There is a strong coupling between output activation and the loss function used during training. The demo program uses sigmoid() output layer activation. This approach assumes that you use BCELoss() ( ...