dvruette/figaro

Source code for "FIGARO: Generating Symbolic Music with Fine-Grained Artistic Control"

figaro's Issues

about your forked MuseMorphose

I am sorry for the off topic.
I saw your forked MuseMorphose.
I can not use original Pre-Trained Models.
So if you don't mind, could you please provide your Pre-Trained Models?
Thank you.

Using pm.time_to_tick() twice on Tempo

While looking through your input_representation.py, I noticed that you use self.pm.time_to_tick() twice when saving the tempo_items.

I am talking about this part:

    max_tick = self.pm.time_to_tick(self.pm.get_end_time())
    existing_ticks = {item.start: item.pitch for item in self.tempo_items}
    wanted_ticks = np.arange(0, max_tick+1, DEFAULT_RESOLUTION)
    output = []
    for tick in wanted_ticks:
      if tick in existing_ticks:
        output.append(Item(
          name='Tempo',
          start=self.pm.time_to_tick(tick),
          end=None,
          velocity=None,
          pitch=existing_ticks[tick]))
      else:
        output.append(Item(
          name='Tempo',
          start=self.pm.time_to_tick(tick),
          end=None,
          velocity=None,
          pitch=output[-1].pitch))
    self.tempo_items = output

In line 145 you use max_tick = self.pm.time_to_tick(self.pm.get_end_time()) and then use a loop to go through the ticks from 0 to max_tick. When you append items to the output, you use start=self.pm.time_to_tick(tick), but tick is already a tick and not a time. This gives far bigger values for the tempo start compared to the chords and notes.

I don't know if changing this will help when using your pretrained weights, since this bug may have been there since training. I just wanted to note it nonetheless.

TypeError: __init__() got an unexpected keyword argument 'vae_run'

When I try to generate midi from the command:
python src/generate.py --model figaro --checkpoint ./checkpoints/figaro.ckpt --vae_checkpoint ./checkpoints/vq-vae.ckpt
The issue can be repeated.
And I am sure that the version of every packages I downloaded are correct.

Generated file not found

Training with extended chord vocabulary

I am trying to train FIGARO with an extended chord vocabulary (using the provided checkpoints).

I edited get_chord_tokens(...) in vocab.py to match the chord qualities in my dataset. However, when loading the checkpoint, I ran into the error of size missmatch for in_layer.weight and out_layer.weight - obviously, as the vocabulary changed.
Do you happen to know which additional steps are needed to continue training from the existing checkpoints, with a dataset that contains more chord qualities than the ones from the paper?

Thank you in advance!

more detailed information about "duration"

Thank you for your work.

Do you mind helping me out with a little bit more information about the "mean_duration" parameter for the expert description?
The only detailed information about duration that I can find in the paper is: "Mean duration is quantized to 32 logarithmically spaced intervals in [0, 128] positions (12 positions per quarter note)."
If I understand this correctly, this would imply that a duration value of 1 is equal to a note being played for 1/12 quarter notes, whereas a duration value of 32 would mean that it is played for 128/12 quarter notes. However, looking at the generated results from the example descriptions, this doesn't seem to be the case.

Kind regards, and thank you in advance!