import argparse import os import cv2 import librosa import numpy as np import soundfile as sf import torch import time from tqdm import tqdm from lib import dataset from lib import spec_utils from lib.model_param_init import ModelParameters class VocalRemover(object): def __init__(self, model, device, window_size): self.model = model self.offset = model.offset self.device = device self.window_size = window_size def _execute(self, X_mag_pad, roi_size, n_window, aggressiveness): self.model.eval() with torch.no_grad(): preds = [] for i in tqdm(range(n_window)): start = i * roi_size X_mag_window = X_mag_pad[None, :, :, start:start + self.window_size] X_mag_window = torch.from_numpy(X_mag_window).to(self.device) pred = self.model.predict(X_mag_window, aggressiveness) pred = pred.detach().cpu().numpy() preds.append(pred[0]) pred = np.concatenate(preds, axis=2) return pred def preprocess(self, X_spec): X_mag = np.abs(X_spec) X_phase = np.angle(X_spec) return X_mag, X_phase def inference(self, X_spec, aggressiveness): X_mag, X_phase = self.preprocess(X_spec) coef = X_mag.max() X_mag_pre = X_mag / coef n_frame = X_mag_pre.shape[2] pad_l, pad_r, roi_size = dataset.make_padding(n_frame, self.window_size, self.offset) n_window = int(np.ceil(n_frame / roi_size)) X_mag_pad = np.pad(X_mag_pre, ((0, 0), (0, 0), (pad_l, pad_r)), mode='constant') pred = self._execute(X_mag_pad, roi_size, n_window, aggressiveness) pred = pred[:, :, :n_frame] return pred * coef, X_mag, np.exp(1.j * X_phase) def inference_tta(self, X_spec, aggressiveness): X_mag, X_phase = self.preprocess(X_spec) coef = X_mag.max() X_mag_pre = X_mag / coef n_frame = X_mag_pre.shape[2] pad_l, pad_r, roi_size = dataset.make_padding(n_frame, self.window_size, self.offset) n_window = int(np.ceil(n_frame / roi_size)) X_mag_pad = np.pad(X_mag_pre, ((0, 0), (0, 0), (pad_l, pad_r)), mode='constant') pred = self._execute(X_mag_pad, roi_size, n_window, aggressiveness) pred = pred[:, :, :n_frame] pad_l += roi_size // 2 pad_r += roi_size // 2 n_window += 1 X_mag_pad = np.pad(X_mag_pre, ((0, 0), (0, 0), (pad_l, pad_r)), mode='constant') pred_tta = self._execute(X_mag_pad, roi_size, n_window, aggressiveness) pred_tta = pred_tta[:, :, roi_size // 2:] pred_tta = pred_tta[:, :, :n_frame] return (pred + pred_tta) * 0.5 * coef, X_mag, np.exp(1.j * X_phase) def main(): p = argparse.ArgumentParser() p.add_argument('--gpu', '-g', type=int, default=-1) p.add_argument('--pretrained_model', '-P', type=str, default='models/baseline.pth') p.add_argument('--input', '-i', required=True) p.add_argument('--nn_architecture', '-n', type=str, default='default') p.add_argument('--model_params', '-m', type=str, default='') p.add_argument('--window_size', '-w', type=int, default=512) p.add_argument('--output_image', '-I', action='store_true') p.add_argument('--postprocess', '-p', action='store_true') p.add_argument('--tta', '-t', action='store_true') p.add_argument('--high_end_process', '-H', type=str, choices=['none', 'bypass', 'correlation'], default='none') p.add_argument('--aggressiveness', '-A', type=float, default=0.09) args = p.parse_args() if args.nn_architecture == 'default': from lib import nets if args.nn_architecture == '33966KB': from lib import nets_33966KB as nets if args.nn_architecture == '123821KB': from lib import nets_123821KB as nets if args.nn_architecture == '129605KB': from lib import nets_129605KB as nets #if '' == args.model_params: # mp = ModelParameters(args.pretrained_model) #else: mp = ModelParameters(args.model_params) start_time = time.time() print('loading model...', end=' ') device = torch.device('cpu') model = nets.CascadedASPPNet(mp.param['bins'] * 2) model.load_state_dict(torch.load(args.pretrained_model, map_location=device)) if torch.cuda.is_available() and args.gpu >= 0: device = torch.device('cuda:{}'.format(args.gpu)) model.to(device) print('done') print('loading & stft of wave source...', end=' ') X_wave, y_wave, X_spec_s, y_spec_s = {}, {}, {}, {} basename = os.path.splitext(os.path.basename(args.input))[0] bands_n = len(mp.param['band']) for d in range(bands_n, 0, -1): bp = mp.param['band'][d] if d == bands_n: # high-end band X_wave[d], _ = librosa.load( args.input, bp['sr'], False, dtype=np.float32, res_type=bp['res_type']) if X_wave[d].ndim == 1: X_wave[d] = np.asarray([X_wave[d], X_wave[d]]) else: # lower bands X_wave[d] = librosa.resample(X_wave[d+1], mp.param['band'][d+1]['sr'], bp['sr'], res_type=bp['res_type']) X_spec_s[d] = spec_utils.wave_to_spectrogram_mt(X_wave[d], bp['hl'], bp['n_fft'], mp.param['mid_side'], mp.param['reverse']) if d == bands_n and args.high_end_process in ['bypass', 'correlation']: input_high_end_h = (bp['n_fft']//2 - bp['crop_stop']) + (mp.param['pre_filter_stop'] - mp.param['pre_filter_start']) input_high_end = X_spec_s[d][:, bp['n_fft']//2-input_high_end_h:bp['n_fft']//2, :] X_spec_m = spec_utils.combine_spectrograms(X_spec_s, mp) del X_wave, X_spec_s print('done') vr = VocalRemover(model, device, args.window_size) if args.tta: pred, X_mag, X_phase = vr.inference_tta(X_spec_m, {'value': args.aggressiveness, 'split_bin': mp.param['band'][1]['crop_stop']}) else: pred, X_mag, X_phase = vr.inference(X_spec_m, {'value': args.aggressiveness, 'split_bin': mp.param['band'][1]['crop_stop']}) if args.postprocess: print('post processing...', end=' ') pred_inv = np.clip(X_mag - pred, 0, np.inf) pred = spec_utils.mask_silence(pred, pred_inv) print('done') if 'is_vocal_model' in mp.param: # swap stems = {'inst': 'Vocals', 'vocals': 'Instruments'} else: stems = {'inst': 'Instruments', 'vocals': 'Vocals'} print('inverse stft of {}...'.format(stems['inst']), end=' ') y_spec_m = pred * X_phase v_spec_m = X_spec_m - y_spec_m if args.high_end_process == 'bypass': wave = spec_utils.cmb_spectrogram_to_wave(y_spec_m, mp, input_high_end_h, input_high_end) elif args.high_end_process == 'correlation': for i in range(input_high_end.shape[2]): for c in range(2): X_mag_max = np.amax(input_high_end[c, :, i]) b1 = mp.param['pre_filter_start']-input_high_end_h//2 b2 = mp.param['pre_filter_start']-1 if X_mag_max > 0 and np.sum(np.abs(v_spec_m[c, b1:b2, i])) / (b2 - b1) > 0.07: y_mag = np.median(y_spec_m[c, b1:b2, i]) input_high_end[c, :, i] = np.true_divide(input_high_end[c, :, i], abs(X_mag_max) / min(abs(y_mag * 4), abs(X_mag_max))) wave = spec_utils.cmb_spectrogram_to_wave(y_spec_m, mp, input_high_end_h, input_high_end) else: wave = spec_utils.cmb_spectrogram_to_wave(y_spec_m, mp) print('done') model_name = os.path.splitext(os.path.basename(args.pretrained_model))[0] sf.write(os.path.join('separated', '{}_{}_{}.wav'.format(basename, model_name, stems['inst'])), wave, mp.param['sr']) if True: print('inverse stft of {}...'.format(stems['vocals']), end=' ') #v_spec_m = X_spec_m - y_spec_m wave = spec_utils.cmb_spectrogram_to_wave(v_spec_m, mp) print('done') sf.write(os.path.join('separated', '{}_{}_{}.wav'.format(basename, model_name, stems['vocals'])), wave, mp.param['sr']) if args.output_image: with open('{}_{}.jpg'.format(basename, stems['inst']), mode='wb') as f: image = spec_utils.spectrogram_to_image(y_spec_m) _, bin_image = cv2.imencode('.jpg', image) bin_image.tofile(f) with open('{}_{}.jpg'.format(basename, stems['vocals']), mode='wb') as f: image = spec_utils.spectrogram_to_image(v_spec_m) _, bin_image = cv2.imencode('.jpg', image) bin_image.tofile(f) print('Total time: {0:.{1}f}s'.format(time.time() - start_time, 1)) if __name__ == '__main__': main()