Fig. 1From: Piano multipitch estimation using sparse coding embedded deep learningTraining structure of MPENet. Following Caffe terminology, this figure shows the computation graph of training phase. Arrows indicate blobs; rectangles and ellipses indicate Caffe built-in computation and loss layers. Hexagons indicate layer collections. Rounded rectangles indicate our implemented layers. All built-in layer names conform to Caffe. More details are explained in Sections 3 and 4. Repeated elements of each modality are omitted by dashed linesBack to article page