ncnn推理时，出现Segmentation fault (core dumped)问题

[cqchenqianqc]

定位到以下发现出错 void NanoDet::decode_infer(ncnn::Mat& cls_pred, ncnn::Mat& dis_pred, int stride, float threshold, std::vector<std::vector>& results) { int feature_h = this->input_size / stride; int feature_w = this->input_size / stride; std::cout<< "feature_h:"<<feature_h<<std::endl; std::cout<< "cls_pred"<<cls_pred<<std::endl; std::cout<< "decode_infer"<<std::endl;

//std::cout<< "cls_pred"<<cls_pred.row()<<std::endl;
//cv::Mat debug_heatmap = cv::Mat(feature_h, feature_w, CV_8UC3);
for (int idx = 0; idx < feature_h * feature_w; idx++)
{
    //std::cout << "w * h: " << feature_h * feature_w << std::endl;
    const float* scores = cls_pred.row(idx);
    int row = idx / feature_w;
    int col = idx % feature_w;
    float score = 0;
    int cur_label = 0;
    for (int label = 0; label < this->num_class; label++)
    {
        if (scores[label] > score)
        {
            score = scores[label];
            cur_label = label;
        }
    }

[cqchenqianqc]

please,help me !! give me some suggestion

[RangiLyu]

你在nanodet.cpp里的类别数量和输入大小是否和训练时设置的保持一致？

[cqchenqianqc]

你在nanodet.cpp里的类别数量和输入大小是否和训练时设置的保持一致？

我在nanodet.cpp里的类别数量和输入大小是否和训练时设置的是 保持一致，现在定位到nanodet.cpp中NanoDet::decode_infer函数 中如下的if语句出错，我将每一次的值score打印出来，但是每一次报错的idx都不一样，且错误都是Segmentation fault (core dumped) const float* scores = cls_pred.row(idx) if (scores[label] > score) { score = scores[label]; cur_label = label; }

[RangiLyu]

这个问题一般来说是越界了，可以检查一下模型最后输出的blob形状和预期的是否一致

[cqchenqianqc]

好的，谢谢！！

[TD-wzw]

I have the same problem as you，You go to download the author's ncnn weight file can run, but with their own training weight file or their own conversion income can not

[TD-wzw]

I have the same problem as you，You go to download the author's ncnn weight file can run, but with their own training weight file or their own conversion income can not

[TD-wzw]

I have checked that the output shape of the file I converted by myself is different from that provided by the author. Why

[hrishikeshps94]

I'm also having the same error, it works perfectly with authors weight file when img_size is 320, but when the img_size is changed the above mentioned error occurs. For a different input shape should we convert the weight to ncnn format with specific img_size? Is it necessary to have fixed input_size when the model is converted from pytorch-->onnx-->ncnn?

[RangiLyu]

I'm also having the same error, it works perfectly with authors weight file when img_size is 320, but when the img_size is changed the above mentioned error occurs. For a different input shape should we convert the weight to ncnn format with specific img_size? Is it necessary to have fixed input_size when the model is converted from pytorch-->onnx-->ncnn?

The Interp layer's parameters in the ncnn model are set to fix values when converting from onnx. If you want to use a dynamic input shape, you need manually modify the .param file. You can refer to the nanodet model in the ncnn repo. https://github.com/nihui/ncnn-assets/blob/master/models/nanodet_m.param https://github.com/Tencent/ncnn/blob/master/examples/nanodet.cpp

Also, notice that the input shape must be divisible by 32.

[cqchenqianqc]

I'm also having the same error, it works perfectly with authors weight file when img_size is 320, but when the img_size is changed the above mentioned error occurs. For a different input shape should we convert the weight to ncnn format with specific img_size? Is it necessary to have fixed input_size when the model is converted from pytorch-->onnx-->ncnn?

Img_size is 608 in my experiment，when cls_pred.row(idx) in void NanoDet::decode_infer(ncnn::Mat cls_pred, ncnn::Mat dis_pred, int stride, float threshold, std::vector<std::vector>& results) may occurs bounds errors. for (int q=0; q<cls_pred.c; q++) { const float* ptr = cls_pred.channel(q); for (int y=0; y<cls_pred.h; y++)

        {
    std::cout << "cls_pred.c: " << cls_pred.c <<" .h:"<< cls_pred.h<<" .w:"<< cls_pred.w<<"\n"<<std::endl;

            for (int x=0; x<cls_pred.w; x++)
            {
                printf("%f ", ptr[x]);
            }
            ptr += cls_pred.w;
            printf("\n");
        }
        printf("------------------------\n");
    }

instead of
const float* scores = cls_pred.row(idx); may be can solve the problem

[DuZzzs]

In my case, because main.cpp set the size of resize to 320x320, change it to your own model size. Run successfully resize_uniform(image, resized_img, cv::Size(input_w, input_h), effect_roi);

[robofisshy]

In my case, I use the confg/nanodet-m.yml train the COCO dataset, no modification in model. But after the deploy pipeline, .pth->.onnx->sim.onnx->ncnn.param, ncnn.bin->ncnn-sim.param, ncnn-sim.bin, my output layers are differnt from author's model. I got similar fault as @cqchenqianqc , after I change the heads_info(like upon image), it works. You can check your .param, maybe the same problem @cqchenqianqc

[cqchenqianqc]

In my case, I use the confg/nanodet-m.yml train the COCO dataset, no modification in model. But after the deploy pipeline, .pth->.onnx->sim.onnx->ncnn.param, ncnn.bin->ncnn-sim.param, ncnn-sim.bin, my output layers are differnt from author's model. I got similar fault as @cqchenqianqc , after I change the heads_info(like upon image), it works. You can check your .param, maybe the same problem @cqchenqianqc

thanks,I try it