EngineX-Iluvatar/enginex-mr_series-sherpa-onnx
Archived
16
0
Fork 0
Code Issues Pull Requests Actions Projects Releases Wiki Activity

Fix nemo feature normalization in test code (#2361)

Browse Source
This commit is contained in:
Fangjun Kuang
2025-07-08 15:41:56 +08:00
committed by GitHub
parent 831aff187d
commit f1405779cf
4 changed files with 9 additions and 14 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
scripts/nemo/canary/test_180m_flash.py
View File

@@ -228,8 +228,8 @@ def main():
features = compute_features(audio, fbank) features = compute_features(audio, fbank)
if model.normalize_type != "": if model.normalize_type != "":
assert model.normalize_type == "per_feature", model.normalize_type assert model.normalize_type == "per_feature", model.normalize_type
mean = features.mean(axis=1, keepdims=True) mean = features.mean(axis=0, keepdims=True)
stddev = features.std(axis=1, keepdims=True) + 1e-5 stddev = features.std(axis=0, keepdims=True) + 1e-5
features = (features - mean) / stddev features = (features - mean) / stddev
features = np.expand_dims(features, axis=0) features = np.expand_dims(features, axis=0)

4
scripts/nemo/fast-conformer-hybrid-transducer-ctc/test-onnx-ctc-non-streaming.py
View File

@@ -140,8 +140,8 @@ def main():
if model.normalize_type != "": if model.normalize_type != "":
assert model.normalize_type == "per_feature", model.normalize_type assert model.normalize_type == "per_feature", model.normalize_type
features = torch.from_numpy(features) features = torch.from_numpy(features)
mean = features.mean(dim=1, keepdims=True) mean = features.mean(dim=0, keepdims=True)
stddev = features.std(dim=1, keepdims=True) + 1e-5 stddev = features.std(dim=0, keepdims=True) + 1e-5
features = (features - mean) / stddev features = (features - mean) / stddev
features = features.numpy() features = features.numpy()

11
scripts/nemo/fast-conformer-hybrid-transducer-ctc/test-onnx-transducer-non-streaming.py
View File

@@ -184,12 +184,7 @@ class OnnxModel:
target = torch.tensor([[token]], dtype=torch.int32).numpy() target = torch.tensor([[token]], dtype=torch.int32).numpy()
target_len = torch.tensor([1], 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()[0].name,
self.decoder.get_outputs()[1].name, self.decoder.get_outputs()[1].name,
@@ -267,8 +262,8 @@ def main():
if model.normalize_type != "": if model.normalize_type != "":
assert model.normalize_type == "per_feature", model.normalize_type assert model.normalize_type == "per_feature", model.normalize_type
features = torch.from_numpy(features) features = torch.from_numpy(features)
mean = features.mean(dim=1, keepdims=True) mean = features.mean(dim=0, keepdims=True)
stddev = features.std(dim=1, keepdims=True) + 1e-5 stddev = features.std(dim=0, keepdims=True) + 1e-5
features = (features - mean) / stddev features = (features - mean) / stddev
features = features.numpy() features = features.numpy()
print(audio.shape) print(audio.shape)

4
scripts/nemo/parakeet-tdt-0.6b-v2/test_onnx.py
View File

@@ -233,8 +233,8 @@ def main():
if model.normalize_type != "": if model.normalize_type != "":
assert model.normalize_type == "per_feature", model.normalize_type assert model.normalize_type == "per_feature", model.normalize_type
features = torch.from_numpy(features) features = torch.from_numpy(features)
mean = features.mean(dim=1, keepdims=True) mean = features.mean(dim=0, keepdims=True)
stddev = features.std(dim=1, keepdims=True) + 1e-5 stddev = features.std(dim=0, keepdims=True) + 1e-5
features = (features - mean) / stddev features = (features - mean) / stddev
features = features.numpy() features = features.numpy()
print(audio.shape) print(audio.shape)