Deep Learning with PyTorch

Loss functions

Once we have defined our network architecture, we are left with two important steps. One is calculating how good our network is at performing a particular task of regression, classification, and the next is optimizing the weight.

The optimizer (gradient descent) generally accepts a scalar value, so our loss function should generate a scalar value that has to be minimized during our training. Certain use cases, such as predicting where an obstacle is on the road and classifying it to a pedestrian or not, would require two or more loss functions. Even in such scenarios, we need to combine the losses to a single scalar for the optimizer to minimize. We will discuss examples of combining multiple losses to a single scalar in detail ...