I have a hard time thinking about how I can build this network with the following problem: I want to build a CNN to classify notes from sheet music. I have tried several models with and without transfer learning, with only one label for each image i.e. 'C quarter note' or 'B half note' (not exactly like that, but you get the idea). Now training the model is going fine, but classifying new images is still rubbish. Now I want to try and split the network into two: one network for training the pitch (label 1) and one network for training the duration of the note (label 2). So each image has two labels. How can I do this best?

1) building and training two separate models and just combine the results.
2) build one multilabel network (you still get label1 x label2 amount of labels right?).
3) a network with two output layers, each with one label and one Softmax function. although I'm not sure how to construct this and if this is different from option 1.

My second question regarding transfer learning: the models trained on imagenet are entirely different from my own dataset, so my logic was to just freeze the first couple of layers for basic features and train the rest of the model, is this correct?




Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.