Small bug in generator

[Waguy02]

Thanks for your code, it is very nice. But I found a little bug in the bar_generator code.
In fact, the last Reshape() layer in the network definition has an incorrect size. As you mentionned, the output shape of the last conv is (batch_size, out_channels, 8*hid_features//hid_channels, n_pitches) and the shape of Reshape layer is output shape: (batch_size, out_channels, 1, n_steps_per_bar, n_pitches). Such configuration works very well only when n_steps_per_bar is set to 16 but One would run experiment with a different n_steps_per_bar. (I run it with 48 steps per bar as far as I am concerned in fact).

For a more general purpose, I propose to add a linear mapping before that Reshape layer as shown in the figure attached.

Yours sincerely.