prediction.py from keras.models import load_model import numpy as np import musdb import librosa import librosa.display import math from time import time def P2R(radii, angles): return radii * (math.cos(angles) + 1j*math.sin(angles)) def get_power_and_angle(audio, orig_sr, new_sr, win_len, hop_len): if orig_sr!= new_sr: audio = librosa.core.resample(audio, orig_sr, new_sr) stft_ar = librosa.stft(audio, win_len, hop_len, center = False) return np.abs(stft_ar), np.angle(stft_ar) def normalize_pow_spec(pow_spec): # Масштабирование спектрограммы к [0;1] pow_spec_norm = np.zeros((pow_spec.shape[0], pow_spec.shape[1])) max = np.max(pow_spec) if max != 0: pow_spec_norm = pow_spec / max return pow_spec_norm def save_audio(power_spec, phases, hop_length, win_length, sampling_rate, path, mask = None): if mask is None: mask = np.ones(power_spec.shape) stft_array = np.empty(power_spec.shape, dtype=complex) for i in range(stft_array.shape[0]): for j in range(stft_array.shape[1]): stft_array[i, j] = P2R(power_spec[i][j]*mask[i][j], phases[i][j]) y = librosa.core.istft(stft_array, hop_length, win_length,) librosa.output.write_wav(path, y, sampling_rate) ORIG_SR = 44100 # Оригинальная частота дискретизации SR = 22050 # Частота дискретизации, с которой работает сеть WIN = 1024 # Размер окна дискретного преобразования Фурье HOP = 256 # Размер шага дискретного преобразования Фурье sample_len = 27 # Размер отрезка спектрограммы, на котором обучена сеть TRGT = 'vocals' # Выделяемый источник # Загрузка модели t0 = time() print("<----------[INFO] model loading...") model = load_model('output/model final 4.hdf5') t1 = time() print("<----------[INFO] model was loaded in " + str(round((t1-t0), 1)) + " seconds...") #Загрузка тестового набора mus = musdb.DB(root='./musdb18', subsets='test', split=None) for track in mus: print(track.name) mix = librosa.core.to_mono(track.audio.T) mix_power_spec, mix_phases = get_power_and_angle(mix, ORIG_SR, SR, WIN, HOP) target = librosa.core.to_mono(track.targets[TRGT].audio.T) target_power_spec, target_phases = get_power_and_angle(target, ORIG_SR, SR, WIN, HOP) # Нормализация и расширение спектрограммы mix_power_spec_norm = normalize_pow_spec(mix_power_spec) pad = np.zeros((mix_power_spec_norm.shape[0], sample_len//2)) mix_power_spec_norm = np.concatenate((pad, mix_power_spec_norm), axis=1) mix_power_spec_norm = np.concatenate((mix_power_spec_norm, pad), axis=1) mask = np.zeros((mix_power_spec_norm.shape[0], 0)) bin_mask = np.zeros((mix_power_spec_norm.shape[0], 0)) set_to_zero_mask = np.zeros((mix_power_spec_norm.shape[0], 0)) for i in range(mix_power_spec.shape[1]): mix = mix_power_spec_norm[:,i:i+sample_len] mix = np.expand_dims(mix, 0) mix = np.expand_dims(mix, len(mix.shape)) pred = model.predict(mix)[0] mask = np.concatenate((mask, np.expand_dims(pred, len(pred.shape))), axis=1) for i in range(mask.shape[0]): for j in range(mask.shape[1]): if mask[i][j] >= 0.5: bin_mask[i][j] = 1 set_to_zero_mask[i][j] = mask[i][j] save_audio(mix_power_spec, mix_phases, HOP, WIN, SR, 'output/estimation/soft/'+ track.name +'.wav', mask) save_audio(mix_power_spec, mix_phases, HOP, WIN, SR, 'output/estimation/bin/'+ track.name +'.wav', mask) save_audio(mix_power_spec, mix_phases, HOP, WIN, SR, 'output/estimation/zero/'+ track.name +'.wav', mask) break;

…

_______________________________________________________________________________________________________- predict.py from keras.models import load_model import numpy as np #import musdb import librosa import librosa.display import math from time import time import matplotlib.pyplot as plt def P2R(radii, angles): return radii * (math.cos(angles) + 1j*math.sin(angles)) def get_power_and_angle(audio, orig_sr, new_sr, win_len, hop_len): if orig_sr!= new_sr: audio = librosa.core.resample(audio, orig_sr, new_sr) stft_ar = librosa.stft(audio, win_len, hop_len, center = False) return np.abs(stft_ar), np.angle(stft_ar) def normalize_pow_spec(pow_spec): # Масштабирование спектрограммы к [0;1] pow_spec_norm = np.zeros((pow_spec.shape[0], pow_spec.shape[1])) max = np.max(pow_spec) if max != 0: pow_spec_norm = pow_spec / max return pow_spec_norm def save_audio(power_spec, phases, hop_length, win_length, sampling_rate, path, mask = None): if mask is None: mask = np.ones(power_spec.shape) stft_array = np.empty(power_spec.shape, dtype=complex) for i in range(stft_array.shape[0]): for j in range(stft_array.shape[1]): stft_array[i, j] = P2R(power_spec[i][j]*mask[i][j], phases[i][j]) y = librosa.core.istft(stft_array, hop_length, win_length,) librosa.output.write_wav(path, y, sampling_rate) def plot_spec(power): plt.subplots(figsize=(power.shape[1]/120,power.shape[0]/120), dpi = 150) librosa.display.specshow(data = librosa.core.amplitude_to_db(power, ref=np.max), y_axis='linear', #log or linear x_axis='time', sr = SR, hop_length = HOP) ORIG_SR = 44100 SR = 22050 # Частота дискретизации, с которой работает сеть WIN = 1024 # Размер окна дискретного преобразования Фурье HOP = 256 # Размер шага дискретного преобразования Фурье sample_len = 27 TRGT = 'vocals' ##Загрузка тестового набора #mus = musdb.DB(root='./musdb18', subsets='test', split=None) #track = mus[1] #print(track.name) #ORIG_SR = track.rate #track.chunk_duration = 15 #track.chunk_start = 0 #mix = librosa.core.to_mono(track.audio.T) #target = librosa.core.to_mono(track.targets[TRGT].audio.T) # #mix_power_spec, mix_phases = get_power_and_angle(mix, ORIG_SR, SR, WIN, HOP) #save_audio(mix_power_spec, mix_phases, HOP, WIN, SR, 'D:/mix.wav') # #target_power_spec, target_phases = get_power_and_angle(target, ORIG_SR, SR, WIN, HOP) #save_audio(target_power_spec, target_phases, HOP, WIN, SR, 'D:/orig_source.wav') # #other_power_spec = np.zeros(mix_power_spec.shape) #for i in range(mix_power_spec.shape[0]): # for j in range(mix_power_spec.shape[1]): # if mix_power_spec[i][j] >= target_power_spec[i][j]: # other_power_spec[i][j] = mix_power_spec[i][j] - target_power_spec[i][j] # else: # other_power_spec[i][j] = 0 #save_audio(other_power_spec, mix_phases, HOP, WIN, SR, 'D:/orig_other.wav') name = 'Кино' mix, ORIG_SR = librosa.core.load(path = 'D:\\tracks\\' + name + '.mp3', sr = 44100, mono = True, offset=21, duration=34-21) mix_power_spec, mix_phases = get_power_and_angle(mix, ORIG_SR, SR, WIN, HOP) save_audio(mix_power_spec, mix_phases, HOP, WIN, SR, 'D:\\tracks\\result\\' + name + ' mix.wav') plot_spec(mix_power_spec) # Нормализация и расширение спектрограммы mix_power_spec_norm = normalize_pow_spec(mix_power_spec) pad = np.zeros((mix_power_spec_norm.shape[0], sample_len//2)) mix_power_spec_norm = np.concatenate((pad, mix_power_spec_norm), axis=1) mix_power_spec_norm = np.concatenate((mix_power_spec_norm, pad), axis=1) # Загрузка модели t0 = time() print("<----------[INFO] model loading...") model = load_model('output/model final 10.hdf5') t1 = time() print("<----------[INFO] model was loaded in " + str(round((t1-t0), 1)) + " seconds...") t0 = time() print("<----------[INFO] mask forming...") mask = np.zeros((mix_power_spec_norm.shape[0], 0)) for i in range(mix_power_spec.shape[1]): mix = mix_power_spec_norm[:,i:i+sample_len] mix = np.expand_dims(mix, 0) mix = np.expand_dims(mix, len(mix.shape)) pred = model.predict(mix)[0] mask = np.concatenate((mask, np.expand_dims(pred, len(pred.shape))), axis=1) # Обработка маски for i in range(mask.shape[0]): for j in range(mask.shape[1]): if mask[i][j] >= 0.5: mask[i][j] = 1 else: mask[i][j] = 0 # if mask[i][j] <= 0.5: # mask[i][j] = 0 plot_spec(mask*1000) t1 = time() print("<----------[INFO] mask was formed in " + str(round((t1-t0), 1)) + " seconds...") save_audio(mix_power_spec, mix_phases, HOP, WIN, SR, 'D:\\tracks\\result\\' + name + ' source.wav', mask) plot_spec(mix_power_spec*mask) reverse_mask = np.zeros(mask.shape) for i in range(mask.shape[0]): for j in range(mask.shape[1]): reverse_mask[i,j] = 1 - mask[i,j] save_audio(mix_power_spec, mix_phases, HOP, WIN, SR, 'D:\\tracks\\result\\' + name + ' other.wav', reverse_mask) сб, 27 мая 2023 г. в 11:02, TomasSand ***@***.***>:

Dear Artyom Ivasik, your project is very interesting. I try to run your code and it's ok. But i don't find the predict file that you mentions in your Thesis. It's possible view this file? Thanks for your support. — Reply to this email directly, view it on GitHub <#2>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AQGSOQJFRAJIYHMCOXAVBFDXIGYJHANCNFSM6AAAAAAYRA47R4> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

Info predict file #2

Description

Activity

Metadata

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

Issue actions