def runit(self, siglen, fmin, fmax, obins, sllen, trlen, real):
sig = rndsig[:siglen]
scale = OctScale(fmin, fmax, obins)
nsgt = NSGT_sliced(scale, fs=44100, sl_len=sllen, tr_area=trlen, real=real)
c = nsgt.forward((sig,))
rc = nsgt.backward(c)
s_r = np.concatenate(list(map(list,rc)))[:len(sig)]
close = np.allclose(sig, s_r, atol=1.e-3)
if not close:
print("Failing params:", siglen, fmin, fmax, obins, sllen, trlen, real)
dev = np.abs(s_r-sig)
print("Error", np.where(dev>1.e-3), np.max(dev))
self.assertTrue(close)
def runit(self, siglen, fmin, fmax, obins, sllen, trlen, real):
sig = rndsig[:siglen]
scale = OctScale(fmin, fmax, obins)
nsgt = NSGT_sliced(scale, fs=44100, sl_len=sllen, tr_area=trlen, real=real)
c = nsgt.forward((sig,))
rc = nsgt.backward(c)
s_r = np.concatenate(map(list,rc))[:len(sig)]
close = np.allclose(sig, s_r, atol=1.e-3)
if not close:
print "Failing params:", siglen, fmin, fmax, obins, sllen, trlen, real
dev = np.abs(s_r-sig)
print "Error", np.where(dev>1.e-3), np.max(dev)
self.assertTrue(close)
def main():
parser = ArgumentParser()
parser.add_argument(
"--mask",
type=str,
default="soft",
choices=("hard", "soft"),
help="mask strategy",
)
parser.add_argument("--outdir", type=str, default="./", help="output directory")
parser.add_argument(
"--stream-size",
type=int,
default=1024,
help="stream size for simulated realtime from wav (default=%(default)s)",
)
parser.add_argument("input", type=str, help="input file")
args = parser.parse_args()
prefix = args.input.split("/")[-1].split("_")[0]
harm_out = os.path.join(args.outdir, prefix + "_harmonic.wav")
perc_out = os.path.join(args.outdir, prefix + "_percussive.wav")
print("writing files to {0}, {1}".format(harm_out, perc_out))
lharm = 17
lperc = 7
# calculate transform parameters
nsgt_scale = OctScale(80, 20000, 12)
trlen = args.stream_size # transition length
sllen = 4 * args.stream_size # slice length
x, fs = librosa.load(args.input, sr=None)
xh = numpy.zeros_like(x)
xp = numpy.zeros_like(x)
hop = trlen
chunk_size = hop
n_chunks = int(numpy.floor(x.shape[0] // hop))
eps = numpy.finfo(numpy.float32).eps
slicq = NSGT_sliced(
nsgt_scale,
sllen,
trlen,
fs,
real=True,
matrixform=True,
)
total_time = 0.0
for chunk in range(n_chunks - 1):
t1 = cputime()
start = chunk * hop
end = start + sllen
s = x[start:end]
signal = (s,)
c = slicq.forward(signal)
c = list(c)
C = numpy.asarray(c)
Cmag = numpy.abs(C)
H = scipy.ndimage.median_filter(Cmag, size=(1, lharm, 1))
P = scipy.ndimage.median_filter(Cmag, size=(1, 1, lperc))
if args.mask == "soft":
# soft mask first
tot = numpy.power(H, 2.0) + numpy.power(P, 2.0) + eps
Mp = numpy.divide(numpy.power(H, 2.0), tot)
Mh = numpy.divide(numpy.power(P, 2.0), tot)
else:
Mh = numpy.divide(H, P + eps) > 2.0
Mp = numpy.divide(P, H + eps) >= 2.0
Cp = numpy.multiply(Mp, C)
Ch = numpy.multiply(Mh, C)
# generator for backward transformation
outseq_h = slicq.backward(Ch)
outseq_p = slicq.backward(Cp)
# make single output array from iterator
sh_r = next(reblock(outseq_h, len(s), fulllast=False))
sh_r = sh_r.real
sp_r = next(reblock(outseq_p, len(s), fulllast=False))
sp_r = sp_r.real
xh[start:end] = sh_r
xp[start:end] = sp_r
t2 = cputime()
total_time += t2 - t1
print("Calculation time per iter: %fs" % (total_time / n_chunks))
scipy.io.wavfile.write(harm_out, fs, xh)
scipy.io.wavfile.write(perc_out, fs, xp)
return 0
matrixform=args.matrixform, reducedform=args.reducedform,
multithreading=args.multithreading
)
t1 = cputime()
signal = (s,)
# generator for forward transformation
c = slicq.forward(signal)
# realize transform from generator
c = list(c)
# generator for backward transformation
outseq = slicq.backward(c)
# make single output array from iterator
s_r = next(reblock(outseq, len(s), fulllast=False))
s_r = s_r.real
t2 = cputime()
norm = lambda x: np.sqrt(np.sum(np.abs(np.square(np.abs(x)))))
rec_err = norm(s-s_r)/norm(s)
print("Reconstruction error: %.3e"%rec_err)
print("Calculation time: %.3fs"%(t2-t1))
# Compare the sliced coefficients with non-sliced ones
if False:
# not implemented yet!
t1 = time()
signal = (s,)
# generator for forward transformation
c = slicq.forward(signal)
# realize transform from generator
c = list(c)
# cl = map(len,c[0])
# print "c",len(cl),cl
# generator for backward transformation
outseq = slicq.backward(c)
# make single output array from iterator
s_r = reblock(outseq,len(s),fulllast=False).next()
s_r = s_r.real
t2 = time()
norm = lambda x: N.sqrt(N.sum(N.abs(N.square(N.abs(x)))))
rec_err = norm(s-s_r)/norm(s)
print "Reconstruction error: %.3e"%rec_err
print "Calculation time: %.3f s"%(t2-t1)
# Compare the sliced coefficients with non-sliced ones
if False:
# not implemented yet!
freq_start = freq_idx
freq_end = C_block.shape[2]
freq_idx += freq_end
C_block_ola = torch.squeeze(overlap_add_slicq(
torch.unsqueeze(C_block, dim=0)),
dim=0)
C_block_flatten = torch.squeeze(overlap_add_slicq(torch.unsqueeze(
C_block, dim=0),
flatten=True),
dim=0)
print(
f'\tblock {i}, f {freq_start}-{freq_start+freq_end-1}: {C_block.shape}, {C_block_ola.shape}, {C_block_flatten.shape}'
)
signal_recon = slicq.backward(c, signal.shape[-1])
print(
f'recon error (mse): {torch.nn.functional.mse_loss(signal_recon, signal)}')
print(f'comparing 4096 stft for fun')
print(f'signal: {signal.shape}')
S = torch.stft(signal,
n_fft=4096,
hop_length=1024,
return_complex=True,
center=False).type(torch.complex64)
print(f'stft with 4096/1024: {S.shape}')
