd)
    print("="*80 + "\n")
    
    # 确认
    response = input("是否开始训练? (y/n): ")
    if response.lower() != 'y':
        print("训练已取消")
        sys.exit(0)
    
    # 执行训练
    print("开始训练...")
    try:
        subprocess.run(cmd_parts, check=True)
        print(f"\n训练完成！模型已保存到: {args.checkpoint_dir}")
    except subprocess.CalledProcessError as e:
        print(f"\n训练失败: {e}")
        sys.exit(1)
    except KeyboardInterrupt:
        print("\n训练被用户中断")
        sys.exit(1)

if __name__ == "__main__":
    main()