PyTorch C++ API中Conv2d析构函数崩溃问题求助

我编写了如下代码，代码运行后结果正确，但在Conv2d的析构函数阶段发生崩溃，请问有没有人遇到过类似情况？

代码 #

torch::Tensor kernel = torch::full({ 1, 32, 1, 2 }, 0, torch::kInt32);
// Define the bias
torch::Tensor bias = torch::tensor({ 0 }, torch::kInt32);
// Define the input image
torch::Tensor image = torch::full({ 32, 6, 7 }, 0, torch::kInt32);

/// Define the convolution operation
//depth_conv = Conv2d(in_channels = 10, out_channels = 10, kernel_size = 3, groups = 10)
torch::nn::Conv2d depth_conv = torch::nn::Conv2d(
    torch::nn::Conv2dOptions(32, 1, {1,2}).groups(1).bias(false)
);
 
// Set the filter for the convolution operation
depth_conv->weight = kernel;

// Set the bias for the convolution operation
depth_conv->bias = bias;
std::cout << "Original Data Tensor:\n" << image << std::endl;
std::cout << "Original Wght Tensor:\n" << depth_conv->weight << std::endl;

// Apply the convolution operation
torch::Tensor output = depth_conv->forward(image);
std::cout << "Output:\n" << output.dtype() << std::endl;

// Print the output shape and values
std::cout << "Output Shape: " << output.sizes() << std::endl;
std::cout << "Output:\n" << output.dtype() << std::endl;

问题原因及解决方案 #

• 核心崩溃原因：初始化Conv2d时设置了.bias(false)，这会让层不分配bias参数的内存空间，但后续强行给depth_conv->bias赋值，导致析构时程序尝试释放未被正确分配的bias资源，直接触发崩溃。
• 输入维度错误：Conv2d要求输入是4维张量（batch_size, in_channels, height, width），当前的image是3维{32,6,7}，需要调整为{1,32,6,7}（batch设为1），否则forward时会有维度不匹配的隐患。
• 数据类型问题：PyTorch的Conv2d默认权重是浮点类型，用kInt32虽然能运行，但可能引发计算精度或内部逻辑的潜在问题，建议统一使用float32类型。

修正后的代码 #

torch::Tensor kernel = torch::full({ 1, 32, 1, 2 }, 0.f, torch::kFloat32);
// 如果需要bias，初始化时不要加.bias(false)
torch::Tensor bias = torch::tensor({ 0.f }, torch::kFloat32);
// 调整输入为4维：batch=1, channels=32, height=6, width=7
torch::Tensor image = torch::full({ 1, 32, 6, 7 }, 0.f, torch::kFloat32);

// 初始化Conv2d时开启bias（如果需要的话）
torch::nn::Conv2d depth_conv = torch::nn::Conv2d(
    torch::nn::Conv2dOptions(32, 1, {1,2}).groups(1).bias(true)
);
 
depth_conv->weight = kernel;
depth_conv->bias = bias;

std::cout << "Original Data Tensor:\n" << image << std::endl;
std::cout << "Original Wght Tensor:\n" << depth_conv->weight << std::endl;

torch::Tensor output = depth_conv->forward(image);
std::cout << "Output:\n" << output.dtype() << std::endl;
std::cout << "Output Shape: " << output.sizes() << std::endl;
std::cout << "Output:\n" << output << std::endl;

如果不需要bias，直接删除depth_conv->bias = bias这一行即可，同时保持.bias(false)的设置。

内容的提问来源于stack exchange，提问作者user1941008