Skip to content

Processed loss parameter and set default value based on the last layer of the network architecture

Carlos Alfredo Yeverino Rodriguez requested to merge process_loss_parameter into master

The loss default value is set based on the last layer of the network architecture. It is cross-entropy loss for a softmax or sigmoid layer and euclidean/squared loss for all other cases. This is achieved by changing the loss default value in the train function from the CNNCreator. All available loss functions should be provided in the CNNCreator as previously agreed in the DL Seminar.

The user has the possibility to specify a different loss function from the default value by providing the loss parameter in the training configuration file using CNNTrain. The explicitly given loss is then stored in the CNNTrainer which overrides the default value by calling the train function. In this way, more combinations are possible such as Sigmoid-Euclidean.

Edited by Carlos Alfredo Yeverino Rodriguez

Merge request reports