Merge pull request #91 from bmcfee/confidence-rated-prediction

Confidence rated prediction

pumpp/task/beat.py

@@ -4,6 +4,7 @@
 
 import numpy as np
 
+from librosa import time_to_frames
 import jams
 from mir_eval.util import boundaries_to_intervals, adjust_intervals
 from sklearn.preprocessing import LabelBinarizer, LabelEncoder
@@ -95,7 +96,11 @@ def inverse(self, encoded, downbeat=None, duration=None):
 
         ann = jams.Annotation(namespace=self.namespace, duration=duration)
 
-        beat_times = [t for t, _ in self.decode_events(encoded) if _]
+        beat_times = np.asarray([t for t, _ in self.decode_events(encoded) if _])
+        beat_frames = time_to_frames(beat_times,
+                                     sr=self.sr,
+                                     hop_length=self.hop_length)
+
         if downbeat is not None:
             downbeat_times = set([t for t, _ in self.decode_events(downbeat)
                                   if _])
@@ -106,12 +111,16 @@ def inverse(self, encoded, downbeat=None, duration=None):
             pickup_beats = 0
 
         value = - pickup_beats - 1
-        for beat in beat_times:
-            if beat in downbeat_times:
+        for beat_t, beat_f in zip(beat_times, beat_frames):
+            if beat_t in downbeat_times:
                 value = 1
             else:
                 value += 1
-            ann.append(time=beat, duration=0, value=value)
+            confidence = encoded[beat_f]
+            ann.append(time=beat_t,
+                       duration=0,
+                       value=value,
+                       confidence=confidence)
 
         return ann
 

pumpp/task/chord.py

@@ -9,6 +9,7 @@
 from sklearn.preprocessing import LabelBinarizer, LabelEncoder
 from sklearn.preprocessing import MultiLabelBinarizer
 
+from librosa import time_to_frames
 import mir_eval
 import jams
 
@@ -492,12 +493,30 @@ def inverse(self, encoded, duration=None):
                                                        duration=duration,
                                                        multi=False,
                                                        sparse=self.sparse):
+
+            # Map start:end to frames
+            f_start, f_end = time_to_frames([start, end],
+                                            sr=self.sr,
+                                            hop_length=self.hop_length)
+
+            # Reverse the index
             if self.sparse:
+                # Compute the confidence
+                if encoded.shape[1] == 1:
+                    # This case is for full-confidence prediction (just the index)
+                    confidence = 1.
+                else:
+                    confidence = np.mean(encoded[f_start:f_end+1, value])
+
                 value_dec = self.encoder.inverse_transform(value)
             else:
+                confidence = np.mean(encoded[f_start:f_end+1, np.argmax(value)])
                 value_dec = self.encoder.inverse_transform(np.atleast_2d(value))
 
             for vd in value_dec:
-                ann.append(time=start, duration=end-start, value=vd)
+                ann.append(time=start,
+                           duration=end-start,
+                           value=vd,
+                           confidence=float(confidence))
 
         return ann

pumpp/task/tags.py

@@ -5,6 +5,8 @@
 import numpy as np
 from sklearn.preprocessing import MultiLabelBinarizer
 
+from librosa import time_to_frames
+
 import jams
 
 from .base import BaseTaskTransformer
@@ -106,10 +108,20 @@ def inverse(self, encoded, duration=None):
         ann = jams.Annotation(namespace=self.namespace, duration=duration)
         for start, end, value in self.decode_intervals(encoded,
                                                        duration=duration):
+            # Map start:end to frames
+            f_start, f_end = time_to_frames([start, end],
+                                            sr=self.sr,
+                                            hop_length=self.hop_length)
+
+            confidence = np.mean(encoded[f_start:f_end+1, value])
+
             value_dec = self.encoder.inverse_transform(np.atleast_2d(value))[0]
 
             for vd in value_dec:
-                ann.append(time=start, duration=end-start, value=vd)
+                ann.append(time=start,
+                           duration=end-start,
+                           value=vd,
+                           confidence=confidence)
 
         return ann
 
@@ -185,9 +197,14 @@ def inverse(self, encoded, duration=None):
         ann = jams.Annotation(namespace=self.namespace, duration=duration)
 
         if np.isrealobj(encoded):
-            encoded = (encoded >= 0.5)
-
-        for vd in self.encoder.inverse_transform(np.atleast_2d(encoded))[0]:
-            ann.append(time=0, duration=duration, value=vd)
-
+            detected = (encoded >= 0.5)
+        else:
+            detected = encoded
+
+        for vd in self.encoder.inverse_transform(np.atleast_2d(detected))[0]:
+            vid = np.flatnonzero(self.encoder.transform(np.atleast_2d(vd)))
+            ann.append(time=0,
+                       duration=duration,
+                       value=vd,
+                       confidence=encoded[vid])
         return ann

tests/test_decode.py

@@ -101,6 +101,8 @@ def test_decode_tags_dynamic_hard(sr, hop_length, ann_tag):
     data = tc.transform_annotation(ann_tag, ann_tag.duration)
 
     inverse = tc.inverse(data['tags'], duration=ann_tag.duration)
+    for obs in inverse:
+        assert 0. <= obs.confidence <= 1.
     data2 = tc.transform_annotation(inverse, ann_tag.duration)
 
     assert np.allclose(data['tags'], data2['tags'])
@@ -119,6 +121,8 @@ def test_decode_tags_dynamic_soft(sr, hop_length, ann_tag):
     # Soften the data, but preserve the decisions
     tags_predict = data['tags'] * 0.51 + 0.1
     inverse = tc.inverse(tags_predict, duration=ann_tag.duration)
+    for obs in inverse:
+        assert 0. <= obs.confidence <= 1.
     data2 = tc.transform_annotation(inverse, ann_tag.duration)
 
     assert np.allclose(data['tags'], data2['tags'])
@@ -130,6 +134,8 @@ def test_decode_tags_static_hard(ann_tag):
 
     data = tc.transform_annotation(ann_tag, ann_tag.duration)
     inverse = tc.inverse(data['tags'], ann_tag.duration)
+    for obs in inverse:
+        assert 0. <= obs.confidence <= 1.
     data2 = tc.transform_annotation(inverse, ann_tag.duration)
 
     assert np.allclose(data['tags'], data2['tags'])
@@ -143,6 +149,8 @@ def test_decode_tags_static_soft(ann_tag):
     tags_predict = data['tags'] * 0.51 + 0.1
 
     inverse = tc.inverse(tags_predict, ann_tag.duration)
+    for obs in inverse:
+        assert 0. <= obs.confidence <= 1.
     data2 = tc.transform_annotation(inverse, ann_tag.duration)
 
     assert np.allclose(data['tags'], data2['tags'])
@@ -154,6 +162,8 @@ def test_decode_beat_hard(sr, hop_length, ann_beat):
 
     data = tc.transform_annotation(ann_beat, ann_beat.duration)
     inverse = tc.inverse(data['beat'], duration=ann_beat.duration)
+    for obs in inverse:
+        assert 0. <= obs.confidence <= 1.
     data2 = tc.transform_annotation(inverse, ann_beat.duration)
 
     assert np.allclose(data['beat'], data2['beat'])
@@ -167,6 +177,8 @@ def test_decode_beat_soft(sr, hop_length, ann_beat):
     beat_pred = data['beat'] * 0.51 + 0.1
 
     inverse = tc.inverse(beat_pred, duration=ann_beat.duration)
+    for obs in inverse:
+        assert 0. <= obs.confidence <= 1.
     data2 = tc.transform_annotation(inverse, ann_beat.duration)
 
     assert np.allclose(data['beat'], data2['beat'])
@@ -179,6 +191,8 @@ def test_decode_beat_downbeat_hard(sr, hop_length, ann_beat):
     data = tc.transform_annotation(ann_beat, ann_beat.duration)
     inverse = tc.inverse(data['beat'], downbeat=data['downbeat'],
                          duration=ann_beat.duration)
+    for obs in inverse:
+        assert 0. <= obs.confidence <= 1.
     data2 = tc.transform_annotation(inverse, ann_beat.duration)
 
     assert np.allclose(data['beat'], data2['beat'])
@@ -193,6 +207,8 @@ def test_decode_beat_downbeat_soft(sr, hop_length, ann_beat):
     dbeat_pred = data['downbeat'] * 0.51 + 0.1
     inverse = tc.inverse(beat_pred, downbeat=dbeat_pred,
                          duration=ann_beat.duration)
+    for obs in inverse:
+        assert 0. <= obs.confidence <= 1.
     data2 = tc.transform_annotation(inverse, ann_beat.duration)
 
     assert np.allclose(data['beat'], data2['beat'])
@@ -250,6 +266,8 @@ def test_decode_chordtag_hard_dense(sr, hop_length, ann_chord):
     data = tc.transform_annotation(ann_chord, ann_chord.duration)
 
     inverse = tc.inverse(data['chord'], duration=ann_chord.duration)
+    for obs in inverse:
+        assert 0 <= obs.confidence <= 1.
     data2 = tc.transform_annotation(inverse, ann_chord.duration)
 
     assert np.allclose(data['chord'], data2['chord'])
@@ -267,6 +285,10 @@ def test_decode_chordtag_soft_dense(sr, hop_length, ann_chord):
 
     chord_predict = data['chord'] * 0.51 + 0.1
     inverse = tc.inverse(chord_predict, duration=ann_chord.duration)
+
+    for obs in inverse:
+        assert 0 <= obs.confidence <= 1.
+
     data2 = tc.transform_annotation(inverse, ann_chord.duration)
 
     assert np.allclose(data['chord'], data2['chord'])
@@ -283,6 +305,8 @@ def test_decode_chordtag_hard_sparse_sparse(sr, hop_length, ann_chord):
     data = tc.transform_annotation(ann_chord, ann_chord.duration)
 
     inverse = tc.inverse(data['chord'], duration=ann_chord.duration)
+    for obs in inverse:
+        assert 0 <= obs.confidence <= 1.
     data2 = tc.transform_annotation(inverse, ann_chord.duration)
 
     assert np.allclose(data['chord'], data2['chord'])
@@ -305,6 +329,8 @@ def test_decode_chordtag_hard_dense_sparse(sr, hop_length, ann_chord):
 
     # Invert using the sparse encoder
     inverse = tcs.inverse(data['chord'], duration=ann_chord.duration)
+    for obs in inverse:
+        assert 0 <= obs.confidence <= 1.
     data2 = tcs.transform_annotation(inverse, ann_chord.duration)
 
     dense_positions = np.where(data['chord'])[1]
@@ -330,6 +356,8 @@ def test_decode_chordtag_soft_dense_sparse(sr, hop_length, ann_chord):
     chord_predict = data['chord'] * 0.51 + 0.1
     # Invert using the sparse encoder
     inverse = tcs.inverse(chord_predict, duration=ann_chord.duration)
+    for obs in inverse:
+        assert 0 <= obs.confidence <= 1.
     data2 = tcs.transform_annotation(inverse, ann_chord.duration)
 
     dense_positions = np.where(data['chord'])[1]