"throughput_x_realtime": 359.0296459040717,
Device: npu
Throughput: 359.03x realtime

.
├── .gitignore
├── README.md
├── assets/accuracy_eval_result.png
├── assets/env_check.png
├── assets/git_submit_result.png
├── assets/inference_result.png
├── assets/performance_eval_result.png
├── eval/eval_accuracy.py
├── eval/eval_performance.py
├── inference.py
├── locked_models.md
├── logs/accuracy_eval.log
├── logs/env_check.log
├── logs/inference.log
├── logs/performance_eval.log
├── requirements.txt
├── results/accuracy_eval.json
├── results/env_info.json
├── results/inference_result.json
└── results/performance_eval.json

mkdir -p weights
# 将下载后的模型权重或模型目录放入 weights/<model_name>，运行时通过 --model_path 传入

pip install -r requirements.txt
python inference.py --model_path <model_path> --audio <audio.wav> --device npu

python eval/eval_accuracy.py --model_path <model_path> --device npu
python eval/eval_performance.py --model_path <model_path> --device npu

python eval/eval_accuracy.py --model_path <model_path> --device npu
python eval/eval_performance.py --model_path <model_path> --device npu

python inference.py --help
python inference.py --model_path <model_path> --audio <audio.wav> --device npu

+------------------------------------------------------------------------------------------------+
| npu-smi 25.5.2                   Version: 25.5.2                                               |
+---------------------------+---------------+----------------------------------------------------+
| NPU   Name                | Health        | Power(W)    Temp(C)           Hugepages-Usage(page)|
| Chip  Phy-ID              | Bus-Id        | AICore(%)   Memory-Usage(MB)  HBM-Usage(MB)        |
+===========================+===============+====================================================+
| 3     Ascend910           | OK            | 167.4       42                0    / 0             |
| 0     6                   | 0000:0A:00.0  | 0           0    / 0          3102 / 65536         |
+------------------------------------------------------------------------------------------------+
| 3     Ascend910           | OK            | -           41                0    / 0             |
| 1     7                   | 0000:0B:00.0  | 0           0    / 0          2870 / 65536         |
+===========================+===============+====================================================+
+---------------------------+---------------+----------------------------------------------------+
| NPU     Chip              | Process id    | Process name             | Process memory(MB)      |
+===========================+===============+====================================================+
| No running processes found in NPU 3                                                            |
+===========================+===============+====================================================+
---
[LOG_WARNING] can not create directory, directory: /home/atomgit/ascend/log, possible reason: No such file or directory.path string is NULLpath string is NULLPyTorch: 2.9.0+cpu
torch_npu: 2.9.0.post1+gitee7ba04
transformers: 4.57.6
torchaudio: 2.9.0
numpy: 1.26.4
NPU available: True
NPU count: 2
NPU name: Ascend910_9362

{
  "os": "Linux-5.10.0-182.0.0.95.r2220_156.hce2.aarch64-aarch64-with-glibc2.35",
  "python": "3.11.14",
  "torch": "2.9.0+cpu",
  "torch_npu": "2.9.0.post1+gitee7ba04",
  "transformers": "4.57.6",
  "torchaudio": "2.9.0",
  "numpy": "1.26.4",
  "npu_available": "True",
  "npu_count": "2",
  "npu_name": "Ascend910_9362",
  "cann_version": "/usr/local/Ascend/cann-8.5.1",
  "soc_version": "ascend910_9391",
  "ascend_visible_devices": "7,6"
}

============================================================
CNN8RNN-AudioSet-SED NPU Inference
============================================================
Device: npu
Dtype: float16
Model path: ./weights
Audio path: ./test_audio.wav
Model classes: 447
Model load time: 1.59s

Top-10 Predictions:
--------------------------------------------------
   1. Tuning fork                               0.8774
   2. Background noise                          0.5771
   3. Mechanisms                                0.2507
   4. Tick                                      0.1886
   5. Sine wave                                 0.1439
   6. Noise                                     0.0600
   7. Generic impact sounds                     0.0570
   8. Wind                                      0.0508
   9. Breathing                                 0.0428
  10. Human sounds                              0.0373

Performance Summary:
--------------------------------------------------
  Audio duration:    3.00s
  Inference time:    0.0084s
  Throughput:        359.03x realtime
  Device info:       Ascend NPU (Ascend910_9362), Memory: 12.5 MB
  Dtype:             float16
  Framewise shape:   (301, 447)
  Clipwise shape:    (447,)

{
  "model": "cnn8rnn-audioset-sed",
  "audio_path": "./test_audio.wav",
  "audio_duration_s": 3.0,
  "inference_time_s": 0.008355855941772461,
  "throughput_x_realtime": 359.0296459040717,
  "device": "npu",
  "device_info": "Ascend NPU (Ascend910_9362), Memory: 12.5 MB",
  "dtype": "float16",
  "top_k_predictions": [
    {
      "rank": 1,
      "class": "Tuning fork",
      "score": 0.87744140625
    },
    {
      "rank": 2,
      "class": "Background noise",
      "score": 0.5771484375
    },
    {
      "rank": 3,
      "class": "Mechanisms",
      "score": 0.250732421875
    },
    {
      "rank": 4,
      "class": "Tick",
      "score": 0.1885986328125
    },
    {
      "rank": 5,
      "class": "Sine wave",
      "score": 0.1439208984375
    },
    {
      "rank": 6,
      "class": "Noise",
      "score": 0.059967041015625
    },
    {
      "rank": 7,
      "class": "Generic impact sounds",
      "score": 0.057037353515625
    },
    {
      "rank": 8,
      "class": "Wind",
      "score": 0.05084228515625
    },
    {
      "rank": 9,
      "class": "Breathing",
      "score": 0.0428466796875
    },
    {
      "rank": 10,
      "class": "Human sounds",
      "score": 0.03729248046875
    }
  ],
  "framewise_shape": [
    301,
    447
  ],
  "clipwise_shape": [
    447
  ]
}

============================================================
CNN8RNN-AudioSet-SED Accuracy Evaluation
============================================================
NPU: Ascend910_9362
Model: ./weights
Audio: ./test_audio.wav
Test device: npu, Reference device: cpu
Dtype: float16, Num runs: 3

Audio shape: torch.Size([96000])
Loading reference model on CPU (float32) ...
Loading test model on NPU (float16) ...

--- Run 1/3 ---
  Clipwise  - max_rel_err: 80.5835%, mean_rel_err: 8.7475%, cos_sim: 0.99999954
  Framewise - max_rel_err: 2.4338%, mean_rel_err: 0.3115%, cos_sim: 0.99999962

--- Run 2/3 ---
  Clipwise  - max_rel_err: 80.5835%, mean_rel_err: 8.7475%, cos_sim: 0.99999954
  Framewise - max_rel_err: 2.4338%, mean_rel_err: 0.3115%, cos_sim: 0.99999962

--- Run 3/3 ---
  Clipwise  - max_rel_err: 80.5835%, mean_rel_err: 8.7475%, cos_sim: 0.99999954
  Framewise - max_rel_err: 2.4338%, mean_rel_err: 0.3115%, cos_sim: 0.99999962

============================================================
AVERAGED RESULTS
============================================================
Clipwise  - avg max_rel_err: 80.5835%, avg mean_rel_err: 8.7475%, avg cos_sim: 0.99999954
Framewise - avg max_rel_err: 2.4338%, avg mean_rel_err: 0.3115%, avg cos_sim: 0.99999962

Framewise mean relative error: 0.3115%
Clipwise mean relative error:  8.7475%
Min cosine similarity: 0.99999954
Threshold: framewise mean_rel_err < 1.0% AND cos_sim > 0.999
Result: PASS
Note: Clipwise mean_rel_err > 1% is expected due to float32->float16 precision
      loss amplified by temporal pooling. Cosine similarity > 0.999 confirms correctness.

Results saved to results/accuracy_eval.json

{
  "model": "cnn8rnn-audioset-sed",
  "audio_path": "./test_audio.wav",
  "test_device": "npu",
  "reference_device": "cpu",
  "test_dtype": "float16",
  "reference_dtype": "float32",
  "num_runs": 3,
  "clipwise_avg": {
    "name": "clipwise_output_avg",
    "max_relative_error_pct": 80.58349945934312,
    "mean_relative_error_pct": 8.747456158062036,
    "cosine_similarity": 0.9999995355344452
  },
  "framewise_avg": {
    "name": "framewise_output_avg",
    "max_relative_error_pct": 2.4338294025746525,
    "mean_relative_error_pct": 0.31153602345792714,
    "cosine_similarity": 0.9999996165961073
  },
  "framewise_mean_rel_err_pct": 0.31153602345792714,
  "clipwise_mean_rel_err_pct": 8.747456158062036,
  "min_cosine_similarity": 0.9999995355344452,
  "threshold_mean_rel_err_pct": 1.0,
  "threshold_cos_sim": 0.999,
  "passed": true,
  "per_run_clipwise": [
    {
      "name": "clipwise_output",
      "shape": [
        447
      ],
      "max_absolute_error": 0.0007358789443969727,
      "mean_absolute_error": 3.063117928506525e-05,
      "max_relative_error": 0.8058349945934311,
      "mean_relative_error": 0.08747456158062036,
      "cosine_similarity": 0.9999995355344452,
      "max_relative_error_pct": 80.58349945934312,
      "mean_relative_error_pct": 8.747456158062036,
      "max_abs_ratio_pct": 0.08379617154798935,
      "significant_values_compared": 392,
      "total_values": 447
    },
    {
      "name": "clipwise_output",
      "shape": [
        447
      ],
      "max_absolute_error": 0.0007358789443969727,
      "mean_absolute_error": 3.063117928506525e-05,
      "max_relative_error": 0.8058349945934311,
      "mean_relative_error": 0.08747456158062036,
      "cosine_similarity": 0.9999995355344452,
      "max_relative_error_pct": 80.58349945934312,
      "mean_relative_error_pct": 8.747456158062036,
      "max_abs_ratio_pct": 0.08379617154798935,
      "significant_values_compared": 392,
      "total_values": 447
    },
    {
      "name": "clipwise_output",
      "shape": [
        447
      ],
      "max_absolute_error": 0.0007358789443969727,
      "mean_absolute_error": 3.063117928506525e-05,
      "max_relative_error": 0.8058349945934311,
      "mean_relative_error": 0.08747456158062036,
      "cosine_similarity": 0.9999995355344452,
      "max_relative_error_pct": 80.58349945934312,
      "mean_relative_error_pct": 8.747456158062036,
      "max_abs_ratio_pct": 0.08379617154798935,
      "significant_values_compared": 392,
      "total_values": 447
    }
  ],
  "per_run_framewise": [
    {
      "name": "framewise_output",
      "shape": [
        301,
        447
      ],
      "max_absolute_error": 0.0017389357089996338,
      "mean_absolute_error": 6.282809046775297e-06,
      "max_relative_error": 0.024338294025746526,
      "mean_relative_error": 0.003115360234579271,
      "cosine_similarity": 0.9999996165961073,
      "max_relative_error_pct": 2.4338294025746525,
      "mean_relative_error_pct": 0.31153602345792714,
      "max_abs_ratio_pct": 0.18362858441429536,
      "significant_values_compared": 73929,
      "total_values": 134547
    },
    {
      "name": "framewise_output",
      "shape": [
        301,
        447
      ],
      "max_absolute_error": 0.0017389357089996338,
      "mean_absolute_error": 6.282809046775297e-06,
      "max_relative_error": 0.024338294025746526,
      "mean_relative_error": 0.003115360234579271,
      "cosine_similarity": 0.9999996165961073,
      "max_relative_error_pct": 2.4338294025746525,
      "mean_relative_error_pct": 0.31153602345792714,
      "max_abs_ratio_pct": 0.18362858441429536,
      "significant_values_compared": 73929,
      "total_values": 134547
    },
    {
      "name": "framewise_output",
      "shape": [
        301,
        447
      ],
      "max_absolute_error": 0.0017389357089996338,
      "mean_absolute_error": 6.282809046775297e-06,
      "max_relative_error": 0.024338294025746526,
      "mean_relative_error": 0.003115360234579271,
      "cosine_similarity": 0.9999996165961073,
      "max_relative_error_pct": 2.4338294025746525,
      "mean_relative_error_pct": 0.31153602345792714,
      "max_abs_ratio_pct": 0.18362858441429536,
      "significant_values_compared": 73929,
      "total_values": 134547
    }
  ]
}

============================================================
CNN8RNN-AudioSet-SED Performance Evaluation
============================================================
NPU: Ascend910_9362
Model: ./weights
Audio: ./test_audio.wav
Device: npu, Dtype: float16
Warmup: 3, Num runs: 10

Audio duration: 3.00s, shape: torch.Size([96000])

Warmup (3 iterations) ...
Warmup done.

Benchmarking (10 iterations) ...
  Run 1: 0.0056s
  Run 2: 0.0056s
  Run 3: 0.0055s
  Run 4: 0.0055s
  Run 5: 0.0055s
  Run 6: 0.0055s
  Run 7: 0.0055s
  Run 8: 0.0055s
  Run 9: 0.0055s
  Run 10: 0.0054s

============================================================
PERFORMANCE RESULTS
============================================================
  Audio duration:      3.00s
  Avg latency:         0.0055s
  Std latency:         0.0000s
  Min latency:         0.0054s
  Max latency:         0.0056s
  P50 latency:         0.0055s
  P90 latency:         0.0056s
  P99 latency:         0.0056s
  Throughput:          545.06x realtime
  NPU memory allocated: 13.1 MB
  NPU memory reserved:  72.0 MB
  NPU memory peak:      35.0 MB

Results saved to results/performance_eval.json

{
  "model": "cnn8rnn-audioset-sed",
  "audio_path": "./test_audio.wav",
  "audio_duration_s": 3.0,
  "device": "npu",
  "dtype": "float16",
  "warmup": 3,
  "num_runs": 10,
  "avg_latency_s": 0.005504012107849121,
  "std_latency_s": 4.8806356303092526e-05,
  "min_latency_s": 0.005448579788208008,
  "max_latency_s": 0.005597352981567383,
  "p50_latency_s": 0.00548398494720459,
  "p90_latency_s": 0.005575251579284668,
  "p99_latency_s": 0.005595142841339111,
  "throughput_x_realtime": 545.0569405037794,
  "npu_memory": {
    "allocated_mb": 13.11572265625,
    "reserved_mb": 72.0,
    "max_allocated_mb": 35.03076171875
  },
  "per_run_latencies_s": [
    0.005597352981567383,
    0.005572795867919922,
    0.005488395690917969,
    0.005547046661376953,
    0.005507946014404297,
    0.005473136901855469,
    0.005473136901855469,
    0.005479574203491211,
    0.005452156066894531,
    0.005448579788208008
  ]
}