fixed #657 by reimplementing the delta function. Deprecates trim= par…

…ameter.

librosa/feature/utils.py

@@ -7,12 +7,12 @@
 
 from .. import cache
 from ..util.exceptions import ParameterError
-
+from ..util.deprecation import Deprecated
 __all__ = ['delta', 'stack_memory']
 
 
 @cache(level=40)
-def delta(data, width=9, order=1, axis=-1, trim=True):
+def delta(data, width=9, order=1, axis=-1, trim=Deprecated()):
     r'''Compute delta features: local estimate of the derivative
     of the input data along the selected axis.
 
@@ -33,13 +33,14 @@ def delta(data, width=9, order=1, axis=-1, trim=True):
         the axis along which to compute deltas.
         Default is -1 (columns).
 
-    trim      : bool
-        set to `True` to trim the output matrix to the original size.
+    trim      : bool [DEPRECATED]
+        This parameter is deprecated in 0.6.0 and will be removed
+        in 0.7.0.
 
     Returns
     -------
-    delta_data   : np.ndarray [shape=(d, t) or (d, t + window)]
-        delta matrix of `data`.
+    delta_data   : np.ndarray [shape=(d, t)]
+        delta matrix of `data` at specified order
 
     Notes
     -----
@@ -97,20 +98,12 @@ def delta(data, width=9, order=1, axis=-1, trim=True):
     # Normalize the window so we're scale-invariant
     window /= np.sum(np.abs(window)**2)
 
-    # Pad out the data by repeating the border values (delta=0)
-    padding = [(0, 0)] * data.ndim
-    width = int(width)
-    padding[axis] = (width, width)
-    delta_x = np.pad(data, padding, mode='edge')
-
+    delta_x = data
     for _ in range(order):
-        delta_x = scipy.signal.lfilter(window, 1, delta_x, axis=axis)
-
-    # Cut back to the original shape of the input data
-    if trim:
-        idx = [slice(None)] * delta_x.ndim
-        idx[axis] = slice(- half_length - data.shape[axis], - half_length)
-        delta_x = delta_x[idx]
+        delta_x = scipy.ndimage.convolve1d(delta_x,
+                                           window,
+                                           axis=axis,
+                                           mode='nearest')
 
     return delta_x
 

tests/test_features.py

@@ -29,34 +29,22 @@ def test_delta():
     #       and width=3 filters
 
     def __test(width, order, axis, x):
-        # Compare trimmed and untrimmed versions
-        delta = librosa.feature.delta(x,
-                                      width=width,
-                                      order=order,
-                                      axis=axis,
-                                      trim=False)
-        delta_t = librosa.feature.delta(x,
+        delta   = librosa.feature.delta(x,
                                         width=width,
                                         order=order,
                                         axis=axis,
                                         trim=True)
 
         # Check that trimming matches the expected shape
-        eq_(x.shape, delta_t.shape)
-
-        # Check that trimming gives the right values in the right places
-        _s = [slice(None)] * delta.ndim
-        _s[axis] = slice(- width//2 - x.shape[axis], -(width//2)-1)
-        delta_retrim = delta[_s]
-        assert np.allclose(delta_t, delta_retrim)
+        eq_(x.shape, delta.shape)
 
         # Once we're sufficiently far into the signal (ie beyond half_len)
-        # (x + delta_t)[t] should approximate x[t+1] if x is actually linear
+        # (x + delta)[t] should approximate x[t+1] if x is actually linear
         slice_orig = [slice(None)] * x.ndim
         slice_out = [slice(None)] * delta.ndim
         slice_orig[axis] = slice(width//2 + 1, -width//2 + 1)
         slice_out[axis] = slice(width//2, -width//2)
-        assert np.allclose((x + delta_t)[slice_out], x[slice_orig])
+        assert np.allclose((x + delta)[slice_out], x[slice_orig])
 
     x = np.vstack([np.arange(100.0)] * 3)