Fix nemo feature normalization in test code (#2361)

csukuangfj · web-flow · commit f1405779cff9 · 2025-07-08T15:41:56.000+08:00
diff --git a/scripts/nemo/canary/test_180m_flash.py b/scripts/nemo/canary/test_180m_flash.py
@@ -228,8 +228,8 @@ def main():
     features = compute_features(audio, fbank)
     if model.normalize_type != "":
         assert model.normalize_type == "per_feature", model.normalize_type
-        mean = features.mean(axis=1, keepdims=True)
-        stddev = features.std(axis=1, keepdims=True) + 1e-5
+        mean = features.mean(axis=0, keepdims=True)
+        stddev = features.std(axis=0, keepdims=True) + 1e-5
         features = (features - mean) / stddev
 
     features = np.expand_dims(features, axis=0)
diff --git a/scripts/nemo/fast-conformer-hybrid-transducer-ctc/test-onnx-ctc-non-streaming.py b/scripts/nemo/fast-conformer-hybrid-transducer-ctc/test-onnx-ctc-non-streaming.py
@@ -140,8 +140,8 @@ def main():
     if model.normalize_type != "":
         assert model.normalize_type == "per_feature", model.normalize_type
         features = torch.from_numpy(features)
-        mean = features.mean(dim=1, keepdims=True)
-        stddev = features.std(dim=1, keepdims=True) + 1e-5
+        mean = features.mean(dim=0, keepdims=True)
+        stddev = features.std(dim=0, keepdims=True) + 1e-5
         features = (features - mean) / stddev
         features = features.numpy()
 
diff --git a/scripts/nemo/fast-conformer-hybrid-transducer-ctc/test-onnx-transducer-non-streaming.py b/scripts/nemo/fast-conformer-hybrid-transducer-ctc/test-onnx-transducer-non-streaming.py
@@ -184,12 +184,7 @@ def run_decoder(
         target = torch.tensor([[token]], dtype=torch.int32).numpy()
         target_len = torch.tensor([1], dtype=torch.int32).numpy()
 
-        (
-            decoder_out,
-            decoder_out_length,
-            state0_next,
-            state1_next,
-        ) = self.decoder.run(
+        (decoder_out, decoder_out_length, state0_next, state1_next,) = self.decoder.run(
             [
                 self.decoder.get_outputs()[0].name,
                 self.decoder.get_outputs()[1].name,
@@ -267,8 +262,8 @@ def main():
     if model.normalize_type != "":
         assert model.normalize_type == "per_feature", model.normalize_type
         features = torch.from_numpy(features)
-        mean = features.mean(dim=1, keepdims=True)
-        stddev = features.std(dim=1, keepdims=True) + 1e-5
+        mean = features.mean(dim=0, keepdims=True)
+        stddev = features.std(dim=0, keepdims=True) + 1e-5
         features = (features - mean) / stddev
         features = features.numpy()
     print(audio.shape)
diff --git a/scripts/nemo/parakeet-tdt-0.6b-v2/test_onnx.py b/scripts/nemo/parakeet-tdt-0.6b-v2/test_onnx.py
@@ -233,8 +233,8 @@ def main():
     if model.normalize_type != "":
         assert model.normalize_type == "per_feature", model.normalize_type
         features = torch.from_numpy(features)
-        mean = features.mean(dim=1, keepdims=True)
-        stddev = features.std(dim=1, keepdims=True) + 1e-5
+        mean = features.mean(dim=0, keepdims=True)
+        stddev = features.std(dim=0, keepdims=True) + 1e-5
         features = (features - mean) / stddev
         features = features.numpy()
     print(audio.shape)
