Disable resampling for rates that divide gset sr without remainder

Sampling rates 11025 Hz and 22050 Hz are safe and do not need resampling.
This simple fix gets rid of many ugly data points that were not present
in the original data annotations.

related to: #7

penn/data/preprocess/core.py

@@ -272,46 +272,50 @@ def gset():
             unvoiced = pitch == 0
             voiced = ~unvoiced
 
-        pitch_list = np.vsplit(pitch, pitch.shape[0])
-        pitch_list_final = []
+        # FOR sampling rates like 11025, 22050, 44100, resampling isn't necessary
+        if GSET_SAMPLE_RATE / penn.SAMPLE_RATE % 1 != 0:
+            printf("Resampling to penn.SAMPLE_RATE")
 
-        voiced_list = np.vsplit(voiced, voiced.shape[0])
-        voiced_list_final = []
+            pitch_list = np.vsplit(pitch, pitch.shape[0])
+            pitch_list_final = []
 
-        for pitch_arr, voiced_arr in zip(pitch_list, voiced_list):
-            # Get target number of frames
-            frames = penn.convert.samples_to_frames(audio.shape[-1])
+            voiced_list = np.vsplit(voiced, voiced.shape[0])
+            voiced_list_final = []
 
-            pitch_arr = pitch_arr[0, :]
-            voiced_arr = voiced_arr[0, :]
+            for pitch_arr, voiced_arr in zip(pitch_list, voiced_list):
+                # Get target number of frames
+                frames = penn.convert.samples_to_frames(audio.shape[-1])
 
-            # Linearly interpolate to target number of frames
-            new_times = penn.HOPSIZE_SECONDS * np.arange(0, frames)
-            new_times += penn.HOPSIZE_SECONDS / 2.
-            pitch_arr = 2. ** np.interp(new_times, times, np.log2(pitch_arr))
+                pitch_arr = pitch_arr[0, :]
+                voiced_arr = voiced_arr[0, :]
 
-            # Linearly interpolate voiced_arr/unvoiced_arr tokens
-            voiced_arr = np.interp(new_times, times, voiced_arr) > .5
+                # Linearly interpolate to target number of frames
+                new_times = penn.HOPSIZE_SECONDS * np.arange(0, frames)
+                new_times += penn.HOPSIZE_SECONDS / 2.
+                pitch_arr = 2. ** np.interp(new_times, times, np.log2(pitch_arr))
 
-            # Check shapes
-            assert (
-                penn.convert.samples_to_frames(audio.shape[-1]) ==
-                pitch_arr.shape[-1] ==
-                voiced_arr.shape[-1])
+                # Linearly interpolate voiced_arr/unvoiced_arr tokens
+                voiced_arr = np.interp(new_times, times, voiced_arr) > .5
 
-            assert np.logical_not(pitch_arr[voiced_arr] == 0).all()
+                # Check shapes
+                assert (
+                    penn.convert.samples_to_frames(audio.shape[-1]) ==
+                    pitch_arr.shape[-1] ==
+                    voiced_arr.shape[-1])
 
-            pitch_list_final.append(pitch_arr)
-            voiced_list_final.append(voiced_arr)
+                assert np.logical_not(pitch_arr[voiced_arr] == 0).all()
 
-        pitch = np.vstack(pitch_list_final)
-        voiced = np.vstack(voiced_list_final)
+                pitch_list_final.append(pitch_arr)
+                voiced_list_final.append(voiced_arr)
 
-        if pitch.shape[0] == 1:
-            pitch = pitch[0, :]
+            pitch = np.vstack(pitch_list_final)
+            voiced = np.vstack(voiced_list_final)
 
-        if voiced.shape[0] == 1:
-            voiced = voiced[0, :]
+            if pitch.shape[0] == 1:
+                pitch = pitch[0, :]
+
+            if voiced.shape[0] == 1:
+                voiced = voiced[0, :]
 
         # Save to cache
         np.save(output_directory / f'{stem}-pitch.npy', pitch)
@@ -559,13 +563,11 @@ def extract_pitch_array_jams(jam: jams.JAMS, track, uniform=True) -> Tuple[np.nd
             freq = np.array([pitch.value['frequency']])
 
             # Don't keep track of zero or unvoiced frequencies
-            if np.sum(freq) == 0 or not pitch.value['voiced']:
-                freq = np.zeros(1)
-
-            # Append the observation time
-            entry_times = np.append(entry_times, pitch.time)
-            # Append the frequency
-            slice_pitch_list.append(freq)
+            if np.sum(freq) != 0 and pitch.value['voiced']:
+                # Append the observation time
+                entry_times = np.append(entry_times, pitch.time)
+                # Append the frequency
+                slice_pitch_list.append(freq)
 
         # Sort the pitch list before resampling just in case it is not already sorted
         entry_times, slice_pitch_array = sort_pitch_list(entry_times, slice_pitch_list)