Merge branch 'master' of https://github.com/danpovey/filtering

Author: zhuang

tests/test_resampler.py

@@ -10,7 +10,7 @@
 sys.path.insert(0, os.path.join(os.path.dirname(__file__), os.pardir))
 
 import lilfilter
-
+import librosa
 
 class TestResampler(unittest.TestCase):
     def test_constructor_and_forward(self):
@@ -37,8 +37,6 @@ def test_energy(self):
             n1, n2 = pair
             print("n1,n2 = {},{}".format(n1, n2))
 
-            a = lilfilter.Resampler(n1, n2, dtype=torch.float32)
-
             nyquist = math.pi * min(n2 / n1, 1)
             omega = 0.85 * nyquist  # enough less than nyquist that energy should be preserved.
             length = 500
@@ -48,38 +46,50 @@ def test_energy(self):
 
             input_energy = (signal * signal).sum().item()
             print("Energy of input signal is ", input_energy)
-            s = a.resample(signal)
-            s2 = lilfilter.resample(signal, n1, n2)
-            assert torch.all(torch.eq(s, s2))
 
-            b = lilfilter.Resampler(n2, n1, dtype=torch.float32)
-            t = b.resample(s)
+            s = lilfilter.resample(signal, n1, n2)
+            s_rosa = torch.tensor(librosa.core.resample(signal.numpy(), n1, n2))
+
+            t = lilfilter.resample(s, n2, n1)
+            t_rosa = torch.tensor(librosa.core.resample(s_rosa.numpy(), n2, n1))
 
             length = min(t.shape[1], signal.shape[1])
 
             sig1 = signal[:,:length]
             sig2 = t[:,:length]
-
-
             prod1 = (sig1 * sig1).sum()
             prod2 = (sig2 * sig2).sum()
             prod3 = (sig1 * sig2).sum()
 
+            length_rosa = min(t_rosa.shape[1], signal.shape[1])
+            sig1_rosa = signal[:,:length_rosa]
+            sig2_rosa = t_rosa[:,:length_rosa]
+            prod1_rosa = (sig1_rosa * sig1_rosa).sum()
+            prod2_rosa = (sig2_rosa * sig2_rosa).sum()
+            prod3_rosa = (sig1_rosa * sig2_rosa).sum()
+
 
-            print("The following numbers should be the same: {},{},{}".format(
+            print("The following [lilfilter] numbers should be the same: {},{},{}".format(
                 prod1, prod2, prod3))
 
             r1 = prod1 / prod2
             r2 = prod2 / prod3
             assert( abs(r1-1.0) < 0.001 and abs(r2-1.0) < 0.001)
 
+            print("The following [librosa] numbers should be the same: {},{},{}".format(
+                prod1_rosa, prod2_rosa, prod3_rosa))
+
+            r1_rosa = prod1_rosa / prod2_rosa
+            r2_rosa = prod2_rosa / prod3_rosa
+            #assert( abs(r1_rosa-1.0) < 0.001 and abs(r2_rosa-1.0) < 0.001)
+
 
             #plt.plot(np.arange(length), sig1.squeeze(0).numpy())
             #plt.plot(np.arange(length), sig2.squeeze(0).numpy())
             #plt.show()
 
-            print("Length of input signal is {}, downsampled {}, reconstructed {}".format(
-                    signal.shape[-1], s.shape[-1], t.shape[-1]))
+            print("Length of input signal is {}, downsampled {}, reconstructed {}, librosa-reconstructed".format(
+                    signal.shape[-1], s.shape[-1], t.shape[-1], t_rosa.shape[-1]))