cd WeSpeaker-ResNet34-LM-MLX-ascend/
python inference.py --precision_test
音频推理
cd WeSpeaker-ResNet34-LM-MLX-ascend/
python3 inference.py --audio_path /tmp/test_audio.wav
参数说明
参数
说明
默认值
--model_path
模型权重路径
WeSpeaker-ResNet34-LM-MLX
--audio_path
待推理音频路径
必需(精度测试时不需要)
--precision_test
运行精度测试
False
--device
运行设备
npu:0
--no_warmup
跳过预热阶段
False
精度测试结果
========================================================
Precision Comparison: CPU vs NPU
========================================================
Max errors: sum=1.91e-06, mean=1.49e-08, std=2.98e-08
PASS: NPU precision within 1% of CPU
========================================================
PRECISION TEST PASSED
========================================================
指标
阈值
实测值
状态
max_error_sum
< 1e-3
1.91e-06
✅ PASS
max_error_mean
< 1e-5
1.49e-08
✅ PASS
max_error_std
< 1e-5
2.98e-08
✅ PASS
输出示例
2026-05-18 08:16:07,750 - INFO - WeSpeaker-ResNet34-LM-MLX Ascend NPU Inference
2026-05-18 08:16:07,600 - INFO - Model loaded on device: npu:0
2026-05-18 08:16:07,600 - INFO - 开始预热...
2026-05-18 08:16:07,834 - INFO - 预热完成
2026-05-18 08:16:07,834 - INFO - 开始推理...
2026-05-18 08:16:07,837 - INFO - Inference time: 0.0045s
2026-05-18 08:16:07,837 - INFO - Embedding shape: torch.Size([1, 256])
2026-05-18 08:16:07,837 - INFO - 推理成功完成!