caffe c++ demo speed is too slow

Hi, First of all, thank your great project.
I convert the pytorch model to caffe, then realized it in c++ code, the accuracy is pretty good But I met the speed problem.

    Caffe::set_mode(Caffe::GPU);
    ...
    double t1 = static_cast<double>(cv::getTickCount());
    net_->Forward();
    double t2 = static_cast<double>(cv::getTickCount());
    double inferenceTime = (t2 - t1) / cv::getTickFrequency() * 1000;
    std::cout << "model infrence time is: " << inferenceTime << std::endl;
   ...

In my GPU card (1080ti), It needs about 30 ms for each inference, Could you please tell the reason or potential fix method ? thanks a lot.

[lxy5513]

@Daniil-Osokin So kind of you! At last I realize the caffe model with depthconv, The speed of inference is about 10ms.

By the way, only in the condition of dilation=1 , you can change convolution to DepthwiseConvolution, or the result will is a little strange.

[Daniil-Osokin]

Great, that it works!