AI-exp-4/nerf/train_nerf.py

import torch
import torch.nn as nn
from torch.utils.data import Dataset, DataLoader
import os
import json
import numpy as np
import argparse


class NeRFDataset(Dataset):
    def __init__(self, data_dir, split='train'):
        self.data_dir = data_dir
        self.split = split

        # 检查必要文件是否存在
        transforms_path = os.path.join(data_dir, 'transforms_train.json')
        if not os.path.exists(transforms_path):
            raise FileNotFoundError(
                f"找不到必要文件: {transforms_path}\n"
                "请确保：\n"
                "1. 数据集已正确下载\n"
                "2. 数据路径配置正确（当前路径：{data_dir}）\n"
                "3. 包含images目录和poses_bounds.npy文件"
            )

        with open(transforms_path, 'r') as f:
            self.transforms = json.load(f)

        # 验证文件格式
        if 'frames' not in self.transforms:
            raise ValueError(
                f"transforms.json格式错误，缺少必要字段'frames'\n"
                "请确保：\n"
                "1. 数据集文件完整\n"
                "2. 使用正确的数据集版本\n"
                "3. transforms.json包含frames数组和intrinsic参数"
            )

        self.image_paths = [os.path.join(data_dir, 'images', frame['file_path'])
                            for frame in self.transforms['frames']]
        self.poses = np.array([frame['transform_matrix'] for frame in self.transforms['frames']])

    def __len__(self):
        return len(self.image_paths)

    def __getitem__(self, idx):
        # 加载图像和位姿
        image_path = self.image_paths[idx]
        pose = self.poses[idx]

        # 简单的数据预处理
        # 生成随机坐标替代真实图像输入（NeRF典型做法）
        height, width = 400, 400  # 降低分辨率以减少计算量
        x = torch.rand(height * width, 3)  # (x, y, z) 坐标

        # 生成随机颜色作为目标值
        y = torch.rand(height * width, 4)  # (r, g, b, a) 目标值

        return {'coords': x, 'targets': y}


class PositionalEncoding(nn.Module):
    """位置编码层，将输入坐标扩展到更高维度"""

    def __init__(self, input_dim=3, max_freq_log2=9, N_freqs=10):
        super(PositionalEncoding, self).__init__()
        self.N_freqs = N_freqs
        self.funcs = [torch.sin, torch.cos]
        self.freq_bands = 2 ** torch.linspace(0, max_freq_log2, steps=N_freqs)
        self.output_dim = input_dim * len(self.funcs) * N_freqs

    def forward(self, x):
        """
        对输入坐标进行位置编码
        x: [N, input_dim] 输入坐标
        returns: [N, output_dim] 编码后的特征
        """
        out = []
        for freq in self.freq_bands:
            for func in self.funcs:
                out.append(func(freq * x))
        return torch.cat(out, dim=-1)


class NeRF(nn.Module):
    def __init__(self, input_ch=60, output_ch=4):
        super(NeRF, self).__init__()
        self.encoding = PositionalEncoding(input_dim=3, N_freqs=10)
        self.net = nn.Sequential(
            nn.Linear(self.encoding.output_dim, 256),
            nn.ReLU(),
            nn.Linear(256, 256),
            nn.ReLU(),
            nn.Linear(256, output_ch)
        )

    def forward(self, x):
        # 确保输入类型一致
        if x.dtype != torch.float32:
            x = x.float()

        # 应用位置编码
        x = self.encoding(x)
        return self.net(x)


if __name__ == '__main__':
    # 添加参数解析
    parser = argparse.ArgumentParser(description='NeRF训练脚本')
    parser.add_argument('--data_path', type=str, default='../data/nerf/lego',
                        help='数据集路径 (default: ../data/nerf/lego)')
    parser.add_argument('--batch_size', type=int, default=1,
                        help='批量大小 (default: 1)')
    parser.add_argument('--num_epochs', type=int, default=10,
                        help='训练轮数 (default: 10)')
    args = parser.parse_args()

    # 数据集路径配置
    try:
        dataset = NeRFDataset(args.data_path)
        dataloader = DataLoader(dataset, batch_size=args.batch_size, shuffle=True)
    except Exception as e:
        print(f"数据加载失败: {str(e)}")
        print("建议解决方案:")
        print("1. 验证数据集路径是否正确")
        print("2. 确保数据集已正确下载并解压")
        print("3. 检查transforms.json文件格式")
        exit(1)

    # 初始化模型
    model = NeRF()
    optimizer = torch.optim.Adam(model.parameters(), lr=5e-4)

    # 训练循环
    for epoch in range(args.num_epochs):
        for i, data in enumerate(dataloader):
            coords = data['coords']
            targets = data['targets']

            # 前向传播
            outputs = model(coords)
            loss = torch.mean((outputs - targets) ** 2)

            # 反向传播
            optimizer.zero_grad()
            loss.backward()
            optimizer.step()

            print(f'Epoch [{epoch + 1}/{args.num_epochs}], Step [{i + 1}/{len(dataloader)}], Loss: {loss.item():.4f}')

    # 保存模型
    os.makedirs(os.path.join(args.data_path, 'checkpoint'), exist_ok=True)
    torch.save(model.state_dict(), os.path.join(args.data_path, 'checkpoint', 'nerf_model.pth'))