Hi,
I am using ResNet
to predict a continuous output of size 100
(the output has the form of a signal with values bounded between 0
and 0.5
and the input is a 2D image).
Obviously, using the CrossEntropyLoss
would not help here as I need to apply some activation function (most likely the Sigmoid
) at the output layer during testing, and things will go wrong as the CrossEntropyLoss
uses Softmax
.
Thus, I need to use one of PyTorch loss functions to enable using the same activation function both in training and testing. I am defining an output layer of size 100, and will thus use BCEWithLogitsLoss
.
Are there other alternatives?