app.py

import gradio as gr
from transformers import AutoTokenizer, TFGPT2LMHeadModel
from transformers import pipeline
import note_seq

from utils import token_sequence_to_note_sequence, create_image_from_note_sequence

SAMPLE_RATE=44100

# Feel free to change this, I am using only three notes here because the model 
# works better this way.
notes = ["A2", "C3", "D3", "E4", "G4", "E5"]
notes_to_midi = {"A2": 45, "C3": 48, "D3": 50, "E4": 64, "G4":67, "E5": 76}
time_signatures = ["4/4", "3/4", "2/4", "6/8"]
time_signature_to_tokens = {"4/4": "4_4", "3/4": "3_4", "2/4": "2_4", "6/8": "6_8"}

# Mutopi model for music generation
mutopia_model = TFGPT2LMHeadModel.from_pretrained("juancopi81/mutopia_guitar_mmm")
mutopia_tokenizer = AutoTokenizer.from_pretrained("juancopi81/mutopia_guitar_mmm")
pipe = pipeline(
    "text-generation", model=mutopia_model, tokenizer=mutopia_tokenizer, device=0
)

# Content for your demo:
title = "Mutopia Guitar Composer"
# I am adding here an image that I generated using DALL-E
description = """
The bot was trained to compose guitar music using the 
[Mutopia Guitar Dataset](https://huggingface.co/datasets/juancopi81/mutopia_guitar_dataset). 
Change the controllers and receive a new guitar piece!
<figure>
<center>
<img src="https://drive.google.com/uc?export=view&id=1F22ofTCeJAHqVag4lJvBZugAE1OyabVA" 
width=200px alt="Robot playing the guitar">
<figcaption>Image generated using DALL-E</figcaption>
</center>
</figure>
"""
article = """
For a complete tutorial on how to create this demo from scratch, check out this
[GitHub Repo](https://github.com/juancopi81/MMM_Mutopia_Guitar). 
This demo is inspired in the work of [Dr. Tristan Behrens](https://www.linkedin.com/in/dr-tristan-behrens-734967a2/)
"""

# Helper function to create the string seed
def create_seed(time_signature: str,
                note: str,
                bpm: int,
                density: int) -> str:
  
  seed = (f"PIECE_START TIME_SIGNATURE={time_signature_to_tokens[time_signature]} "
          f"BPM={bpm} TRACK_START INST=0 DENSITY={density} "
          f"BAR_START NOTE_ON={notes_to_midi[note]} ")
  return seed
          
def generate_guitar_piece(time_signature: str,
                          note: str,
                          bpm: int,
                          density: int):
  seed = create_seed(time_signature, note, bpm, density)
  piece = pipe(seed, max_length=250)[0]["generated_text"]
  
  # Convert text of notes to audio
  note_sequence = token_sequence_to_note_sequence(piece)
  synth = note_seq.midi_synth.fluidsynth
  array_of_floats = synth(note_sequence, sample_rate=SAMPLE_RATE)
  int16_data = note_seq.audio_io.float_samples_to_int16(array_of_floats)
  piano_roll = create_image_from_note_sequence(note_sequence)
  return (SAMPLE_RATE, int16_data), piano_roll

# Create a block object
demo = gr.Blocks()

# Use your Block object as a context
with demo:
  gr.Markdown("<h1 style='text-align: center; margin-bottom: 1rem'>" 
              + title 
              + "</h1>")
  gr.Markdown(description)

  # UI for the inputs of the model
  gr.Markdown("Select the generation parameters.")
  with gr.Row():
    time_signature = gr.Dropdown(time_signatures, value="6/8", label="Time signature") 
    note = gr.Dropdown(notes, value="E4", label="First note")
    bpm = gr.Slider(minimum=60, maximum=130, step=10, value=120, label="Tempo")
    density = gr.Slider(minimum=0, maximum=4, step=1, value=2, label="Density")
  with gr.Row():
    btn = gr.Button("Compose")
  with gr.Row():
    audio_output = gr.Audio()
    image_output = gr.Image()
  btn.click(generate_guitar_piece, 
            inputs = [
                time_signature, 
                note, 
                bpm, 
                density
            ],
            outputs=[audio_output, image_output])
  
  gr.Markdown(article)

# Launch your demo
demo.launch(debug=False)