import os # os.system("pip install ./ort_nightly_gpu-1.17.0.dev20240118002-cp310-cp310-manylinux_2_28_x86_64.whl") os.system("pip install ort-nightly-gpu --index-url=https://aiinfra.pkgs.visualstudio.com/PublicPackages/_packaging/ort-cuda-12-nightly/pypi/simple/") import gc import hashlib import queue import threading import json import shlex import sys import subprocess import librosa import numpy as np import soundfile as sf import torch from tqdm import tqdm from utils import ( remove_directory_contents, create_directories, download_manager, ) import random import spaces from utils import logger import onnxruntime as ort import warnings import spaces import gradio as gr import logging import time warnings.filterwarnings("ignore") title = "
Audio🔹separator
" description = "This demo uses the MDX-Net models for vocal and background sound separation." theme = "NoCrypt/miku" stem_naming = { "Vocals": "Instrumental", "Other": "Instruments", "Instrumental": "Vocals", "Drums": "Drumless", "Bass": "Bassless", } class MDXModel: def __init__( self, device, dim_f, dim_t, n_fft, hop=1024, stem_name=None, compensation=1.000, ): self.dim_f = dim_f self.dim_t = dim_t self.dim_c = 4 self.n_fft = n_fft self.hop = hop self.stem_name = stem_name self.compensation = compensation self.n_bins = self.n_fft // 2 + 1 self.chunk_size = hop * (self.dim_t - 1) self.window = torch.hann_window( window_length=self.n_fft, periodic=True ).to(device) out_c = self.dim_c self.freq_pad = torch.zeros( [1, out_c, self.n_bins - self.dim_f, self.dim_t] ).to(device) def stft(self, x): x = x.reshape([-1, self.chunk_size]) x = torch.stft( x, n_fft=self.n_fft, hop_length=self.hop, window=self.window, center=True, return_complex=True, ) x = torch.view_as_real(x) x = x.permute([0, 3, 1, 2]) x = x.reshape([-1, 2, 2, self.n_bins, self.dim_t]).reshape( [-1, 4, self.n_bins, self.dim_t] ) return x[:, :, : self.dim_f] def istft(self, x, freq_pad=None): freq_pad = ( self.freq_pad.repeat([x.shape[0], 1, 1, 1]) if freq_pad is None else freq_pad ) x = torch.cat([x, freq_pad], -2) # c = 4*2 if self.target_name=='*' else 2 x = x.reshape([-1, 2, 2, self.n_bins, self.dim_t]).reshape( [-1, 2, self.n_bins, self.dim_t] ) x = x.permute([0, 2, 3, 1]) x = x.contiguous() x = torch.view_as_complex(x) x = torch.istft( x, n_fft=self.n_fft, hop_length=self.hop, window=self.window, center=True, ) return x.reshape([-1, 2, self.chunk_size]) class MDX: DEFAULT_SR = 44100 # Unit: seconds DEFAULT_CHUNK_SIZE = 0 * DEFAULT_SR DEFAULT_MARGIN_SIZE = 1 * DEFAULT_SR def __init__( self, model_path: str, params: MDXModel, processor=0 ): # Set the device and the provider (CPU or CUDA) self.device = ( torch.device(f"cuda:{processor}") if processor >= 0 else torch.device("cpu") ) self.provider = ( ["CUDAExecutionProvider"] if processor >= 0 else ["CPUExecutionProvider"] ) self.model = params # Load the ONNX model using ONNX Runtime self.ort = ort.InferenceSession(model_path, providers=self.provider) # Preload the model for faster performance self.ort.run( None, {"input": torch.rand(1, 4, params.dim_f, params.dim_t).numpy()}, ) self.process = lambda spec: self.ort.run( None, {"input": spec.cpu().numpy()} )[0] self.prog = None @staticmethod def get_hash(model_path): try: with open(model_path, "rb") as f: f.seek(-10000 * 1024, 2) model_hash = hashlib.md5(f.read()).hexdigest() except: # noqa model_hash = hashlib.md5(open(model_path, "rb").read()).hexdigest() return model_hash @staticmethod def segment( wave, combine=True, chunk_size=DEFAULT_CHUNK_SIZE, margin_size=DEFAULT_MARGIN_SIZE, ): """ Segment or join segmented wave array Args: wave: (np.array) Wave array to be segmented or joined combine: (bool) If True, combines segmented wave array. If False, segments wave array. chunk_size: (int) Size of each segment (in samples) margin_size: (int) Size of margin between segments (in samples) Returns: numpy array: Segmented or joined wave array """ if combine: # Initializing as None instead of [] for later numpy array concatenation processed_wave = None for segment_count, segment in enumerate(wave): start = 0 if segment_count == 0 else margin_size end = None if segment_count == len(wave) - 1 else -margin_size if margin_size == 0: end = None if processed_wave is None: # Create array for first segment processed_wave = segment[:, start:end] else: # Concatenate to existing array for subsequent segments processed_wave = np.concatenate( (processed_wave, segment[:, start:end]), axis=-1 ) else: processed_wave = [] sample_count = wave.shape[-1] if chunk_size <= 0 or chunk_size > sample_count: chunk_size = sample_count if margin_size > chunk_size: margin_size = chunk_size for segment_count, skip in enumerate( range(0, sample_count, chunk_size) ): margin = 0 if segment_count == 0 else margin_size end = min(skip + chunk_size + margin_size, sample_count) start = skip - margin cut = wave[:, start:end].copy() processed_wave.append(cut) if end == sample_count: break return processed_wave def pad_wave(self, wave): """ Pad the wave array to match the required chunk size Args: wave: (np.array) Wave array to be padded Returns: tuple: (padded_wave, pad, trim) - padded_wave: Padded wave array - pad: Number of samples that were padded - trim: Number of samples that were trimmed """ n_sample = wave.shape[1] trim = self.model.n_fft // 2 gen_size = self.model.chunk_size - 2 * trim pad = gen_size - n_sample % gen_size # Padded wave wave_p = np.concatenate( ( np.zeros((2, trim)), wave, np.zeros((2, pad)), np.zeros((2, trim)), ), 1, ) mix_waves = [] for i in range(0, n_sample + pad, gen_size): waves = np.array(wave_p[:, i:i + self.model.chunk_size]) mix_waves.append(waves) mix_waves = torch.tensor(mix_waves, dtype=torch.float32).to( self.device ) return mix_waves, pad, trim def _process_wave(self, mix_waves, trim, pad, q: queue.Queue, _id: int): """ Process each wave segment in a multi-threaded environment Args: mix_waves: (torch.Tensor) Wave segments to be processed trim: (int) Number of samples trimmed during padding pad: (int) Number of samples padded during padding q: (queue.Queue) Queue to hold the processed wave segments _id: (int) Identifier of the processed wave segment Returns: numpy array: Processed wave segment """ mix_waves = mix_waves.split(1) with torch.no_grad(): pw = [] for mix_wave in mix_waves: self.prog.update() spec = self.model.stft(mix_wave) processed_spec = torch.tensor(self.process(spec)) processed_wav = self.model.istft( processed_spec.to(self.device) ) processed_wav = ( processed_wav[:, :, trim:-trim] .transpose(0, 1) .reshape(2, -1) .cpu() .numpy() ) pw.append(processed_wav) processed_signal = np.concatenate(pw, axis=-1)[:, :-pad] q.put({_id: processed_signal}) return processed_signal def process_wave(self, wave: np.array, mt_threads=1): """ Process the wave array in a multi-threaded environment Args: wave: (np.array) Wave array to be processed mt_threads: (int) Number of threads to be used for processing Returns: numpy array: Processed wave array """ self.prog = tqdm(total=0) chunk = wave.shape[-1] // mt_threads waves = self.segment(wave, False, chunk) # Create a queue to hold the processed wave segments q = queue.Queue() threads = [] for c, batch in enumerate(waves): mix_waves, pad, trim = self.pad_wave(batch) self.prog.total = len(mix_waves) * mt_threads thread = threading.Thread( target=self._process_wave, args=(mix_waves, trim, pad, q, c) ) thread.start() threads.append(thread) for thread in threads: thread.join() self.prog.close() processed_batches = [] while not q.empty(): processed_batches.append(q.get()) processed_batches = [ list(wave.values())[0] for wave in sorted( processed_batches, key=lambda d: list(d.keys())[0] ) ] assert len(processed_batches) == len( waves ), "Incomplete processed batches, please reduce batch size!" return self.segment(processed_batches, True, chunk) @spaces.GPU() def run_mdx( model_params, output_dir, model_path, filename, exclude_main=False, exclude_inversion=False, suffix=None, invert_suffix=None, denoise=False, keep_orig=True, m_threads=2, device_base="cuda", ): if device_base == "cuda": device = torch.device("cuda:0") processor_num = 0 device_properties = torch.cuda.get_device_properties(device) vram_gb = device_properties.total_memory / 1024**3 m_threads = 1 if vram_gb < 8 else (8 if vram_gb > 32 else 2) logger.info(f"threads: {m_threads} vram: {vram_gb}") else: device = torch.device("cpu") processor_num = -1 m_threads = 1 model_hash = MDX.get_hash(model_path) mp = model_params.get(model_hash) model = MDXModel( device, dim_f=mp["mdx_dim_f_set"], dim_t=2 ** mp["mdx_dim_t_set"], n_fft=mp["mdx_n_fft_scale_set"], stem_name=mp["primary_stem"], compensation=mp["compensate"], ) mdx_sess = MDX(model_path, model, processor=processor_num) wave, sr = librosa.load(filename, mono=False, sr=44100) # normalizing input wave gives better output peak = max(np.max(wave), abs(np.min(wave))) wave /= peak if denoise: wave_processed = -(mdx_sess.process_wave(-wave, m_threads)) + ( mdx_sess.process_wave(wave, m_threads) ) wave_processed *= 0.5 else: wave_processed = mdx_sess.process_wave(wave, m_threads) # return to previous peak wave_processed *= peak stem_name = model.stem_name if suffix is None else suffix main_filepath = None if not exclude_main: main_filepath = os.path.join( output_dir, f"{os.path.basename(os.path.splitext(filename)[0])}_{stem_name}.wav", ) sf.write(main_filepath, wave_processed.T, sr) invert_filepath = None if not exclude_inversion: diff_stem_name = ( stem_naming.get(stem_name) if invert_suffix is None else invert_suffix ) stem_name = ( f"{stem_name}_diff" if diff_stem_name is None else diff_stem_name ) invert_filepath = os.path.join( output_dir, f"{os.path.basename(os.path.splitext(filename)[0])}_{stem_name}.wav", ) sf.write( invert_filepath, (-wave_processed.T * model.compensation) + wave.T, sr, ) if not keep_orig: os.remove(filename) del mdx_sess, wave_processed, wave gc.collect() torch.cuda.empty_cache() return main_filepath, invert_filepath MDX_DOWNLOAD_LINK = "https://github.com/TRvlvr/model_repo/releases/download/all_public_uvr_models/" UVR_MODELS = [ "UVR-MDX-NET-Voc_FT.onnx", "UVR_MDXNET_KARA_2.onnx", "Reverb_HQ_By_FoxJoy.onnx", "UVR-MDX-NET-Inst_HQ_4.onnx", ] BASE_DIR = "." # os.path.dirname(os.path.dirname(os.path.abspath(__file__))) mdxnet_models_dir = os.path.join(BASE_DIR, "mdx_models") output_dir = os.path.join(BASE_DIR, "clean_song_output") def convert_to_stereo_and_wav(audio_path): wave, sr = librosa.load(audio_path, mono=False, sr=44100) # check if mono if type(wave[0]) != np.ndarray or audio_path[-4:].lower() != ".wav": # noqa stereo_path = f"{os.path.splitext(audio_path)[0]}_stereo.wav" stereo_path = os.path.join(output_dir, stereo_path) command = shlex.split( f'ffmpeg -y -loglevel error -i "{audio_path}" -ac 2 -f wav "{stereo_path}"' ) sub_params = { "stdout": subprocess.PIPE, "stderr": subprocess.PIPE, "creationflags": subprocess.CREATE_NO_WINDOW if sys.platform == "win32" else 0, } process_wav = subprocess.Popen(command, **sub_params) output, errors = process_wav.communicate() if process_wav.returncode != 0 or not os.path.exists(stereo_path): raise Exception("Error processing audio to stereo wav") return stereo_path else: return audio_path def get_hash(filepath): with open(filepath, 'rb') as f: file_hash = hashlib.blake2b() while chunk := f.read(8192): file_hash.update(chunk) return file_hash.hexdigest()[:18] def random_sleep(): sleep_time = round(random.uniform(5.2, 7.9), 1) time.sleep(sleep_time) def process_uvr_task( orig_song_path: str = "aud_test.mp3", main_vocals: bool = False, dereverb: bool = True, song_id: str = "mdx", # folder output name only_voiceless: bool = False, remove_files_output_dir: bool = False, ): device_base = "cuda" if torch.cuda.is_available() else "cpu" logger.info(f"Device: {device_base}") if remove_files_output_dir: remove_directory_contents(output_dir) with open(os.path.join(mdxnet_models_dir, "data.json")) as infile: mdx_model_params = json.load(infile) song_output_dir = os.path.join(output_dir, song_id) create_directories(song_output_dir) orig_song_path = convert_to_stereo_and_wav(orig_song_path) logger.info(f"onnxruntime device >> {ort.get_device()}") if only_voiceless: logger.info("Voiceless Track Separation...") return run_mdx( mdx_model_params, song_output_dir, os.path.join(mdxnet_models_dir, "UVR-MDX-NET-Inst_HQ_4.onnx"), orig_song_path, suffix="Voiceless", denoise=False, keep_orig=True, exclude_inversion=True, device_base=device_base, ) logger.info("Vocal Track Isolation...") vocals_path, instrumentals_path = run_mdx( mdx_model_params, song_output_dir, os.path.join(mdxnet_models_dir, "UVR-MDX-NET-Voc_FT.onnx"), orig_song_path, denoise=True, keep_orig=True, device_base=device_base, ) if main_vocals: random_sleep() msg_main = "Main Voice Separation from Supporting Vocals..." logger.info(msg_main) gr.Info(msg_main) try: backup_vocals_path, main_vocals_path = run_mdx( mdx_model_params, song_output_dir, os.path.join(mdxnet_models_dir, "UVR_MDXNET_KARA_2.onnx"), vocals_path, suffix="Backup", invert_suffix="Main", denoise=True, device_base=device_base, ) except Exception as e: if "0:00:" in str(e): gr.Info("Waiting 60 seconds for GPU quota") time.sleep(56) random_sleep() backup_vocals_path, main_vocals_path = run_mdx( mdx_model_params, song_output_dir, os.path.join(mdxnet_models_dir, "UVR_MDXNET_KARA_2.onnx"), vocals_path, suffix="Backup", invert_suffix="Main", denoise=True, device_base=device_base, ) else: raise e else: backup_vocals_path, main_vocals_path = None, vocals_path if dereverb: random_sleep() msg_dereverb = "Vocal Clarity Enhancement through De-Reverberation..." logger.info(msg_dereverb) gr.Info(msg_dereverb) try: _, vocals_dereverb_path = run_mdx( mdx_model_params, song_output_dir, os.path.join(mdxnet_models_dir, "Reverb_HQ_By_FoxJoy.onnx"), main_vocals_path, invert_suffix="DeReverb", exclude_main=True, denoise=True, device_base=device_base, ) except Exception as e: if "0:00:" in str(e): gr.Info("Waiting 60 seconds for GPU quota") time.sleep(56) random_sleep() _, vocals_dereverb_path = run_mdx( mdx_model_params, song_output_dir, os.path.join(mdxnet_models_dir, "Reverb_HQ_By_FoxJoy.onnx"), main_vocals_path, invert_suffix="DeReverb", exclude_main=True, denoise=True, device_base=device_base, ) else: raise e else: vocals_dereverb_path = main_vocals_path return ( vocals_path, instrumentals_path, backup_vocals_path, main_vocals_path, vocals_dereverb_path, ) def sound_separate(media_file, stem, main, dereverb): if not media_file: raise ValueError("The audio pls") if not stem: raise ValueError("Select vocal or background...") hash_audio = str(get_hash(media_file)) outputs = [] start_time = time.time() if stem == "vocal": try: _, _, _, _, vocal_audio = process_uvr_task( orig_song_path=media_file, song_id=hash_audio+"mdx", main_vocals=main, dereverb=dereverb, remove_files_output_dir=False, ) outputs.append(vocal_audio) except Exception as error: gr.Info(str(error)) logger.error(str(error)) if stem == "background": background_audio, _ = process_uvr_task( orig_song_path=media_file, song_id=hash_audio+"voiceless", only_voiceless=True, remove_files_output_dir=False, ) # copy_files(background_audio, ".") outputs.append(background_audio) end_time = time.time() execution_time = end_time - start_time logger.info(f"Execution time: {execution_time} seconds") if not outputs: raise Exception("Error in sound separate") return outputs def audio_conf(): return gr.File( label="Audio file", # file_count="multiple", type="filepath", container=True, ) def stem_conf(): return gr.Radio( choices=["vocal", "background"], value="vocal", label="Vocal", # info="", ) def main_conf(): return gr.Checkbox( False, label="Main", # info="", ) def dereverb_conf(): return gr.Checkbox( False, label="Dereverb", # info="", visible=True, ) def button_conf(): return gr.Button( "Inference", variant="primary", ) def output_conf(): return gr.File( label="Result", file_count="multiple", interactive=False, ) def show_vocal_components(input_bool): param = True if input_bool == "vocal" else False return gr.update(visible=param), gr.update( visible=param ) def get_gui(theme): with gr.Blocks(theme=theme) as app: gr.Markdown(title) gr.Markdown(description) aud = audio_conf() with gr.Column(): with gr.Row(): stem_gui = stem_conf() with gr.Column(): with gr.Row(): main_gui = main_conf() dereverb_gui = dereverb_conf() stem_gui.change( show_vocal_components, [stem_gui], [main_gui, dereverb_gui], ) button_base = button_conf() output_base = output_conf() button_base.click( sound_separate, inputs=[ aud, stem_gui, main_gui, dereverb_gui, ], outputs=[output_base], ) gr.Examples( examples=[ [ "./test.mp3", "vocal", False, False, ], ], fn=sound_separate, inputs=[ aud, stem_gui, main_gui, dereverb_gui, ], outputs=[output_base], cache_examples=False, ) return app if __name__ == "__main__": for id_model in UVR_MODELS: download_manager( os.path.join(MDX_DOWNLOAD_LINK, id_model), mdxnet_models_dir ) app = get_gui(theme) app.queue(default_concurrency_limit=40) app.launch( max_threads=40, share=False, show_error=True, quiet=False, debug=False, )