def __init__(self, nfft, win_step):
def tfc(x):
return np.dstack([spectrogram(x[:, ci], nfft, win_step) for ci in range(x.shape[1])])
BaseNode.__init__(self)
self.nfft, self.win_step = nfft, win_step
self.n = FeatMap(tfc)
class TFC(BaseNode):
def __init__(self, nfft, win_step):
def tfc(x):
return np.dstack([spectrogram(x[:, ci], nfft, win_step) for ci in range(x.shape[1])])
BaseNode.__init__(self)
self.nfft, self.win_step = nfft, win_step
self.n = FeatMap(tfc)
def apply_(self, d):
assert len(d.feat_shape) == 2 # [frames x channels]
if d.feat_dim_lab != None:
assert d.feat_dim_lab[0] == "time"
tfc = self.n.apply(d)
feat_dim_lab = ["time", "frequency", d.feat_dim_lab[1]]
if d.feat_nd_lab != None:
old_time = np.asarray([float(i) for i in d.feat_nd_lab[0]])
time = np.mean(sliding_window(old_time, self.nfft, self.win_step), axis=1)
time = ["%.1f" % i for i in time]
dt = np.mean(np.diff(old_time))
dt = (np.max(old_time) - np.min(old_time)) / old_time.size
freqs = np.fft.fftfreq(self.nfft, dt)
freqs = ["%d" % abs(i) for i in freqs[: self.nfft / 2 + 1]]
channels = d.feat_nd_lab[1]
feat_nd_lab = [time, freqs, channels]
else:
feat_nd_lab = None
return DataSet(feat_dim_lab=feat_dim_lab, feat_nd_lab=feat_nd_lab, default=tfc)