Part2_Music_Generation: model prediction inputs

[maple24]

Maybe this gets a little bit late, but I think this is also a good chance to say thank you to Alex and Ava for this great course.

Here is my questions:

When I went through Music_Generation codes, the answer here was quite confusing to me. Only one character passes to the model each time although it is updating but the previous information is missing. (I think this also part the reason why the songs generated are always invalid.)

~Pass the prediction along with the previous hidden state
~as the next inputs to the model
input_eval = tf.expand_dims([predicted_id], 0)

So I save the initial input and concatenate with each output as the next input. This makes more sense to me and the results start to be much better, but I'm not sure if I make something wrong or there are better ways, like taking the previous state as the next initializer state.

out_eval = tf.expand_dims([predicted_id], 0)
input_eval = tf.concat([input_eval, output_eval], 1)