Merge pull request #13 from bmcfee/tag-vocab

added automatic tag vocab inference

pumpp/task/tags.py

@@ -5,14 +5,16 @@
 import numpy as np
 from sklearn.preprocessing import MultiLabelBinarizer
 
+import jams
+
 from .base import BaseTaskTransformer
 
 __all__ = ['DynamicLabelTransformer', 'StaticLabelTransformer']
 
 
 class DynamicLabelTransformer(BaseTaskTransformer):
 
-    def __init__(self, namespace, name, labels, sr=22050, hop_length=512):
+    def __init__(self, namespace, name, labels=None, sr=22050, hop_length=512):
         '''Initialize a time-series label transformer
 
         Parameters
@@ -32,6 +34,9 @@ def __init__(self, namespace, name, labels, sr=22050, hop_length=512):
                                                       sr=sr,
                                                       hop_length=hop_length)
 
+        if labels is None:
+            labels = jams.schema.values(namespace)
+
         self.encoder = MultiLabelBinarizer()
         self.encoder.fit([labels])
         self._classes = set(self.encoder.classes_)
@@ -63,7 +68,7 @@ def transform_annotation(self, ann, duration):
 
 class StaticLabelTransformer(BaseTaskTransformer):
 
-    def __init__(self, namespace, name, labels):
+    def __init__(self, namespace, name, labels=None):
         '''Initialize a global label transformer
 
         Parameters
@@ -76,6 +81,9 @@ def __init__(self, namespace, name, labels):
                                                      namespace=namespace,
                                                      sr=1, hop_length=1)
 
+        if labels is None:
+            labels = jams.schema.values(namespace)
+
         self.encoder = MultiLabelBinarizer()
         self.encoder.fit([labels])
         self._classes = set(self.encoder.classes_)

tests/test_task.py

@@ -202,7 +202,7 @@ def test_task_dlabel_present(SR, HOP_LENGTH):
 
     ann.append(time=0, duration=1.0, value='alpha')
     ann.append(time=0, duration=1.0, value='beta')
-    ann.append(time=1, duration=1.0, value='23')
+    ann.append(time=1, duration=1.0, value='some nonsense')
     ann.append(time=3, duration=1.0, value='disco')
 
     jam.annotations.append(ann)
@@ -259,6 +259,27 @@ def test_task_dlabel_absent(SR, HOP_LENGTH):
         assert type_match(output[key].dtype, trans.fields[key].dtype)
 
 
+def test_task_dlabel_auto(SR, HOP_LENGTH):
+    jam = jams.JAMS(file_metadata=dict(duration=4.0))
+    trans = pumpp.task.DynamicLabelTransformer(namespace='tag_gtzan',
+                                               name='genre')
+
+    output = trans.transform(jam)
+
+    # Mask should be false since we have no matching namespace
+    assert not np.any(output['genre/_valid'])
+
+    y = output['genre/tags']
+
+    # Check the shape
+    assert y.shape == (1, 4 * (SR // HOP_LENGTH), 10)
+
+    # Make sure it's empty
+    assert not np.any(y)
+    for key in trans.fields:
+        assert shape_match(output[key].shape[1:], trans.fields[key].shape)
+        assert type_match(output[key].dtype, trans.fields[key].dtype)
+
 
 def test_task_slabel_absent():
     labels = ['alpha', 'beta', 'psycho', 'aqua', 'disco']
@@ -294,7 +315,7 @@ def test_task_slabel_present():
 
     ann.append(time=0, duration=1.0, value='alpha')
     ann.append(time=0, duration=1.0, value='beta')
-    ann.append(time=1, duration=1.0, value='23')
+    ann.append(time=1, duration=1.0, value='some nonsense')
     ann.append(time=3, duration=1.0, value='disco')
 
     jam.annotations.append(ann)
@@ -323,6 +344,27 @@ def test_task_slabel_present():
         assert type_match(output[key].dtype, trans.fields[key].dtype)
 
 
+def test_task_slabel_auto():
+    jam = jams.JAMS(file_metadata=dict(duration=4.0))
+    trans = pumpp.task.StaticLabelTransformer(namespace='tag_gtzan',
+                                              name='genre')
+
+    output = trans.transform(jam)
+
+    # Mask should be false since we have no matching namespace
+    assert not np.any(output['genre/_valid'])
+
+    # Check the shape
+    assert output['genre/tags'].ndim == 2
+    assert output['genre/tags'].shape[1] == 10
+
+    # Make sure it's empty
+    assert not np.any(output['genre/tags'])
+
+    for key in trans.fields:
+        assert shape_match(output[key].shape[1:], trans.fields[key].shape)
+        assert type_match(output[key].dtype, trans.fields[key].dtype)
+
 
 @pytest.mark.parametrize('dimension', [1, 2, 4])
 @pytest.mark.parametrize('name', ['collab', 'vec'])
@@ -532,7 +574,7 @@ def test_transform_query():
 
     ann.append(time=0, duration=1.0, value='alpha')
     ann.append(time=0, duration=1.0, value='beta')
-    ann.append(time=1, duration=1.0, value='23')
+    ann.append(time=1, duration=1.0, value='some nonsense')
     ann.append(time=3, duration=1.0, value='disco')
 
     jam.annotations.append(ann)