Hey,
For this music generation problem, the training data has Tx=Ty=30, with the following architecture as presented in the lecture:
For the test problem, x<0> is a single music note of shape (m=1, Tx=1, n_values), and the prediction is a list of length Ty=50, where each list element has shape (1, n_values). It thus seems to me that the test problem architecture is one-to-many:
Thus, the many-to-many LSTM trained on 30 sequences of music is being used to forward propagate a single input x<0> (a single music note) to generate a sequence with 50 music notes.
Please let me know if my understanding is correct. Thanks,