Tim Cooijmans

> If we are indeed going to normalize sequence wise and batch wise, then the words that are "really loud" will be somewhat dampened. This would be skewed even more...

A text search for -1 in `Jsb16thSeparated.json` gives me several results. Are you sure you're not looking at `jsb-chorales-16th.json`? The former has voices separated and therefore always has four pitches...

@jaweriaamjad it depends on what you want to compute. `tf.gradients` computes vector-Jacobian products `u^T (dy/dx)`, and this trick computes Jacobian-vector products `(dy/dx) u`. The former is "how does a change...

@jaweriaamjad right, but to be clear, summing the columns means summing across `x`, not across `y`. It will not correspond to summing losses but to summing parameters. The shape of...