app.py

import argparse
import glob
import os.path

import torch
import torch.nn.functional as F

import gradio as gr

import onnxruntime as rt
import tqdm

from midi_synthesizer import synthesis
import TMIDIX

import matplotlib.pyplot as plt

in_space = os.getenv("SYSTEM") == "spaces"
      
#=================================================================================================

@torch.no_grad()
def GenerateMIDI(progress=gr.Progress()):

    start_tokens = [3087, 3073+1, 3075+1]
    seq_len = 512
    max_seq_len = 2048
    temperature = 1.0
    verbose=False
    return_prime=False
    

    out = torch.FloatTensor([start_tokens])

    st = len(start_tokens)

    if verbose:
      print("Generating sequence of max length:", seq_len)

    progress(0, desc="Starting...")
    step = 0
    
    for i in progress.tqdm(range(seq_len)):

        try:
            x = out[:, -max_seq_len:]
    
            torch_in = x.tolist()[0]
    
            logits = torch.FloatTensor(session.run(None, {'input': [torch_in]})[0])[:, -1]

            probs = F.softmax(logits / temperature, dim=-1)
    
            sample = torch.multinomial(probs, 1)
    
            out = torch.cat((out, sample), dim=-1)

            if step % 16 == 0:
                print(step, '/', seq_len)

            step += 1
        
            if step >= seq_len:
                break
            
        except Exception as e:
            print('Error', e)
            break

    if return_prime:
      melody_chords_f = out[:, :]
    
    else:
      melody_chords_f = out[:, st:]

    melody_chords_f = melody_chords_f.tolist()[0]
    
    print('=' * 70)
    print('Sample INTs', melody_chords_f[:12])
    print('=' * 70)
    
    if len(melody_chords_f) != 0:
    
        song = melody_chords_f
        song_f = []
        time = 0
        dur = 0
        vel = 0
        pitch = 0
        channel = 0
    
        for ss in song:
    
          if ss > 0 and ss < 256:
    
              time += ss * 8
    
          if ss >= 256 and ss < 1280:
    
              dur = ((ss-256) // 8) * 32
              vel = (((ss-256) % 8)+1) * 15
    
          if ss >= 1280 and ss < 2816:
              channel = (ss-1280) // 128
              pitch = (ss-1280) % 128
    
              song_f.append(['note', int(time), int(dur), int(channel), int(pitch), int(vel) ])
    
    output_signature = 'Allegro Music Transformer'
    output_file_name = 'Allegro-Music-Transformer-Music-Composition'
    track_name='Project Los Angeles'
    list_of_MIDI_patches=[0, 24, 32, 40, 42, 46, 56, 71, 73, 0, 53, 19, 0, 0, 0, 0]
    number_of_ticks_per_quarter=500
    text_encoding='ISO-8859-1'
    
    output_header = [number_of_ticks_per_quarter, 
            [['track_name', 0, bytes(output_signature, text_encoding)]]]                                                    

    patch_list = [['patch_change', 0, 0, list_of_MIDI_patches[0]], 
                    ['patch_change', 0, 1, list_of_MIDI_patches[1]],
                    ['patch_change', 0, 2, list_of_MIDI_patches[2]],
                    ['patch_change', 0, 3, list_of_MIDI_patches[3]],
                    ['patch_change', 0, 4, list_of_MIDI_patches[4]],
                    ['patch_change', 0, 5, list_of_MIDI_patches[5]],
                    ['patch_change', 0, 6, list_of_MIDI_patches[6]],
                    ['patch_change', 0, 7, list_of_MIDI_patches[7]],
                    ['patch_change', 0, 8, list_of_MIDI_patches[8]],
                    ['patch_change', 0, 9, list_of_MIDI_patches[9]],
                    ['patch_change', 0, 10, list_of_MIDI_patches[10]],
                    ['patch_change', 0, 11, list_of_MIDI_patches[11]],
                    ['patch_change', 0, 12, list_of_MIDI_patches[12]],
                    ['patch_change', 0, 13, list_of_MIDI_patches[13]],
                    ['patch_change', 0, 14, list_of_MIDI_patches[14]],
                    ['patch_change', 0, 15, list_of_MIDI_patches[15]],
                    ['track_name', 0, bytes(track_name, text_encoding)]]

    output = output_header + [patch_list + song_f]

    midi_data = TMIDIX.score2midi(output, text_encoding)
    
    with open(f"Allegro-Music-Transformer-Music-Composition.mid", 'wb') as f:
        f.write(midi_data)

    output1 = []
    itrack = 1
    
    opus =  TMIDIX.score2opus(output)
          
    while itrack < len(opus):
        for event in opus[itrack]:
            if (event[0] == 'note_on') or (event[0] == 'note_off'): 
                output1.append(event)
        itrack += 1
        
    audio = synthesis([500, output1], 'SGM-v2.01-YamahaGrand-Guit-Bass-v2.7.sf2')
    
    x = []
    y =[]
    c = []
    
    colors = ['red', 'yellow', 'green', 'cyan', 'blue', 'pink', 'orange', 'purple', 'gray', 'white', 'gold', 'silver']
    
    for s in song_f:
      x.append(s[1] / 1000)
      y.append(s[4])
      c.append(colors[s[3]])

    plt.figure(figsize=(14,5))
    ax=plt.axes(title='Allegro Music Transformer')
    ax.set_facecolor('black')
    
    plt.scatter(x,y, c=c)
    plt.xlabel("Time")
    plt.ylabel("Pitch")

    yield [500, output1], plt, "Allegro-Music-Transformer-Music-Composition.mid", (44100, audio)
        
#=================================================================================================

if __name__ == "__main__":

    parser = argparse.ArgumentParser()
    parser.add_argument("--share", action="store_true", default=False, help="share gradio app")
    parser.add_argument("--port", type=int, default=7860, help="gradio server port")
    opt = parser.parse_args()

    print('Loading model...')
    session = rt.InferenceSession('Allegro_Music_Transformer_Small_Trained_Model_56000_steps_0.9399_loss_0.7374_acc.onnx', providers=['CUDAExecutionProvider', 'CPUExecutionProvider'])
    print('Done!')
    
    app = gr.Blocks()
    with app:
        gr.Markdown("<h1 style='text-align: center; margin-bottom: 1rem'>Allegro Music Transformer</h1>")
        gr.Markdown("![Visitors](https://api.visitorbadge.io/api/visitors?path=asigalov61.Allegro-Music-Transformer&style=flat)\n\n"
                    "Full-attention multi-instrumental music transformer featuring asymmetrical encoding with octo-velocity, and chords counters tokens, optimized for speed and performance\n\n"
                    "Check out [Allegro Music Transformer](https://github.com/asigalov61/Allegro-Music-Transformer) on GitHub!\n\n"
                    "[Open In Colab]"
                    "(https://colab.research.google.com/github/asigalov61/Allegro-Music-Transformer/blob/main/Allegro_Music_Transformer_Composer.ipynb)"
                    " for faster execution and endless generation"
                        )
                
        run_btn = gr.Button("generate", variant="primary")

        output_midi_seq = gr.Variable()
        output_audio = gr.Audio(label="output audio", format="mp3", elem_id="midi_audio")
        output_plot = gr.Plot(label="output plot")
        output_midi = gr.File(label="output midi", file_types=[".mid"])
        run_event = run_btn.click(GenerateMIDI, [], [output_midi_seq, output_plot, output_midi, output_audio])
        
        app.queue(2).launch(server_port=opt.port, share=opt.share, inbrowser=True)