$ mkdir build && cd build
$ cmake .. && make -j 
$ make install

$ cd install/lighttrack_demo
$ ./LightTrack [videopath(file or camera)]

class Super_model_DP_retrain(Super_model):
    def __init__(self, search_size=256, template_size=128, stride=16):
        super(Super_model_DP_retrain, self).__init__(search_size=search_size, template_size=template_size,
                                                     stride=stride)

    def template(self, z):
        # print("z shape is:", z.shape)
        self.zf = self.features(z)

    def track(self, x):
        # print("x shape is:", x.shape)
        # supernet backbone
        xf = self.features(x)

        # print("xf shape is:", xf.shape)
        # print("zf shape is:", self.zf.shape)
        # BN before Pointwise Corr
        zf, xf = self.neck(self.zf, xf)
        # Point-wise Correlation
        feat_dict = self.feature_fusor(zf, xf)
        # supernet head
        oup = self.head(feat_dict)
        return oup['cls'], oup['reg']

    # when convert model, annotate this function
    def forward(self, template, search, label, reg_target, reg_weight):
        '''backbone_index: which layer's feature to use'''
        zf = self.features(template)
        xf = self.features(search)
        # Batch Normalization before Corr
        zf, xf = self.neck(zf, xf)
        # Point-wise Correlation
        feat_dict = self.feature_fusor(zf, xf)
        # supernet head
        oup = self.head(feat_dict)
        # compute loss
        reg_loss = self.add_iouloss(oup['reg'], reg_target, reg_weight)
        cls_loss = self._weighted_BCE(oup['cls'], label)
        return cls_loss, reg_loss

    # # convert init-net
    # def forward(self, z):   # init
    #     self.zf = self.features(z)
    #     return self.zf

    # # convert update
    # def forward(self, zf, x):   # update
    #     xf = self.features(x)
    #     # BN before Pointwise Corr
    #     zf, xf = self.neck(zf, xf)
    #     # Point-wise Correlation
    #     feat_dict = self.feature_fusor(zf, xf)
    #     # supernet head
    #     oup = self.head(feat_dict)
    #     return oup['cls'], oup['reg']

    # # convert init-net
    # x = torch.randn(1, 3, 127, 127)
    # mod = torch.jit.trace(siam_net, x)
    # mod.save("ligthtrack_init.pt")
    # os.system("/home/pxierra/nas/6-Software/pnnx-20211223-ubuntu/pnnx ligthtrack_init.pt inputshape=[1,3,127,127]")

    # convert update
    zf = torch.randn(1, 96, 8, 8)
    x = torch.randn(1, 3, 288, 288)
    mod = torch.jit.trace(siam_net, (zf, x))
    mod.save("ligthtrack_update.pt")
    os.system("/home/pxierra/nas/6-Software/pnnx-20211223-ubuntu/pnnx ligthtrack_update.pt inputshape=[1,96,8,8],[1,3,288,288]")

$ cd LightTrack/tracking
$ python torch2pnnx.py --arch LightTrackM_Subnet --resume ../snapshot/LightTrackM/LightTrackM.pth --stride 16 --path_name back_04502514044521042540+cls_211000022+reg_100000111_ops_32