A Novel Diminish Smooth L1 Loss Model with Generative Adversarial Network

Abstract

The training process of GAN can be regarded as a process in which the generation network and the identification network play against each other and finally reach a state where it cannot be further improved if the opponent does not change. At the same time, the start of the gradient descent method will choose a direction to reduce the defined loss. The loss function plays a key role in the performance of the model. Choosing the right loss function can help your model learn how to focus on the correct set of features in the data to achieve optimal and faster convergence. In this work, we propose a novel loss function scheme, namely, Diminish Smooth L1 loss. We improve a robust L1 loss called Smooth L1 loss by lowering the threshold so that the network can converge to a lower minimum. From our experimental results on several benchmark data, we found that our algorithm often outperforms the previous approaches.