Fix librosa calls (#2208)

Summary:
Yesterday's release of librosa 0.9.0 made args keyword-only and changed default padding from "reflect" to "zero" for some functions. This PR adjusts callsites in our tutorials and tests accordingly.

Pull Request resolved: https://github.com/pytorch/audio/pull/2208

Reviewed By: mthrok

Differential Revision: D34099793

Pulled By: hwangjeff

fbshipit-source-id: 4e2642cdda8aae6d0a928befaf1bbb3873d229bc

examples/tutorials/audio_feature_extractions_tutorial.py

@@ -306,8 +306,8 @@ plot_mel_fbank(mel_filters, "Mel Filter Bank - torchaudio")
 
 
 mel_filters_librosa = librosa.filters.mel(
-    sample_rate,
-    n_fft,
+    sr=sample_rate,
+    n_fft=n_fft,
     n_mels=n_mels,
     fmin=0.0,
     fmax=sample_rate / 2.0,
@@ -365,7 +365,7 @@ plot_spectrogram(melspec[0], title="MelSpectrogram - torchaudio", ylabel="mel fr
 
 
 melspec_librosa = librosa.feature.melspectrogram(
-    waveform.numpy()[0],
+    y=waveform.numpy()[0],
     sr=sample_rate,
     n_fft=n_fft,
     hop_length=hop_length,

examples/tutorials/audio_resampling_tutorial.py

@@ -190,7 +190,7 @@ def benchmark_resample(
         waveform_np = waveform.squeeze().numpy()
         begin = time.time()
         for _ in range(iters):
-            librosa.resample(waveform_np, sample_rate, resample_rate, res_type=librosa_type)
+            librosa.resample(waveform_np, orig_sr=sample_rate, target_sr=resample_rate, res_type=librosa_type)
         elapsed = time.time() - begin
         return elapsed / iters
 
@@ -343,7 +343,7 @@ resampled_waveform = F.resample(
 plot_sweep(resampled_waveform, resample_rate, title="Kaiser Window Best (torchaudio)")
 
 librosa_resampled_waveform = torch.from_numpy(
-    librosa.resample(waveform.squeeze().numpy(), sample_rate, resample_rate, res_type="kaiser_best")
+    librosa.resample(waveform.squeeze().numpy(), orig_sr=sample_rate, target_sr=resample_rate, res_type="kaiser_best")
 ).unsqueeze(0)
 plot_sweep(librosa_resampled_waveform, resample_rate, title="Kaiser Window Best (librosa)")
 
@@ -363,7 +363,7 @@ resampled_waveform = F.resample(
 plot_specgram(resampled_waveform, resample_rate, title="Kaiser Window Fast (torchaudio)")
 
 librosa_resampled_waveform = torch.from_numpy(
-    librosa.resample(waveform.squeeze().numpy(), sample_rate, resample_rate, res_type="kaiser_fast")
+    librosa.resample(waveform.squeeze().numpy(), orig_sr=sample_rate, target_sr=resample_rate, res_type="kaiser_fast")
 ).unsqueeze(0)
 plot_sweep(librosa_resampled_waveform, resample_rate, title="Kaiser Window Fast (librosa)")
 

test/torchaudio_unittest/functional/librosa_compatibility_test_impl.py

@@ -62,6 +62,7 @@ class Functional(TestBaseMixin):
             momentum=momentum,
             init=None,
             length=waveform.size(1),
+            pad_mode="reflect",
         )[None, ...]
         self.assertEqual(result, torch.from_numpy(expected), atol=5e-5, rtol=1e-07)
 

test/torchaudio_unittest/transforms/librosa_compatibility_test_impl.py

@@ -36,7 +36,7 @@ class TransformsTestBase(TestBaseMixin):
         ).to(self.device, self.dtype)
 
         expected = librosa.core.spectrum._spectrogram(
-            y=waveform[0].cpu().numpy(), n_fft=n_fft, hop_length=hop_length, power=power
+            y=waveform[0].cpu().numpy(), n_fft=n_fft, hop_length=hop_length, power=power, pad_mode="reflect"
         )[0]
 
         result = T.Spectrogram(n_fft=n_fft, hop_length=hop_length, power=power,).to(self.device, self.dtype)(
@@ -54,7 +54,7 @@ class TransformsTestBase(TestBaseMixin):
         ).to(self.device, self.dtype)
 
         expected = librosa.core.spectrum._spectrogram(
-            y=waveform[0].cpu().numpy(), n_fft=n_fft, hop_length=hop_length, power=1
+            y=waveform[0].cpu().numpy(), n_fft=n_fft, hop_length=hop_length, power=1, pad_mode="reflect"
         )[0]
 
         result = T.Spectrogram(n_fft=n_fft, hop_length=hop_length, power=None, return_complex=True,).to(
@@ -86,6 +86,7 @@ class TransformsTestBase(TestBaseMixin):
             n_mels=n_mels,
             norm=norm,
             htk=mel_scale == "htk",
+            pad_mode="reflect",
         )
         result = T.MelSpectrogram(
             sample_rate=sample_rate,
@@ -136,6 +137,7 @@ class TransformsTestBase(TestBaseMixin):
             n_mels=n_mels,
             htk=True,
             norm=None,
+            pad_mode="reflect",
         )
         expected = librosa.feature.mfcc(
             S=librosa.core.spectrum.power_to_db(melspec), n_mfcc=n_mfcc, dct_type=2, norm="ortho"
@@ -157,6 +159,6 @@ class TransformsTestBase(TestBaseMixin):
             self.device, self.dtype
         )(waveform)
         expected = librosa.feature.spectral_centroid(
-            y=waveform[0].cpu().numpy(), sr=sample_rate, n_fft=n_fft, hop_length=hop_length
+            y=waveform[0].cpu().numpy(), sr=sample_rate, n_fft=n_fft, hop_length=hop_length, pad_mode="reflect"
         )
         self.assertEqual(result, torch.from_numpy(expected), atol=5e-4, rtol=1e-5)