#============================================================================================
# https://huggingface.co/spaces/projectlosangeles/Orpheus-MIDI-Loops-Generator
#============================================================================================
print('=' * 70)
print('Orpheus MIDI Loops Generator Gradio App')
print('=' * 70)
print('Loading core Orpheus MIDI Loops Generator modules...')
import os
import copy
import time as reqtime
import datetime
from pytz import timezone
print('=' * 70)
print('Loading main Orpheus MIDI Loops Generator 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 MIDI Loops Generator 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_Loops_Fine_Tuned_Model_3441_steps_0.7715_loss_0.7992_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 MIDI Loops dataset...')
orpheus_loops_dataset_file = hf_hub_download(repo_id='asigalov61/Orpheus-Music-Transformer',
filename='orpheus_data/230414_Select_Orpheus_MIDI_Loops_Dataset_CC_BY_NC_SA.pickle'
)
loops_data = TMIDIX.Tegridy_Any_Pickle_File_Reader(orpheus_loops_dataset_file)
print('=' * 70)
print('Done!')
print('=' * 70)
print('Loaded', len(loops_data), 'loops')
print('=' * 70)
#==================================================================================
def tokens_to_score(tokens):
song_f = []
time = 0
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
#==================================================================================
@spaces.GPU
def generate_loops(start_loop_seq,
num_prime_toks,
num_loops_to_generate,
model_temperature,
model_sampling_top_p
):
prime_seq = start_loop_seq[:num_prime_toks]
x = torch.LongTensor([prime_seq] * num_loops_to_generate).cuda()
with ctx:
out = model.generate(x,
SEQ_LEN-x.shape[1],
temperature=model_temperature,
filter_logits_fn=top_p,
filter_kwargs={'thres': model_sampling_top_p},
return_prime=True,
eos_token=18818,
verbose=True)
y = out.tolist()
outputs = []
for seq in y:
try:
eidx = seq.index(18818)+1
if len(seq[:eidx]) - seq.index(18817) == 122:
outputs.append(seq[:eidx-120])
except:
continue
sidx = start_loop_seq.index(18817)+2
song = start_loop_seq[:sidx]
for o in outputs:
song.extend(o[2:])
#==============================================================================
print('=' * 70)
print('Done!')
print('=' * 70)
print('Song has', len(outputs)+1, 'loops')
print('=' * 70)
return song
#==================================================================================
def Generate_MIDI_Loops(num_loops_to_generate,
num_prime_toks,
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)
print('Num loops to generate:', num_loops_to_generate)
print('Num of prime toks:', num_prime_toks)
print('Model temperature:', model_temperature)
print('Model top k:', model_sampling_top_p)
print('=' * 70)
#==================================================================
print('Generating...')
#==============================================================================
start_loop_idx = random.randint(0, len(loops_data))
start_loop = loops_data[start_loop_idx]
#==============================================================================
print('=' * 70)
print('Song:', start_loop[0])
print('Artist:', start_loop[1])
print('=' * 70)
#==============================================================================
song = generate_loops(start_loop[2],
num_prime_toks,
num_loops_to_generate,
model_temperature,
model_sampling_top_p
)
print('=' * 70)
print('Done!')
print('=' * 70)
#===============================================================================
print('Rendering results...')
#===============================================================================
print('=' * 70)
print('Sample INTs', song[:15])
print('=' * 70)
#===============================================================================
output_score = tokens_to_score(song)
#===============================================================================
patched_score, patches, overflow_patches = TMIDIX.patch_enhanced_score_notes(output_score)
fn1 = "Orpheus-MIDI-Loops-Generator-Composition"
detailed_stats = TMIDIX.Tegridy_ms_SONG_to_MIDI_Converter(patched_score,
output_signature = 'Orpheus MIDI Loops Generator',
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,
output_for_gradio=True
)
#===============================================================================
print('Done!')
print('=' * 70)
#========================================================
output_title_artist = 'Song title: ' + start_loop[0] + '\n'
output_title_artist += 'Artist: ' + start_loop[1]
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(output_title_artist)
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_artist, 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 MIDI Loops Generator
")
gr.Markdown("Generate awesome MIDI loops!
")
gr.HTML("""
for faster execution and endless generation!
""")
#==================================================================================
gr.Markdown("## Generation options")
num_loops_to_generate = gr.Slider(2, 32, value=16, step=1, label="Number of loops to generate")
num_prime_toks = gr.Slider(32, 256, value=128, step=1, label="Number of prime tokens")
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 Loops", variant="primary")
gr.Markdown("## Generation results")
output_title_artist = gr.Textbox(label="MIDI loops title/artist", lines=2)
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_MIDI_Loops,
[num_loops_to_generate,
num_prime_toks,
model_temperature,
model_sampling_top_p
],
[output_title_artist,
output_audio,
output_plot,
output_midi
]
)
#==================================================================================
demo.launch()
#==================================================================================