In this chapter, we used the datasets that we prepared in the previous chapter to train the Magenta models on different instruments and genres. We first compared different models and configurations for specific use cases and then showed how to create a new one if necessary.
Then, we launched different training runs and looked at how to tune the model for better training. We showed how to launch the training and evaluation jobs and how to check the resulting metrics on the TensorBoard. Then, we saw a case of overfitting and explained how to fix overfitting and underfitting. We also showed how to define the proper network size and hyperparameters, by looking at problems such as incorrect batch size, memory errors, a model not converging, and not having enough training data. Using our newly trained model, we've generated sequences and showed how to package and handle...
