"Object Detection Using Deep Learning" toturial doesn't use GPU

I am trying to do the "Object Detection Using Deep Learning" toturial from Matlab.

When i get to the "Train CNN Using CIFAR-10 Data" chapter, it starts to train the CNN, to accelerate the training it uses my GPU (an Nvidia Quadro M2200) and it works fine.

In one of the next chapters "Train R-CNN Stop Sign Detector" i have to retrain the network using new data. For retraining you can also use the GPU to accelerate the process, however when i run the code to retrain it says that it uses my GPU but using the task manager 'preformance' tab (in Windows 10) i can clearly see it doesn't and it retrains only using the cpu (which takes a lot longer).

My question is, why doesn't it use my GPU and how can i change it to use my GPU?

ps. my GPU supports CUDA version 5.2.

[Joss Knight]

I'm speculating, but R-CNN also does a lot of stuff on the CPU, processing candidate regions of interest. It's possible that you're expecting the GPU to be being used continuously while in fact the GPU computation of weight gradients is not the most compute-intensive part in terms of wall-clock time. What exactly are you looking at that makes you think the GPU isn't being used at all? Have you tried forcing training to happen on the CPU (using training option 'ExecutionEnvironment', 'cpu') to see whether there is any difference in training time?

[Tico Laugs]

Joss Knight thank you for the reply, i have tried focring the training on the CPU and it seems that doesn't make a difference. If it is the case that R-CNN does (almost) everything on the CPU then i don't understand why the toturial states that "a CUDA-capable NVIDIA™ GPU with compute capability 3.0 or higher is highly recommended for training".