app.py

#============================================================================================
# https://huggingface.co/spaces/projectlosangeles/Orpheus-Humanizing-Transformer
#============================================================================================

print('=' * 70)
print('Orpheus Humanizing Transformer Gradio App')

print('=' * 70)
print('Loading core Orpheus Humanizing Transformer modules...')

import os
import copy

import time as reqtime
import datetime
from pytz import timezone

print('=' * 70)
print('Loading main Orpheus Humanizing Transformer modules...')

os.environ['USE_FLASH_ATTENTION'] = '1'

import torch

torch.set_float32_matmul_precision('high')
torch.backends.cuda.matmul.allow_tf32 = True # allow tf32 on matmul
torch.backends.cudnn.allow_tf32 = True # allow tf32 on cudnn
torch.backends.cuda.enable_flash_sdp(True)

from huggingface_hub import hf_hub_download

import TMIDIX

from midi_to_colab_audio import midi_to_colab_audio

from x_transformer_2_3_1 import *

import random

import tqdm

print('=' * 70)
print('Loading aux Orpheus Humanizing Transformer modules...')

import matplotlib.pyplot as plt

import gradio as gr
import spaces

print('=' * 70)
print('PyTorch version:', torch.__version__)
print('=' * 70)
print('Done!')
print('Enjoy! :)')
print('=' * 70)

#==================================================================================

MODEL_CHECKPOINT = 'Orpheus_Music_Transformer_Trained_Model_96332_steps_0.82_loss_0.748_acc.pth'

SOUDFONT_PATH = 'SGM-v2.01-YamahaGrand-Guit-Bass-v2.7.sf2'

#==================================================================================

print('=' * 70)
print('Instantiating model...')

device_type = 'cuda'
dtype = 'bfloat16'

ptdtype = {'bfloat16': torch.bfloat16, 'float16': torch.float16}[dtype]
ctx = torch.amp.autocast(device_type=device_type, dtype=ptdtype)

SEQ_LEN = 8192
PAD_IDX = 18819

model = TransformerWrapper(num_tokens = PAD_IDX+1,
                           max_seq_len = SEQ_LEN,
                           attn_layers = Decoder(dim = 2048,
                                                 depth = 8,
                                                 heads = 32,
                                                 rotary_pos_emb = True,
                                                 attn_flash = True
                                                 )
                           )

model = AutoregressiveWrapper(model, ignore_index=PAD_IDX, pad_value=PAD_IDX)

print('=' * 70)
print('Loading model checkpoint...')      

model_checkpoint = hf_hub_download(repo_id='asigalov61/Orpheus-Music-Transformer', filename=MODEL_CHECKPOINT)

model.load_state_dict(torch.load(model_checkpoint, map_location=device_type, weights_only=True))

model = torch.compile(model, mode='max-autotune')

model.to(device_type)
model.eval()

print('=' * 70)
print('Done!')
print('=' * 70)
print('Model will use', dtype, 'precision...')
print('=' * 70)

#==================================================================================

def load_midi(input_midi):

    raw_score = TMIDIX.midi2single_track_ms_score(input_midi)
    
    escore_notes = TMIDIX.advanced_score_processor(raw_score, return_enhanced_score_notes=True, apply_sustain=True)
    
    if escore_notes:
    
        escore_notes = TMIDIX.augment_enhanced_score_notes(escore_notes[0], sort_drums_last=True)
        
        dscore = TMIDIX.delta_score_notes(escore_notes)
        
        dcscore = TMIDIX.chordify_score([d[1:] for d in dscore])
        
        melody_chords = [18816]
    
        chords = []
        
        #=======================================================
        # MAIN PROCESSING CYCLE
        #=======================================================
        
        for i, c in enumerate(dcscore):
        
            delta_time = c[0][0]
        
            melody_chords.append(delta_time)
    
            cho = []
            
            cho.append(delta_time)
        
            for e in c:
            
                #=======================================================
                
                # Durations
                dur = max(1, min(255, e[1]))
        
                # Patches
                pat = max(0, min(128, e[5]))
                
                # Pitches
                ptc = max(1, min(127, e[3]))
                
                # Velocities
                # Calculating octo-velocity
                
                vel = max(8, min(127, e[4]))
                velocity = round(vel / 15)-1
                
                #=======================================================
                # FINAL NOTE SEQ
                #=======================================================
                
                # Writing final note
                pat_ptc = (128 * pat) + ptc 
                dur_vel = (8 * dur) + velocity
        
                melody_chords.extend([pat_ptc+256, dur_vel+16768]) # 18816
                cho.extend([pat_ptc+256, dur_vel+16768])
    
            chords.append(cho)
        
        print('Done!')
        print('=' * 70)
        print('Score has', len(melody_chords), 'tokens')
        print('Score has', len(chords), 'chords')
        print('=' * 70)

        return melody_chords, chords

    else:
        return None

#==================================================================================

@spaces.GPU
def Humanize_MIDI(input_midi,
                  num_prime_toks,
                  num_hum_notes,
                  humanize_durations,
                  humanize_velocities,
                  model_temperature,
                  model_sampling_top_p
                 ):

    #===============================================================================

    print('=' * 70)
    print('Req start time: {:%Y-%m-%d %H:%M:%S}'.format(datetime.datetime.now(PDT)))
    start_time = reqtime.time()
    print('=' * 70)

    print('=' * 70)
    print('Requested settings:')
    print('=' * 70)
    fn = os.path.basename(input_midi)
    fn1 = fn.split('.')[0]
    print('Input MIDI file name:', fn)
    print('Number of prime tokens:', num_prime_toks)
    print('Number of notes to humanize:', num_hum_notes)
    print('Humanize durations:', humanize_durations)
    print('Humanize velocities:', humanize_velocities)
    print('Model temperature:', model_temperature)
    print('Model top p:', model_sampling_top_p)
   
    print('=' * 70)

    #==================================================================

    if input_midi is not None:

        print('Loading MIDI...')
    
        score, chords = load_midi(input_midi.name)
    
        if score is not None and chords is not None:
            
            print('Sample score tokens', score[:10])
            print('=' * 70)
        
            #==================================================================

            dur_vel_toks_num = len([t for t in score[num_prime_toks:] if 16767 < t < 18816])

            print('Number of tokens to humanize:', dur_vel_toks_num)

            #==================================================================
            
            print('=' * 70)
            print('Generating...')

            final_song = score[:num_prime_toks]

            hn_count = 0
        
            for t in tqdm.tqdm(score[num_prime_toks:]):

                if t < 16767 or t > 18815:
                    final_song.append(t)

                else:

                    fdur = ((t-16768) // 8)
                    fvel = ((t-16768) % 8)

                    x = torch.LongTensor(final_song).to(device_type)
                    
                    with ctx:
                        out = model.generate(x,
                                             1,
                                             temperature=model_temperature,
                                             filter_logits_fn=top_p,
                                             filter_kwargs={'thres': model_sampling_top_p},
                                             return_prime=False,
                                             eos_token=18818,
                                             verbose=False)
        
                    y = out.tolist()[0]

                    gdur = ((y-16768) // 8)
                    gvel = ((y-16768) % 8)                    
                    
                    if humanize_durations:
                        fdur = gdur

                    if humanize_velocities:
                        fvel = gvel
                        
                    dur_vel_tok = ((8 * fdur) + fvel) + 16768
                    
                    final_song.append(dur_vel_tok)

                    hn_count += 1

                    if hn_count == num_hum_notes:
                        break

            #==================================================================
           
            print('=' * 70)
            print('Done!')
            print('=' * 70)
            
            #===============================================================================
            
            print('Rendering results...')
            
            print('=' * 70)
            print('Sample INTs', final_song[:15])
            print('=' * 70)
        
            song_f = []
            
            if len(final_song) != 0:
            
                time = 0
                dur = 1
                vel = 90
                pitch = 60
                channel = 0
                patch = 0
            
                patches = [-1] * 16
            
                channels = [0] * 16
                channels[9] = 1
            
                for ss in final_song:
            
                    if 0 <= ss < 256:
            
                        time += ss * 16
            
                    if 256 <= ss < 16768:
            
                        patch = (ss-256) // 128
            
                        if patch < 128:
            
                            if patch not in patches:
                              if 0 in channels:
                                  cha = channels.index(0)
                                  channels[cha] = 1
                              else:
                                  cha = 15
            
                              patches[cha] = patch
                              channel = patches.index(patch)
                            else:
                              channel = patches.index(patch)
            
                        if patch == 128:
                            channel = 9
            
                        pitch = (ss-256) % 128
            
            
                    if 16768 <= ss < 18816:
            
                        dur = ((ss-16768) // 8) * 16
                        vel = (((ss-16768) % 8)+1) * 15
            
                        song_f.append(['note', time, dur, channel, pitch, vel, patch])
            
                patches = [0 if x==-1 else x for x in patches]

            output_score, patches, overflow_patches = TMIDIX.patch_enhanced_score_notes(song_f)
        
            fn1 = "Orpheus-Humanizing-Transformer-Composition"
            
            detailed_stats = TMIDIX.Tegridy_ms_SONG_to_MIDI_Converter(output_score,
                                                                      output_signature = 'Orpheus Humanizing Transformer',
                                                                      output_file_name = fn1,
                                                                      track_name='Project Los Angeles',
                                                                      list_of_MIDI_patches=patches
                                                                      )
            
            new_fn = fn1+'.mid'
                    
            
            audio = midi_to_colab_audio(new_fn, 
                                soundfont_path=SOUDFONT_PATH,
                                sample_rate=16000,
                                volume_scale=10,
                                output_for_gradio=True
                                )
            
            print('Done!')
            print('=' * 70)
        
            #========================================================
        
            output_midi = str(new_fn)
            output_audio = (16000, audio)
            output_plot = TMIDIX.plot_ms_SONG(song_f, plot_title=output_midi, return_plt=True)
        
            print('Output MIDI file name:', output_midi)
            print('=' * 70) 
            
            #========================================================
    
        else:
            return None, None, None
    
        print('-' * 70)
        print('Req end time: {:%Y-%m-%d %H:%M:%S}'.format(datetime.datetime.now(PDT)))
        print('-' * 70)
        print('Req execution time:', (reqtime.time() - start_time), 'sec')
    
        return output_audio, output_plot, output_midi

    else:
        return None, None, None
    
#==================================================================================

PDT = timezone('US/Pacific')

print('=' * 70)
print('App start time: {:%Y-%m-%d %H:%M:%S}'.format(datetime.datetime.now(PDT)))
print('=' * 70)

#==================================================================================

with gr.Blocks() as demo:

    #==================================================================================

    gr.Markdown("<h1 style='text-align: left; margin-bottom: 1rem'>Orpheus Humanizing Transformer</h1>")
    gr.Markdown("<h1 style='text-align: left; margin-bottom: 1rem'>Humanize durations and/or velocities in any MIDI score</h1>")
    gr.HTML("""            
            <p> 
                <a href="https://huggingface.co/spaces/projectlosangeles/Orpheus-Humanizing-Transformer?duplicate=true">
                    <img src="https://huggingface.co/datasets/huggingface/badges/resolve/main/duplicate-this-space-md.svg" alt="Duplicate in Hugging Face">
                </a>
            </p>
            
            for faster execution and endless generation!
            """)
    
    #==================================================================================
    
    gr.Markdown("## Upload source MIDI or select a sample MIDI on the bottom of the page")
    
    input_midi = gr.File(label="Input MIDI", 
                         file_types=[".midi", ".mid", ".kar"]
                        )
    
    gr.Markdown("## Generation options")
    
    humanize_durations = gr.Checkbox(value=False, label="Humanize durations")
    humanize_velocities = gr.Checkbox(value=True, label="Humanize velocities")
    
    num_prime_toks = gr.Slider(0, 1024, value=0, step=1, label="Number of prime tokens")
    num_hum_notes = gr.Slider(128, 2048, value=512, step=1, label="Number of notes to humanize")

    model_temperature = gr.Slider(0.1, 1, value=0.9, step=0.01, label="Model temperature")
    model_sampling_top_p = gr.Slider(0.1, 0.99, value=0.96, step=0.01, label="Model sampling top p value")
    
    generate_btn = gr.Button("Generate", variant="primary")

    gr.Markdown("## Generation results")

    output_title = gr.Textbox(label="MIDI melody title")
    output_audio = gr.Audio(label="MIDI audio", format="wav", elem_id="midi_audio")
    output_plot = gr.Plot(label="MIDI score plot")
    output_midi = gr.File(label="MIDI file", file_types=[".mid"])

    generate_btn.click(Humanize_MIDI, 
                       [input_midi,
                        num_prime_toks,
                        num_hum_notes,
                        humanize_durations,
                        humanize_velocities,
                        model_temperature,
                        model_sampling_top_p
                       ], 
                       [output_audio,
                        output_plot,
                        output_midi                          
                       ]
                      )

    gr.Examples(
                [["Sharing The Night Together.kar", 0, 1024, False, True, 0.9, 0.96]
                ],
                [input_midi,
                 num_prime_toks,
                 num_hum_notes,
                 humanize_durations,
                 humanize_velocities,
                 model_temperature,
                 model_sampling_top_p
                ],
                [output_audio,
                 output_plot,
                 output_midi
                ],
                Humanize_MIDI
    )
    
#==================================================================================

demo.launch()

#==================================================================================