[WIP] Decoding from the categorical posterior distribution

crepe/core.py

@@ -25,7 +25,7 @@
 def build_and_load_model(model_capacity):
     """
     Build the CNN model and load the weights
-    
+
     Parameters
     ----------
     model_capacity : 'tiny', 'small', 'medium', 'large', or 'full'
@@ -92,24 +92,29 @@ def output_path(file, suffix, output_dir):
     return path
 
 
-def to_local_average_cents(salience, center=None):
+def bin_to_cents(bin):
     """
-    find the weighted average cents near the argmax bin
+    map bin numbers to cents values
     """
-
-    if not hasattr(to_local_average_cents, 'cents_mapping'):
-        # the bin number-to-cents mapping
-        to_local_average_cents.mapping = (
+    if not hasattr(bin_to_cents, 'cents_mapping'):
+        bin_to_cents.mapping = (
                 np.linspace(0, 7180, 360) + 1997.3794084376191)
 
+    return bin_to_cents.mapping[bin]
+
+
+def to_local_average_cents(salience, center=None):
+    """
+    find the weighted average cents near the argmax bin
+    """
     if salience.ndim == 1:
         if center is None:
             center = int(np.argmax(salience))
         start = max(0, center - 4)
         end = min(len(salience), center + 5)
         salience = salience[start:end]
         product_sum = np.sum(
-            salience * to_local_average_cents.mapping[start:end])
+            salience * bin_to_cents(np.arange(start, end)))
         weight_sum = np.sum(salience)
         return product_sum / weight_sum
     if salience.ndim == 2:
@@ -124,43 +129,32 @@ def to_viterbi_cents(salience):
     Find the Viterbi path using a transition prior that induces pitch
     continuity.
     """
-    from hmmlearn import hmm
-
-    # uniform prior on the starting pitch
-    starting = np.ones(360) / 360
+    from librosa.sequence import viterbi
 
     # transition probabilities inducing continuous pitch
     xx, yy = np.meshgrid(range(360), range(360))
     transition = np.maximum(12 - abs(xx - yy), 0)
     transition = transition / np.sum(transition, axis=1)[:, None]
 
-    # emission probability = fixed probability for self, evenly distribute the
-    # others
-    self_emission = 0.1
-    emission = (np.eye(360) * self_emission + np.ones(shape=(360, 360)) *
-                ((1 - self_emission) / 360))
-
-    # fix the model parameters because we are not optimizing the model
-    model = hmm.MultinomialHMM(360, starting, transition)
-    model.startprob_, model.transmat_, model.emissionprob_ = \
-        starting, transition, emission
+    # compute the posterior distribution from the logits
+    posterior = np.exp(salience) / np.sum(np.exp(salience), axis=1,
+                                          keepdims=True)
 
-    # find the Viterbi path
-    observations = np.argmax(salience, axis=1)
-    path = model.predict(observations.reshape(-1, 1), [len(observations)])
+    # determine the path through the posterior distribution
+    path = viterbi(posterior.T, transition)
 
-    return np.array([to_local_average_cents(salience[i, :], path[i]) for i in
-                     range(len(observations))])
+    # convert bin indices to cents
+    return bin_to_cents(path)
 
 
 def get_activation(audio, sr, model_capacity='full', center=True, step_size=10,
                    verbose=1):
     """
-    
+
     Parameters
     ----------
     audio : np.ndarray [shape=(N,) or (N, C)]
-        The audio samples. Multichannel audio will be downmixed. 
+        The audio samples. Multichannel audio will be downmixed.
     sr : int
         Sample rate of the audio samples. The audio will be resampled if
         the sample rate is not 16 kHz, which is expected by the model.
@@ -216,11 +210,11 @@ def predict(audio, sr, model_capacity='full',
             viterbi=False, center=True, step_size=10, verbose=1):
     """
     Perform pitch estimation on given audio
-    
+
     Parameters
     ----------
     audio : np.ndarray [shape=(N,) or (N, C)]
-        The audio samples. Multichannel audio will be downmixed. 
+        The audio samples. Multichannel audio will be downmixed.
     sr : int
         Sample rate of the audio samples. The audio will be resampled if
         the sample rate is not 16 kHz, which is expected by the model.
@@ -242,7 +236,7 @@ def predict(audio, sr, model_capacity='full',
     Returns
     -------
     A 4-tuple consisting of:
-    
+
         time: np.ndarray [shape=(T,)]
             The timestamps on which the pitch was estimated
         frequency: np.ndarray [shape=(T,)]

setup.py

@@ -73,8 +73,8 @@
         'matplotlib>=2.1.0',
         'resampy>=0.2.0,<0.3.0',
         'h5py>=2.7.0,<3.0.0',
-        'hmmlearn>=0.2.0,<0.3.0',
         'imageio>=2.3.0',
+        'librosa>=0.6.2',
         'scikit-learn>=0.16'
     ],
     package_data={