app.py

#==================================================================================
# https://huggingface.co/spaces/asigalov61/Guided-Accompaniment-Transformer
#==================================================================================

print('=' * 70)
print('Guided Accompaniment Transformer Gradio App')

print('=' * 70)
print('Loading core Guided Accompaniment Transformer modules...')

import os
import copy

import time as reqtime
import datetime
from pytz import timezone

print('=' * 70)
print('Loading main Guided Accompaniment 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_mem_efficient_sdp(True)
torch.backends.cuda.enable_math_sdp(True)
torch.backends.cuda.enable_flash_sdp(True)
torch.backends.cuda.enable_cudnn_sdp(True)

from huggingface_hub import hf_hub_download

import TMIDIX

from midi_to_colab_audio import midi_to_colab_audio

from x_transformer_1_23_2 import *

import random

from collections import defaultdict

import tqdm

print('=' * 70)
print('Loading aux Guided Accompaniment 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 = 'Guided_Accompaniment_Transformer_Trained_Model_36457_steps_0.5384_loss_0.8417_acc.pth'

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

MAX_MELODY_NOTES = 72

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

print('=' * 70)
print('Loading popular hook melodies dataset...')

popular_hook_melodies_pickle = hf_hub_download(repo_id='asigalov61/Guided-Accompaniment-Transformer', 
                                               filename='popular_hook_melodies_24_64_CC_BY_NC_SA.pickle'
                                              )

popular_hook_melodies = TMIDIX.Tegridy_Any_Pickle_File_Reader(popular_hook_melodies_pickle)

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

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

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 = 4096
PAD_IDX = 1794

model = TransformerWrapper(
        num_tokens = PAD_IDX+1,
        max_seq_len = SEQ_LEN,
        attn_layers = Decoder(dim = 2048,
                              depth = 4,
                              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/Guided-Accompaniment-Transformer', filename=MODEL_CHECKPOINT)

model.load_state_dict(torch.load(model_checkpoint, map_location='cpu', 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, melody_patch=-1):

    raw_score = TMIDIX.midi2single_track_ms_score(input_midi)

    escore_notes = TMIDIX.advanced_score_processor(raw_score, return_enhanced_score_notes=True)[0]
    escore_notes = TMIDIX.augment_enhanced_score_notes(escore_notes, timings_divider=32)
    
    sp_escore_notes = TMIDIX.solo_piano_escore_notes(escore_notes)

    if melody_patch == -1:
        zscore = TMIDIX.recalculate_score_timings(sp_escore_notes)

    else:
        mel_score = [e for e in sp_escore_notes if e[6] == melody_patch]

        if mel_score:
            zscore = TMIDIX.recalculate_score_timings(mel_score)

        else:
            zscore = TMIDIX.recalculate_score_timings(sp_escore_notes)
    
    cscore = TMIDIX.chordify_score([1000, zscore])[:MAX_MELODY_NOTES]
    
    score = []
    
    pc = cscore[0]
    
    for c in cscore:
        score.append(max(0, min(127, c[0][1]-pc[0][1])))

        n = c[0]
        
        score.extend([max(1, min(127, n[2]))+128, max(1, min(127, n[4]))+256])
    
        pc = c
        
    score_len = len(score) // 3
    score_ptcs = [t-256 for t in score if t > 256]
    
    return score, score_len, score_ptcs

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

def tokens_to_scores(tokens, mel_len):

    time = 0
    dur = 0
    vel = 90
    pitch = 0
    channel = 0
    patch = 0

    channels_map = [0, 1, 2, 3, 4, 5, 6, 7, 8, 10, 11, 9, 12, 13, 14, 15]
    patches_map = [40, 0, 10, 19, 24, 35, 40, 52, 56, 9, 65, 73, 0, 0, 0, 0]
    velocities_map = [125, 80, 100, 80, 90, 100, 100, 80, 110, 110, 110, 110, 80, 80, 80, 80]

    melody = []
    accompaniment = []

    mel_notes_count = 0
    final_time = 9999999

    for m in tokens:

        if 0 <= m < 128:
            time += m * 32

        elif 128 < m < 256:
            dur = (m-128) * 32

        elif 256 < m < 1792:
            cha = (m-256) // 128
            pitch = (m-256) % 128

            channel = channels_map[cha]
            patch = patches_map[channel]
            vel = velocities_map[channel]

            if time > final_time:
                break

            if channel == 0:    
                melody.append(['note', time, dur, channel, pitch, vel, patch])
                mel_notes_count += 1

                if mel_notes_count == mel_len:
                    final_time = time

            else:                    
                accompaniment.append(['note', time, dur, channel, pitch, vel, patch])

    return melody, accompaniment

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

def check_tones(mel, acc):

    mel_times = [e[1] for e in mel]
    mel_tones = [e[4] % 12 for e in mel]

    groups = defaultdict(list)
    
    for sub in acc:
        key = sub[1]
        if key in mel_times:
            groups[key].append(sub)

    matches = []
    
    for k, v in groups.items():
        ktone = mel_tones[mel_times.index(k)]
        vtones = sorted(set([e[4] % 12 for e in v]))
    
        if ktone in vtones:
            matches.append(True)
    
        else:
            matches.append(False)

    return sum(matches)

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

def check_seq(seq, mel_len, mel_ptcs):

    mel, acc = tokens_to_scores(seq, mel_len)

    trg_mel_ptcs = [e[4] for e in mel]

    ptcs_good = mel_ptcs == trg_mel_ptcs

    mel_times = [e[1] for e in mel]
    acc_times = [e[1] for e in acc if e[3] != 9]

    times_good = all(mel_times[i] in acc_times for i in range(len(mel_times)))

    if ptcs_good and times_good:
        return mel, acc

    else:
        return None, None

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

@spaces.GPU
def generate_sequences(score_seq,
                       score_len, 
                       temperature=0.9, 
                       top_k_value=15,
                       num_batches=48,
                       verbose=True
                      ):

    x = torch.LongTensor([score_seq] * num_batches).cuda()
    
    with ctx:
        out = model.generate(x,
                             32*score_len,
                             filter_logits_fn=top_k,
                             filter_kwargs={'k': top_k_value},
                             temperature=temperature,
                             return_prime=False,
                             verbose=verbose)
    
    output = out.tolist()

    return output

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

def Generate_Accompaniment(input_midi,
                           input_melody,
                           melody_patch,
                           model_temperature,
                           model_sampling_top_k
                          ):

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

    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)
    if input_midi:
        fn = os.path.basename(input_midi)
        fn1 = fn.split('.')[0]
        print('Input MIDI file name:', fn)

    else:
        print('Input sample melody:', input_melody)
    print('Source melody patch:', melody_patch)
    print('Model temperature:', model_temperature)
    print('Model top k:', model_sampling_top_k)
   
    print('=' * 70)

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

    print('Prepping melody...')

    if input_midi:
        inp_mel = 'Custom MIDI'
        score, score_len, score_ptcs = load_midi(input_midi.name, melody_patch)

    else:
        mel_list = [m[0].lower() for m in popular_hook_melodies]

        inp_mel = random.choice(mel_list).title()

        for m in mel_list:
            if input_melody.lower().strip() in m:
                inp_mel = m.title()
                break

        score = popular_hook_melodies[[m[0] for m in popular_hook_melodies].index(inp_mel)][1]
        score_len = len(score) // 3
        score_ptcs = [t-256 for t in score if t > 256]
        
        print('Selected melody:', inp_mel)
        
    print('Sample score events', score[:12])

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

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

    score_seq = [1792] + score + [1793]

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

    output = generate_sequences(score_seq,
                                score_len,
                                temperature=model_temperature,
                                top_k_value=model_sampling_top_k,
                               )

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

    print('Processing generated sequences...')   

    good_scores = []
    
    for seq in output:
        mel, acc = check_seq(seq, score_len, score_ptcs)
    
        if mel and acc:
            if check_tones(mel, acc) / len(mel) >= 0.7:
                score = sorted(mel + acc, key=lambda x: x[1])
                good_scores.append(score)

    print('Number of good scores:', len(good_scores))

    if len(good_scores) > 0:
        final_score = random.choice(good_scores)

    else:
        mel, acc = tokens_to_scores(random.choice(output), score_len)
        final_score = sorted(mel + acc, key=lambda x: x[1])
                
    print('Done!')
    print('=' * 70)
    
    #===============================================================================
    
    print('Rendering results...')

    fn1 = "Guided-Accompaniment-Transformer-Composition"

    patches_map = [40, 0, 10, 19, 24, 35, 40, 52, 56, 9, 65, 73, 0, 0, 0, 0]
    
    detailed_stats = TMIDIX.Tegridy_ms_SONG_to_MIDI_Converter(final_score,
                                                              output_signature = 'Guided Accompaniment Transformer',
                                                              output_file_name = fn1,
                                                              track_name='Project Los Angeles',
                                                              list_of_MIDI_patches=patches_map
                                                              )
    
    new_fn = fn1+'.mid'
            
    
    audio = midi_to_colab_audio(new_fn, 
                        soundfont_path=SOUNDFONT_PATH,
                        sample_rate=16000,
                        output_for_gradio=True
                        )
    
    print('Done!')
    print('=' * 70)

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

    output_title = str(inp_mel)
    output_midi = str(new_fn)
    output_audio = (16000, audio)
    
    output_plot = TMIDIX.plot_ms_SONG(final_score,
                                      plot_title=output_midi,
                                      return_plt=True
                                     )

    print('Output MIDI file name:', output_midi)
    print('Output MIDI melody title:', output_title)
    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_title, output_audio, output_plot, output_midi
    
#==================================================================================

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: center; margin-bottom: 1rem'>Guided Accompaniment Transformer</h1>")
    gr.Markdown("<h1 style='text-align: center; margin-bottom: 1rem'>Guided melody accompaniment generation with transformers</h1>")
    gr.HTML("""            
            <p> 
                <a href="https://huggingface.co/spaces/asigalov61/Guided-Accompaniment-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 melody MIDI or enter a search query for a sample melody below")
    
    input_midi = gr.File(label="Input MIDI", 
                         file_types=[".midi", ".mid", ".kar"]
                        )

    input_melody = gr.Textbox(value="Hotel California", 
                              label="Popular melodies database search query",
                              info='If the query is not found, random melody will be selected. Custom MIDI overrides search query'
                             )
    
    gr.Markdown("## Generation options")
    
    melody_patch = gr.Slider(-1, 127, value=-1, step=1, label="Source melody MIDI patch")
    model_temperature = gr.Slider(0.1, 1, value=0.9, step=0.01, label="Model temperature")
    model_sampling_top_k = gr.Slider(1, 100, value=15, step=1, label="Model sampling top k 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(Generate_Accompaniment, 
                       [input_midi,
                        input_melody,
                        melody_patch,
                        model_temperature,
                        model_sampling_top_k
                       ], 
                       [output_title,
                        output_audio,
                        output_plot,
                        output_midi                          
                       ]
                      )

    gr.Examples(
                [["USSR-National-Anthem-Seed-Melody.mid", "Custom MIDI", -1, 0.9, 15],
                 ["Sparks-Fly-Seed-Melody.mid", "Custom MIDI", -1, 0.9, 15]
                ],
                [input_midi,
                 input_melody,
                 melody_patch,
                 model_temperature,
                 model_sampling_top_k
                ],
                [output_title,
                 output_audio,
                 output_plot,
                 output_midi
                ],
                Generate_Accompaniment
    )
        
#==================================================================================

demo.launch()

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