Merge dbbb932 into 09381cc

.travis.yml

@@ -1,6 +1,8 @@
 cache:
+  apt: True
   directories:
   - $HOME/env
+  - $HOME/Library/Caches/Homebrew
 
 language: python
 
@@ -22,13 +24,25 @@ jobs:
           language: generic
           osx_image: xcode11
 
+addons:
+    apt:
+        packages:
+        - libsamplerate0
+
+    homebrew:
+        packages:
+        - libsamplerate
+
 before_install:
     - bash .travis_dependencies.sh
     - export PATH="$HOME/env/miniconda-$TRAVIS_OS_NAME$TRAVIS_PYTHON_VERSION/bin:$PATH";
     - hash -r
     - source activate test-environment
     - conda list
 
+before_cache:
+    - brew cleanup
+
 install:
     # install your own package into the environment
     # pip install -e rather than setup.py, so that coverage can find the source

.travis_dependencies.sh

@@ -25,14 +25,9 @@ if [ ! -d "$src" ]; then
         # Download miniconda packages
         if [ "$TRAVIS_OS_NAME" = "osx" ]; then
             wget https://repo.continuum.io/miniconda/Miniconda3-latest-MacOSX-x86_64.sh -O miniconda.sh;
-            # Platform-specific dependencies
-            brew update
-            brew install libsamplerate
         fi
         if [ "$TRAVIS_OS_NAME" = "linux" ]; then
             wget https://repo.continuum.io/miniconda/Miniconda3-latest-Linux-x86_64.sh -O miniconda.sh;
-            sudo apt update
-            sudo apt install libsamplerate0
         fi
         # Install both environments
         bash miniconda.sh -b -p $src

librosa/feature/utils.py

@@ -4,6 +4,7 @@
 
 import numpy as np
 import scipy.signal
+from numba import jit
 
 from .._cache import cache
 from ..util.exceptions import ParameterError
@@ -135,7 +136,7 @@ def stack_memory(data, n_steps=2, delay=1, **kwargs):
 
     Parameters
     ----------
-    data : np.ndarray [shape=(t,) or (d, t)]
+    data : np.ndarray [shape=(d, t)]
         Input data matrix.  If `data` is a vector (`data.ndim == 1`),
         it will be interpreted as a row matrix and reshaped to `(1, t)`.
 
@@ -219,12 +220,19 @@ def stack_memory(data, n_steps=2, delay=1, **kwargs):
     if n_steps < 1:
         raise ParameterError('n_steps must be a positive integer')
 
+    if data.ndim > 2:
+        raise ParameterError('Input must be at most 2-dimensional. '
+                             'Given data.shape={}'.format(data.shape))
+
     if delay == 0:
         raise ParameterError('delay must be a non-zero integer')
 
     data = np.atleast_2d(data)
-
-    t = data.shape[1]
+    t = data.shape[-1]
+
+    if t < 1:
+        raise ParameterError('Cannot stack memory when input data has '
+                             'no columns. Given data.shape={}'.format(data.shape))
     kwargs.setdefault('mode', 'constant')
 
     if kwargs['mode'] == 'constant':
@@ -238,13 +246,56 @@ def stack_memory(data, n_steps=2, delay=1, **kwargs):
 
     data = np.pad(data, [(0, 0), padding], **kwargs)
 
-    history = np.vstack([np.roll(data, -i * delay, axis=1) for i in range(n_steps)[::-1]])
+    # Construct the shape of the target array
+    shape = list(data.shape)
+    shape[0] = shape[0] * n_steps
+    shape[1] = t
+    shape = tuple(shape)
+
+    # Construct the output array to match layout and dtype of input
+    history = np.empty_like(data, shape=shape)
+
+    # Populate the output array
+    __stack(history, data, n_steps, delay)
+
+    return history
+
+
+@jit(nopython=True, cache=True)
+def __stack(history, data, n_steps, delay):
+    '''Memory-stacking helper function.
+
+    Parameters
+    ----------
+    history : output array (2-dimensional)
+
+    data : pre-padded input array (2-dimensional)
+
+    n_steps : int > 0, the number of steps to stack
+
+    delay : int != 0, the amount of delay between steps
+
+    Returns
+    -------
+    None
+        Output is stored directly in the history array
+    '''
+    # Dimension of each copy of the data
+    d = data.shape[0]
+
+    # Total number of time-steps to output
+    t = history.shape[1]
 
-    # Trim to original width
     if delay > 0:
-        history = history[:, :t]
+        for step in range(n_steps):
+            q = n_steps - 1 - step
+            # nth block is original shifted left by n*delay steps
+            history[step * d:(step + 1) * d] = data[:, q*delay:q*delay+t]
     else:
-        history = history[:, -t:]
+        # Handle the last block separately to avoid -t:0 empty slices
+        history[-d:, :] = data[:, -t:]
 
-    # Make contiguous
-    return np.asfortranarray(history)
+        for step in range(n_steps-1):
+            # nth block is original shifted right by n*delay steps
+            q = n_steps - 1 - step
+            history[step * d:(step + 1) * d] = data[:, -t + q*delay:q*delay]

tests/test_features.py

@@ -69,7 +69,7 @@ def test_delta_badwidthaxis(x, width, axis):
     librosa.feature.delta(x, width=width, axis=axis)
 
 
-@pytest.mark.parametrize("data", [np.random.randn(5), np.random.randn(5, 5), np.remainder(np.arange(10000), 24)])
+@pytest.mark.parametrize("data", [np.arange(5.), np.remainder(np.arange(10000), 24)])
 @pytest.mark.parametrize("delay", [-4, -2, -1, 1, 2, 4])
 @pytest.mark.parametrize("n_steps", [1, 2, 3, 300])
 def test_stack_memory(data, n_steps, delay):
@@ -104,6 +104,18 @@ def test_stack_memory_fail(data, n_steps, delay):
     librosa.feature.stack_memory(data, n_steps=n_steps, delay=delay)
 
 
+@pytest.mark.xfail(raises=librosa.ParameterError)
+def test_stack_memory_ndim_toobig():
+    librosa.feature.stack_memory(np.zeros((2,2,2)), n_steps=3, delay=1)
+
+@pytest.mark.parametrize('data', [np.zeros((2, 0))])
+@pytest.mark.xfail(raises=librosa.ParameterError)
+@pytest.mark.parametrize('delay', [-2, -1, 1, 2])
+@pytest.mark.parametrize('n_steps', [1, 2])
+def test_stack_memory_ndim_badshape(data, delay, n_steps):
+    librosa.feature.stack_memory(data, n_steps=n_steps, delay=delay)
+
+
 @pytest.fixture(scope="module")
 def S_ideal():
     # An idealized spectrum with all zero energy except at one DFT band