#============================================================================================
# https://huggingface.co/spaces/projectlosangeles/Orpheus-Mono-Melodies-Mixer
#============================================================================================
print('=' * 70)
print('Orpheus Mono Melodies Mixer Gradio App')
print('=' * 70)
print('Loading core Orpheus Mono Melodies Mixer modules...')
import os
import copy
import time as reqtime
import datetime
from pytz import timezone
print('=' * 70)
print('Loading main Orpheus Mono Melodies Mixer 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 Mono Melodies Mixer 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_Bridge_Music_Transformer_Trained_Model_43450_steps_0.8334_loss_0.7629_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 = 1668
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)
#==================================================================================
print('=' * 70)
print('Loading Orpheus Mono Melodies dataset...')
orpheus_mono_mels_dataset_file = hf_hub_download(repo_id='asigalov61/Orpheus-Music-Transformer',
filename='orpheus_data/58870_Orpheus_Mono_Melodies_40_88_Dataset_CC_BY_NC_SA.pickle'
)
mono_mels_data = TMIDIX.Tegridy_Any_Pickle_File_Reader(orpheus_mono_mels_dataset_file)
mono_mels_data = [[str(l[0]), list(l[1])] + l[2:] for l in mono_mels_data]
print('=' * 70)
print('Done!')
print('=' * 70)
print('Loaded', len(mono_mels_data), 'mels')
print('=' * 70)
#==================================================================================
def tokens_to_score(tokens, abs_time):
song_f = []
time = abs_time
dur = 1
vel = 90
pitch = 60
channel = 0
patch = 0
patches = [-1] * 16
channels = [0] * 16
channels[9] = 1
for ss in tokens:
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])
return song_f, time
#==================================================================================
@spaces.GPU
def Mix_Mono_Melodies(num_mels_to_mix,
use_one_mel,
mel_patch_num,
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)
print('Num mels to mix:', num_mels_to_mix)
print('Use one mel:', use_one_mel)
print('Melody MIDI patch number:', mel_patch_num)
print('Model temperature:', model_temperature)
print('Model top k:', model_sampling_top_k)
print('=' * 70)
#==================================================================
print('Generating...')
song = []
song_indexes = []
song_titles = []
song_parts = []
while len(song) <= 512:
lidx = random.randint(0, len(mono_mels_data)-1)
song = mono_mels_data[lidx][1]
song_indexes.append(lidx)
song_titles.append(mono_mels_data[lidx][0])
song_parts.append(mono_mels_data[lidx][1])
for i in tqdm.tqdm(range(num_mels_to_mix-1)):
left_chunk = [1] + mono_mels_data[lidx][1][2:]
if use_one_mel:
right_chunk = [1] + mono_mels_data[lidx][1][2:]
else:
right_chunk = []
ridx = [-1]
rlen = -1
while ridx and rlen <= 512:
rlen = len(mono_mels_data[ridx[0]][1])
ridx = [l for l in mono_mels_data[lidx][2] if l not in song_indexes]
if ridx:
ridx = ridx[0]
right_chunk = [1] + mono_mels_data[ridx][1][2:]
lidx = ridx
song_titles.append(mono_mels_data[lidx][0])
song_indexes.append(lidx)
else:
break
seq = [18815] + left_chunk[-512:] + [18816] + right_chunk[:512] + [18817] + left_chunk[-64:]
x = torch.LongTensor(seq).cuda()
y_val = []
rcount = 0
while y_val != right_chunk[:64]:
with ctx:
out = model.generate(x,
576,
temperature=model_temperature,
filter_logits_fn=top_k,
filter_kwargs={'k': model_sampling_top_k},
eos_token=18818,
return_prime=False,
verbose=False)
y = out.tolist()
y_val = y[-64:]
if y_val != right_chunk[:64]:
rcount += 1
print('Regenerating attempt #', rcount)
if rcount == 3:
break
song = song + y[:-64] + right_chunk
song_parts.append(y[:-64])
song_parts.append(right_chunk)
#==================================================================
print('=' * 70)
print('Done!')
print('=' * 70)
#===============================================================================
print('Rendering results...')
used_mels_titles = 'Composition uses ' + str(len(song_titles)) + ' monophonic melodies with the following indexes:\n\n'
for i, t in enumerate(song_titles):
used_mels_titles += 'Melody #' + str(i+1) + ': ' + str(t) + '\n'
#===============================================================================
print('=' * 70)
print('Sample INTs', song[:15])
print('=' * 70)
output_score = []
abs_time = 1000
for i, part in enumerate(song_parts):
if i == 0:
part = part[1:]
if not use_one_mel:
part_idx = song_indexes[i // 2]
else:
part_idx = song_indexes[0]
if i % 2 == 0:
output_score.append(['text_event', abs_time + (part[0] * 16), 'Melody #' + str((i // 2)+1) + ' / IDX #' + str(part_idx)])
else:
tidx = [i for i in range(20) if part[i] < 256][0]
output_score.append(['text_event', abs_time + (part[tidx] * 16), 'AI-generated bridge'])
score, abs_time= tokens_to_score(part, abs_time)
output_score.extend(score)
#===============================================================================
TMIDIX.adjust_score_velocities(output_score, 100, adj_per_channel=True)
#===============================================================================
for e in output_score:
if e[0] == 'note':
if e[6] == 40:
e[5] = 103 + (e[4] % 24)
e[6] = mel_patch_num
#===============================================================================
patched_score, patches, overflow_patches = TMIDIX.patch_enhanced_score_notes(output_score,
reserved_patch=mel_patch_num,
reserved_patch_channel=3
)
fn1 = "Orpheus-Mono-Melodies-Mixer-Composition"
detailed_stats = TMIDIX.Tegridy_ms_SONG_to_MIDI_Converter(patched_score,
output_signature = 'Orpheus Mono Melodies Mixer',
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(patched_score,
plot_title=output_midi,
return_plt=True
)
#===============================================================================
print(used_mels_titles)
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 used_mels_titles, 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("Orpheus Mono Melodies Mixer
")
gr.Markdown("Mix several monophonic melodies into one composition by bridging
")
gr.HTML("""
for faster execution and endless generation!
""")
#==================================================================================
gr.Markdown("## Generation options")
num_mels_to_mix = gr.Slider(2, 8, value=5, step=1, label="Number of melodies to mix")
use_one_mel = gr.Checkbox(value=False, label="Use only one randomly selected melody")
mel_patch_num = gr.Slider(0, 127, value=40, step=1, label="Melody MIDI patch number")
model_temperature = gr.Slider(0.1, 1, value=1.0, step=0.01, label="Model temperature")
model_sampling_top_k = gr.Slider(1, 100, value=5, step=1, label="Model sampling top k value")
generate_btn = gr.Button("Mix Melodies", variant="primary")
gr.Markdown("## Generation results")
used_mels_titles = gr.Textbox(label="MIDI summary")
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(Mix_Mono_Melodies,
[num_mels_to_mix,
use_one_mel,
mel_patch_num,
model_temperature,
model_sampling_top_k
],
[used_mels_titles,
output_audio,
output_plot,
output_midi
]
)
#==================================================================================
demo.launch()
#==================================================================================