Real-time 实时性

[zryzds]

您好，请问现在TCTrack有通过TensorRT算法加速的版本了吗

[assertdebug]

Hello, I made some mistakes in transferring Onnx

[zryzds]

Hello, I made some mistakes in transferring Onnx

i transferred model to onnx type,what's your mistake

[iangiu]

Hello, I made some mistakes in transferring Onnx

i transferred model to onnx type,what's your mistake

@zryzds 你好，请教一下，如何将模型导出onnx？因为我看作者的代码，初始化内容较多，我修改了forward的部分代码，如下：

def forward(self, presearch, template, search):
        presearch=presearch.cuda()
        template = template.cuda()
        search = search.cuda()

        videorange = 3
        presearch=t.cat((presearch[ : , 5 - videorange:,:,:,:],search.unsqueeze(1)),1)    
            
        zf = self.backbone(template.unsqueeze(1))
        xf = self.backbone(presearch) ###b l c w h
        xf=xf.view(1, videorange+1,xf.size(-3),xf.size(-2),xf.size(-1))
       
        loc,cls2,cls3=self.grader(xf[:,-1,:,:,:],zf,xf[:,:-1,:,:,:].permute(1,0,2,3,4))
     
        return  loc, cls2, cls3

然后export onnx，代码如下：

torch.onnx.export(
        model,
        (presearch, template, search),
        "tctrackplus.onnx",
        export_params=True,
        opset_version=11,
        do_constant_folding=True,
        input_names=['presearch', 'template', 'search'],
        output_names=['loc', 'cls2', 'cls3'],
        dynamic_axes={
            'presearch': {0: 'batch_size'},
            'template': {0: 'batch_size'},
            'search': {0: 'batch_size'},
            'loc': {0: 'batch_size'},
            'cls2': {0: 'batch_size'},
            'cls3': {0: 'batch_size'}
        }
    )

最后报错为：Failed to export a node '%338

如果方便的话，希望能帮助一下，非常感谢