请问能加个联系方式吗。 想问下，官方训练好的模型是那个模型。

We have uploaded the pre-trained weights of the SuperNets(for both ImageNet classification and object tracking) to Google Drive. Users can use them as initialization for future research on efficient object tracking.

是网盘提供的这个文件吗。

您好，我在替换了lib/models/super_model_DP.py和tracking/torch2onnx.py后，无论是转换backbone还是neck_head模型，程序都会在torch.onnx.export内突然结束，没有明显的报错，只有Process finished with exit code 132 (interrupted by signal 4: SIGILL)，前面的读模型都是正常的，也能print(siam_net) super_model_DP.py代码 torch2onnx.py代码 运行结果

我通过单步调试发现程序在F.conv2d()函数内突然中止了，不清楚是否跟我的LightTrack环境有关，我没有完全安装官方的安装脚本的库，只装了pytorch，cuda，cudnn，torchvision等转换格式必要的库，因为按照官方脚本有些库装不上。 如能解答，不胜感激

#if NCNN_VULKAN and USE_GPU std::cout << NCNN_VULKAN << std::endl; ex_backbone.set_vulkan_compute(true); #endif

// net_init.opt.use_vulkan_compute= true; // net_init.set_vulkan_device(0); net_init.opt.use_fp16_packed = true; net_init.opt.use_fp16_storage = true; net_init.opt.use_fp16_arithmetic = true; net_init.opt.use_int8_storage = true; net_backbone.opt.use_fp16_packed = true; net_backbone.opt.use_fp16_storage = true; net_backbone.opt.use_fp16_arithmetic = true; net_backbone.opt.use_int8_storage = true; net_neck_head.opt.use_fp16_packed = true; net_neck_head.opt.use_fp16_storage = true; net_neck_head.opt.use_fp16_arithmetic = true; net_neck_head.opt.use_int8_storage = true;

我增加了这些部分，但gpu推理失败了 错误信息为 compile spir-v module failed ERROR: 0:10: 'constant_id' : only allowed when generating SPIR-V ERROR: 0:10: '' : compilation terminated ERROR: 2 compilation errors. No code generated.

出错位置为 ex_backbone.extract("output.1", xf);

你好,我这里遇到如下的报错: Update stage ---- output cls_score and bbox_pred extracting cost time : 6 ms Update stage ---- postprocess cost time : 0 ms pscore_window max score is: nan

通过单步调试发现, cls_score_data, bbox_pred_data1,bbox_pred_data2, bbox_pred_data3, bbox_pred_data4都是 NaN

` float* cls_score_data = (float*)cls_score.data; cls_score_sigmoid.clear();

int cols = cls_score.w;
int rows = cls_score.h;

for (int i = 0; i < cols*rows; i++)   // 18 * 18
{
    cls_score_sigmoid.push_back(sigmoid(cls_score_data[i]));
}

std::vector<float> pred_x1(cols*rows, 0), pred_y1(cols*rows, 0), pred_x2(cols*rows, 0), pred_y2(cols*rows, 0);

float* bbox_pred_data1 = bbox_pred.channel(0);
float* bbox_pred_data2 = bbox_pred.channel(1);
float* bbox_pred_data3 = bbox_pred.channel(2);
float* bbox_pred_data4 = bbox_pred.channel(3);`

难道是在网络的输出地方就出错了? ex_neck_head.extract("output.1", cls_score); // [c, w, h] = [1, 18, 18] ex_neck_head.extract("output.2", bbox_pred); // [c, w, h] = [4, 18, 18]