How can I use the function 'Predict' in deep learning toolbox with full GPU capacity

I'm currently learning to use the deep learning toolbox to train a U-net for image regression tasks. When I was trying to use the 'predict' function with a minibatch size of 512 to generate some output, I found that the GPU memory usage was fluctuating periodically from 3.3/11GB to 9/11GB. I am pretty new to the matlab so I am wondering what is the cause of the phenomenon and how can I maximize the GPU useage so as to further accelerate the computation. Could anyone give me a little hint or suggestions?

Answer 1

[Srivardhan Gadila]

As per my knowledge, the fluctuation in the memory usage is due to the different computational and memory requirement across the layers of the U-Net network i.e., based on the layer properties there may be more number of computations involved for a layer when compared to another layer within the same network.

Refer to Deep Learning with Big Data on GPUs and in Parallel for other information.

[Huhufly]

Thank you Srivardhan, this actually makes a lot of sense to me.