Merge pull request #47 from bmcfee/random-output-selection

randomized target annotation selection

pumpp/sampler.py

@@ -85,13 +85,19 @@ def sample(self, data, interval):
         data_slice = dict()
 
         for key in data:
-            if key in self._time:
-                index = [slice(None)] * data[key].ndim
+            if '_valid' in key:
+                continue
 
-                if self._time[key] is not None:
-                    index[self._time[key]] = interval
+            index = [slice(None)] * data[key].ndim
 
-                data_slice[key] = data[key][index]
+            # if we have multiple observations for this key, pick one
+            index[0] = np.random.randint(0, data[key].shape[0])
+            index[0] = slice(index[0], index[0] + 1)
+
+            if self._time.get(key, None) is not None:
+                index[self._time[key]] = interval
+
+            data_slice[key] = data[key][index]
 
         return data_slice
 

tests/test_sampler.py

@@ -33,8 +33,8 @@ def ops(sr, hop_length):
                                        win_length=hop_length))
 
     # A time-varying annotation
-    ops.append(pumpp.task.BeatTransformer(name='beat', sr=sr,
-                                          hop_length=hop_length))
+    ops.append(pumpp.task.ChordTransformer(name='chord', sr=sr,
+                                           hop_length=hop_length))
 
     # And a static annotation
     ops.append(pumpp.task.VectorTransformer(namespace='vector',
@@ -80,10 +80,12 @@ def test_sampler(data, ops, n_samples, duration):
         # Now test that shape is preserved in the right way
         for key in datum:
             ref_shape = list(data[key].shape)
-            if sampler._time[key] is not None:
+            if sampler._time.get(key, None) is not None:
                 ref_shape[sampler._time[key]] = duration
 
-            assert list(datum[key].shape) == ref_shape
+            # Check that all keys have length=1
+            assert datum[key].shape[0] == 1
+            assert list(datum[key].shape[1:]) == ref_shape[1:]
 
     # Test that we got the right number of samples out
     if n_samples is None: