Fig. 4From: Training audio transformers for cover song identificationThe training loss with or without average poolingBack to article page