app.py

#=========================================================================
# https://huggingface.co/spaces/asigalov61/Score-2-Performance-Transformer
#=========================================================================

import os
import time as reqtime
import datetime
from pytz import timezone

import copy
from itertools import groupby
import tqdm

import spaces
import gradio as gr

import torch
from x_transformer_1_23_2 import *
import random

import TMIDIX

from midi_to_colab_audio import midi_to_colab_audio

from huggingface_hub import hf_hub_download

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

print('Loading model...')

SEQ_LEN = 1802
PAD_IDX = 771
DEVICE = 'cuda' # 'cpu'

# instantiate the model

model = TransformerWrapper(
    num_tokens = PAD_IDX+1,
    max_seq_len = SEQ_LEN,
    attn_layers = Decoder(dim = 1024, 
                          depth = 8, 
                          heads = 8, 
                          rotary_pos_emb=True, 
                          attn_flash = True
                         )
    )

model = AutoregressiveWrapper(model, ignore_index = PAD_IDX)


print('=' * 70)

print('Loading model checkpoint...')

model_checkpoint = hf_hub_download(repo_id='asigalov61/Score-2-Performance-Transformer', 
                                   filename='Score_2_Performance_Transformer_Final_Small_Trained_Model_4496_steps_1.5185_loss_0.5589_acc.pth'
                                  )

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

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

dtype = torch.bfloat16

ctx = torch.amp.autocast(device_type=DEVICE, dtype=dtype)

print('=' * 70)
print('Done!')
print('=' * 70)

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

def load_midi(midi_file):

    print('Loading MIDI...')

    raw_score = TMIDIX.midi2single_track_ms_score(midi_file)
    
    escore_notes = TMIDIX.advanced_score_processor(raw_score, return_enhanced_score_notes=True)
    
    if escore_notes[0]:
    
        escore_notes = TMIDIX.augment_enhanced_score_notes(escore_notes[0], timings_divider=16)
    
        pe = escore_notes[0]
    
        melody_chords = []
    
        seen = []
    
        for e in escore_notes:
    
            if e[3] != 9:
        
                #=======================================================
        
                dtime = max(0, min(255, e[1]-pe[1]))
        
                if dtime != 0:
                    seen = []
                
                # Durations
                dur = max(1, min(255, e[2]))
        
                # Pitches
                ptc = max(1, min(127, e[4]))
                
                vel = max(1, min(127, e[5]))
        
                if ptc not in seen:
        
                    melody_chords.append([dtime, dur, ptc, vel])
        
                    seen.append(ptc)
        
                pe = e
    
    print('=' * 70)
    print('Number of notes in a composition:', len(melody_chords))
    print('=' * 70)

    src_melody_chords_f = []
    
    for i in range(0, len(melody_chords), 150):
        
        chunk = melody_chords[i:i+300]
        
        src = []
        
        for mm in chunk:
            src.append([mm[0], mm[2]+256, mm[1]+384, mm[3]+640])
    
        clen = len(src)
    
        if clen < 300:
    
            chunk_mult = (300 // clen) + 1
    
            src += src * chunk_mult
    
        src_melody_chords_f.append([clen, src[:300]])
            
    print('Done!')
    print('=' * 70)
    print('Number of composition chunks:', len(src_melody_chords_f))
    print('=' * 70)

    return src_melody_chords_f

# =================================================================================================
                       
@spaces.GPU
def Convert_Score_to_Performance(input_midi,
                                 input_midi_type,
                                 input_conv_type,
                                 input_number_prime_notes,
                                 input_number_conv_notes,
                                 input_model_dur_top_k,
                                 input_model_dur_temperature,
                                 input_model_vel_temperature
                                ):

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

    fn = os.path.basename(input_midi)
    fn1 = fn.split('.')[0]

    print('=' * 70)
    print('Requested settings:')
    print('=' * 70)
    print('Input MIDI file name:', fn)
    print('Input MIDI type:', input_midi_type)
    print('Conversion type:', input_conv_type)
    print('Number of prime notes:', input_number_prime_notes)
    print('Number of notes to convert:', input_number_conv_notes)
    print('Model durations sampling top value:', input_model_dur_top_k)
    print('Model durations temperature:', input_model_dur_temperature)
    print('Model velocities temperature:', input_model_vel_temperature)
    
    print('=' * 70)

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

    src_melody_chords_f = load_midi(input_midi.name)

    #==================================================================
    
    print('Sample output events', src_melody_chords_f[0][1][:3])
    print('=' * 70)
    print('Generating...')

    model.to(DEVICE)
    model.eval()

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

    num_prime_notes = input_number_prime_notes # Priming improves the results but it is not necessary and you can set it to zero
    dur_top_k = input_model_dur_top_k # Use k == 1 if src composition is score and k > 1 if src composition is performance
    
    dur_temperature = input_model_dur_temperature # For best results, durations temperature should be more than 1.0 but less than velocities temperature
    vel_temperature = input_model_vel_temperature # For best results, velocities temperature must be larger than 1.3 and larger than durations temperature
    
    #==================================================================
    
    if input_midi_type == 'Score':
        
        dur_top_k = 1
        dur_temperature = 1.1
        vel_temperature = 1.5
    
    elif input_midi_type == 'Performance':
        
        dur_top_k = 100
        dur_temperature = 1.5
        vel_temperature = 1.9
    
    else:
        
        dur_top_k = input_model_dur_top_k # Use k == 1 if src composition is score and k > 1 if src composition is performance
        
        dur_temperature = input_model_dur_temperature # For best results, durations temperature should be more than 1.0 but less than velocities temperature
        vel_temperature = input_model_vel_temperature
    
    final_song = []
    
    for cc, (song_chunk_len, song_chunk) in enumerate(src_melody_chords_f):
    
        print('=' * 70)
        print('Rendering song chunk #', cc)
        print('=' * 70)
    
        #========================================================================
    
        song = [768]
    
        if cc == 0:
    
            for m in song_chunk:
                song.extend(m[:2])
            
            song.append(769)
    
            sidx = 0
            eidx = 300
    
        else:
            for m in song_chunk[:150]:
                psrc.extend(m[:2])
    
            psrc.append(769)
    
            song = copy.deepcopy(psrc + ptrg)
    
            sidx = 150
            eidx = 300
        
        #========================================================================
        
        for i in tqdm.tqdm(range(sidx, eidx)):
        
            song.extend(song_chunk[i][:2])
    
            if 'Durations' in input_conv_type:
        
                if i < num_prime_notes and cc == 0:
                    song.append(song_chunk[i][2])
            
                else:
            
                    # Durations
                
                    x = torch.LongTensor(song).cuda()
            
                    y = 0 
            
                    while not 384 < y < 640:
                    
                        with ctx:
                            out = model.generate(x,
                                                 1,
                                                 temperature=dur_temperature,
                                                 filter_logits_fn=top_k,
                                                 filter_kwargs={'k': dur_top_k},
                                                 return_prime=False,
                                                 verbose=False)
                        
                        y = out.tolist()[0][0]
                
                    song.append(y)
    
            else:
                song.append(song_chunk[i][2])
    
            #========================================================================
    
            if 'Velocities' in input_conv_type:
        
        
                if i < num_prime_notes and cc == 0:
                    song.append(song_chunk[i][3])
            
                else:
            
                    # Velocities
                
                    x = torch.LongTensor(song).cuda()
                    
                    y = 0 
            
                    while not 640 < y < 768:
                            
                        with ctx:
                            out = model.generate(x,
                                                 1,
                                                 temperature=vel_temperature,
                                                 return_prime=False,
                                                 verbose=False)
                        
                        y = out.tolist()[0][0]
                
                    song.append(y)
                    
            else:
                song.append(song_chunk[i][3])
                
        #========================================================================
    
        if cc == 0:
            final_song.extend(song[602:][:(song_chunk_len * 4)])
    
        else:
            final_song.extend(song[602:][600:(song_chunk_len * 4)])
    
        psrc = copy.deepcopy(song[1:301])
        ptrg = copy.deepcopy(song[602:][:600])
    
        #========================================================================
    
        if len(final_song) >= input_number_conv_notes * 4:
            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 = 0
        vel = 90
        pitch = 60
        channel = 0
        patch = 0
        
        patches = [0] * 16
        
        for ss in final_song:
        
            if 0 <= ss < 256:
        
                time += ss * 16
        
            if 256 <= ss < 384:
        
                pitch = ss-256
        
            if 384 <= ss < 640:
        
                dur = (ss-384) * 16
        
            if 640 <= ss < 768:
                
                vel = (ss-640)
            
                song_f.append(['note', time, dur, channel, pitch, vel, patch])

    fn1 = "Score-2-Performance-Transformer-Composition"
    
    detailed_stats = TMIDIX.Tegridy_ms_SONG_to_MIDI_Converter(song_f,
                                                              output_signature = 'Score 2 Performance 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=soundfont,
                        sample_rate=16000,
                        volume_scale=10,
                        output_for_gradio=True
                        )
    
    print('Done!')
    print('=' * 70)

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

    output_midi_title = str(fn1)
    output_midi_summary = str(song_f[:3])
    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('Output MIDI title:', output_midi_title)
    print('Output MIDI summary:', output_midi_summary)
    print('=' * 70) 
    

    #========================================================
    
    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_midi_title, output_midi_summary, output_midi, output_audio, output_plot

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

if __name__ == "__main__":
    
    PDT = timezone('US/Pacific')
    
    print('=' * 70)
    print('App start time: {:%Y-%m-%d %H:%M:%S}'.format(datetime.datetime.now(PDT)))
    print('=' * 70)

    soundfont = "SGM-v2.01-YamahaGrand-Guit-Bass-v2.7.sf2"

    app = gr.Blocks()
    
    with app:
        
        gr.Markdown("<h1 style='text-align: center; margin-bottom: 1rem'>Score 2 Performance Transformer</h1>")
        gr.Markdown("<h1 style='text-align: center; margin-bottom: 1rem'>Convert any MIDI score to a nice performance</h1>")

        gr.Markdown("## Upload your MIDI or select a sample example MIDI below")
        
        input_midi = gr.File(label="Input MIDI", file_types=[".midi", ".mid", ".kar"])

        gr.Markdown("## Select MIDI type")

        input_midi_type = gr.Radio(["Score", "Performance", "Custom"], 
                                   value="Score", 
                                   label="Input MIDI type",
                                   info="Select 'Custom' option to enable model top_k and temperature settings below"
                                  )
        gr.Markdown("## Select conversion type")

        input_conv_type = gr.Radio(["Durations and Velocities", "Durations", "Velocities"], 
                                   value="Durations and Velocities", 
                                   label="Conversion type"
                                  )

        gr.Markdown("## Conversion options")

        input_number_prime_notes = gr.Slider(0, 512, value=0, step=8, label="Number of prime notes")
        input_number_conv_notes = gr.Slider(8, 2048, value=512, step=8, label="Number of notes to convert")

        gr.Markdown("## Custom MIDI type model options")

        input_model_dur_top_k = gr.Slider(1, 100, value=1, step=1, label="Model sampling top k value for durations")
        input_model_dur_temperature = gr.Slider(0.5, 1.5, value=1.1, step=0.05, label="Model temperature for durations")
        input_model_vel_temperature = gr.Slider(0.5, 1.5, value=1.5, step=0.05, label="Model temperature for velocities")
        
        run_btn = gr.Button("convert", variant="primary")

        gr.Markdown("## Generation results")

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

        run_event = run_btn.click(Convert_Score_to_Performance, [input_midi,
                                                                 input_midi_type,
                                                                 input_conv_type,
                                                                 input_number_prime_notes,
                                                                 input_number_conv_notes,
                                                                 input_model_dur_top_k,
                                                                 input_model_dur_temperature,
                                                                 input_model_vel_temperature
                                                                ],
                                  [output_midi_title, output_midi_summary, output_midi, output_audio, output_plot])

        gr.Examples(
            [["asap_midi_score_21.mid", "Score", "Durations and Velocities", 8, 600, 1, 1.1, 1.5],
             ["asap_midi_score_45.mid", "Score", "Durations and Velocities", 8, 600, 1, 1.1, 1.5],
             ["asap_midi_score_69.mid", "Score", "Durations and Velocities", 8, 600, 1, 1.1, 1.5],
             ["asap_midi_score_118.mid", "Score", "Durations and Velocities", 8, 600, 1, 1.1, 1.5],
             ["asap_midi_score_167.mid", "Score", "Durations and Velocities", 8, 600, 1, 1.1, 1.5],
            ],
            [input_midi,
             input_midi_type,
             input_conv_type,
             input_number_prime_notes,
             input_number_conv_notes,
             input_model_dur_top_k,
             input_model_dur_temperature,
             input_model_vel_temperature
            ],
            [output_midi_title, output_midi_summary, output_midi, output_audio, output_plot],
            Convert_Score_to_Performance
        )
        
        app.queue().launch()