a error about the model of pytorch transform to libtorch

jit
1 
    # load net
    num_classes = len(labelmap) + 1                      # +1 for background
    net = build_refinedet('test', int(args.input_size), num_classes)            # initialize SSD
    net.load_state_dict(torch.load(args.trained_model,map_location='cuda:0'))    # model = torch.load(model_path, map_location='cuda:0')
    net.eval()
    print('Finished loading model!')

    # 向模型中输入数据以得到模型参数
    example = torch.rand(1,3,320,320).cuda()
    traced_script_module = torch.jit.trace(net,example)

    # 保存模型
    traced_script_module.save("torch_script_eval.pt")

i want to make the refinedet model transform to libtorch,but it have a error about “RuntimeError: Attempted to trace Detect_RefineDet, but tracing of legacy functions is not supported”

Thank you for helping me！

2

As the error message says, the model is using PyTorch 0.1.2-style Functions. (Newer autograd.Functions cannot be traced either, but that is another story…)
You probably want to convert the model to use the nms functions provided by TorchVision, those have tracing support.

Best regards

Thomas