How could I use a designed loss function in training deep learning networks? Such as in faster r-cnn and cnn.

Hello,

This is a problem that has puzzled me for a long time.

How could I use a designed loss function in training deep learning networks? Now, when training a deep neural network, we just use the 'trainFasterRCNNObjectDetector' or the 'trainNetwork' function. The loss function in each network is default.

But I want to use the loss function I designed.

Has anyone do it?