Use a stronger threshold in tensorflowpredictmaest

This was achieves by making sure that the input melspectrogram was computed exactly in the same way in TensorflowPredictMAEST and the original pytorch implementation, using MonoLoader with resampleQuality=1. While using resampleQuality=4 has a numerical impact in the results, it is not enough to change the prediction ranking according to informal observations (could not be the same in corner cases).
MTG · Oct 20, 2023 · 6b043f1 · 6b043f1
1 parent 53148b2
commit 6b043f1
Show file tree

Hide file tree

Showing 2 changed files with 2 additions and 2 deletions.
diff --git a/...c/unittests/machinelearning/tensorflowpredictmaest/preds_maest_discogs-maest-30s-pw-1.npy b/...c/unittests/machinelearning/tensorflowpredictmaest/preds_maest_discogs-maest-30s-pw-1.npy
diff --git a/test/src/unittests/machinelearning/test_tensorflowpredictmaest.py b/test/src/unittests/machinelearning/test_tensorflowpredictmaest.py
@@ -55,12 +55,12 @@ def testRegression(self):
         )
 
         filename = join(testdata.audio_dir, "recorded", "techno_loop.wav")
-        audio = MonoLoader(filename=filename, sampleRate=16000, resampleQuality=4)()
+        audio = MonoLoader(filename=filename, sampleRate=16000)()
 
         activations = self.model30s(audio)
         found = numpy.mean(activations, axis=0).squeeze()
 
-        self.assertAlmostEqualVector(found, expected, 1e-1)
+        self.assertAlmostEqualVector(found, expected, 1e-3)
 
     def testInvalidParam(self):
         self.assertConfigureFails(