in train.py, if config.inference['normalize'] is True, transform estimates after demix

[hunterhogan]

When the new parameter in config.inference, normalize, was added in this commit, logic was added to valid.py and inference.py to remove the normalization transformation from estimates. In train.py, the logic to transform mix was added, however, nothing was added for estimates.

Music-Source-Separation-Training/train.py

Lines 182 to 212 in 54bd725

	mix_orig = mix.copy()
	mix = mix.T # (channels, waveform)
	if 'normalize' in config.inference:
	if config.inference['normalize'] is True:
	mono = mix.mean(0)
	mean = mono.mean()
	std = mono.std()
	mix = (mix - mean) / std
	folder = os.path.dirname(path)
	folder_name = os.path.abspath(folder)
	if verbose:
	print('Song: {}'.format(folder_name))
	res = demix(config, model, mix, device, model_type=args.model_type)
	if 1:
	pbar_dict = {}
	for instr in instruments:
	if instr != 'other' or config.training.other_fix is False:
	try:
	track, sr1 = sf.read(folder + '/{}.wav'.format(instr))
	except Exception as e:
	# print('No data for stem: {}. Skip!'.format(instr))
	continue
	else:
	# other is actually instrumental
	track, sr1 = sf.read(folder + '/{}.wav'.format('vocals'))
	track = mix_orig - track
	references = np.expand_dims(track, axis=0)
	estimates = np.expand_dims(res[instr].T, axis=0)
	sdr_val = sdr(references, estimates)[0]

Would this affect the SDR calculation in sdr_val = sdr(references, estimates)[0]?

[ZFTurbo]

I added normalize only to support one model (SCNet) for inference. I didn't change training to support normalization.

[Kenus23]

how use training for stem separation?