NPU 标签依据

本仓库作为昇腾 NPU 模型仓库发布。本 README 顶部的模型卡片元数据使用了确切的标量字段 hardware: NPU，且标签列表包含 NPU、Ascend 和 ascend-npu。仓库描述或模型卡片在 AtomGit 或 GitCode 上还应包含 #+NPU 标签。

项目	数值
仓库	https://gitcode.com/nanyizjm/hubert-base-deepfake-npu
竞赛任务	Track 1 model adaptation
硬件元数据	hardware: NPU
所需标签	#+NPU
README 数据策略	推理、精度和性能数值以文本形式写入本 README；不使用图像替代数据。

Track 1 模型卡片摘要

项目	数值
模型仓库	https://gitcode.com/nanyizjm/hubert-base-deepfake-npu
原始模型或权重来源	https://gitcode.com/hf_mirrors/abhishtagatya/hubert-base-960h-itw-deepfake
竞赛赛道	Track 1: model adaptation
目标硬件	Ascend NPU
所需功能	NPU 推理成功运行或明确记录阻塞原因
所需精度	与 CPU/GPU 参考结果相比，NPU 结果误差小于 1%
所需标签	#+NPU

交付物清单

交付物	状态
inference.py	已提供
readme.md / README.md	已提供
eval/eval_accuracy.py	已提供
eval/eval_performance.py	已提供
logs 目录	已提供
results 目录	已提供
assets 或截图证据	已提供

精度证据要求

README 必须包含明确的 CPU/GPU 与 NPU 对比数值数据。关键验收目标是误差小于 1%。相应的结构化证据应在可用时保存于 results/accuracy_eval.json 和 logs/accuracy_eval.log 中。

#+NPU

hubert-base-deepfake on Ascend NPU

平台审核证据摘要（直接文本）

本部分直接写入 README 供平台审核使用。仅使用本仓库中已签入的日志和 JSON 结果文件，不依赖嵌入式图片。

审核项	直接结果
仓库	`hubert-base-deepfake-npu`
硬件元数据	本 README 中存在 `hardware: NPU` 和 `#+NPU`
正常 NPU 推理输出	通过 - 已签入的 NPU 推理输出如下所示。
精度要求	通过 - 选定的可复现误差 0.00003410931458347477% 低于 1%。
性能证据	可用 - 已签入的性能指标如下所示。
证据文件	`logs/inference.log`、`results/accuracy_eval.json`、`results/performance_eval.json`、`logs/accuracy_eval.log`、`logs/performance_eval.log`

正常 NPU 推理输出证据

{"model": "hubert-base-960h-itw-deepfake", "model_path": "./weights", "audio_path": "./test_audio_5s.wav", "audio_duration_s": 5.0, "sample_rate": 16000, "device": "npu:0", "dtype": "torch.float32", "prediction": "bona-fide", "prediction_id

NPU 推理指标

来源	指标	数值
`logs/inference.log`	`audio_path`	`./test_audio_5s.wav`
`logs/inference.log`	`audio_duration_s`	`5`
`logs/inference.log`	`device`	`npu:0`

CPU/GPU 参考与 NPU 精度验证

来源	指标	数值
`results/accuracy_eval.json`	`evaluations[0].logits.cpu_logits`	`[3.1859610080718994,-3.1582372188568115]`
`results/accuracy_eval.json`	`evaluations[0].logits.npu_logits`	`[3.1860363483428955,-3.157881021499634]`
`results/accuracy_eval.json`	`evaluations[0].logits.max_relative_error`	`0.00005639497976517305`
`results/accuracy_eval.json`	`evaluations[0].logits.mean_relative_error`	`0.00003410931458347477`
`results/accuracy_eval.json`	`evaluations[0].logits.filtered_mean_relative_error`	`0.00003410931458347477`
`results/accuracy_eval.json`	`evaluations[0].logits.cosine_similarity`	`1.0000001192092896`
`results/accuracy_eval.json`	`evaluations[0].prediction.cpu_prediction`	`bona-fide`
`results/accuracy_eval.json`	`evaluations[0].prediction.npu_prediction`	`bona-fide`
`results/accuracy_eval.json`	`evaluations[0].pass_criteria.filtered_mre_lt_1pct`	`true`
`results/accuracy_eval.json`	`evaluations[0].pass_criteria.cosine_similarity_gt_09999`	`true`

精度结论：PASS - 所选可复现误差 0.00003410931458347477% 低于 1%。

性能验证

来源	指标	数值
`results/performance_eval.json`	`device`	`npu:0`
`results/performance_eval.json`	`dtype`	`float32`
`results/performance_eval.json`	`warmup_runs`	`3`
`results/performance_eval.json`	`num_runs`	`10`
`results/performance_eval.json`	`benchmarks[0].avg_latency_s`	`0.0067`
`results/performance_eval.json`	`benchmarks[0].std_latency_s`	`0`
`results/performance_eval.json`	`benchmarks[0].min_latency_s`	`0.0066`
`results/performance_eval.json`	`benchmarks[0].max_latency_s`	`0.0067`
`results/performance_eval.json`	`benchmarks[0].p50_latency_s`	`0.0067`
`results/performance_eval.json`	`benchmarks[0].p95_latency_s`	`0.0067`

HuBERT-Base-Deepfake on Ascend NPU

1. 简介

本文档记录 HuBERT-Base-Deepfake 在华为昇腾 NPU 环境下的适配验证、推理部署与评测结果整理。

HuBERT-Base-Deepfake 的当前适配任务类型为：模型推理适配。仓库围绕 赛道一模型适配 交付要求，提供 NPU 推理脚本、精度评测、性能评测、运行日志、结果文件和文本化自验证证据。

2. 适配内容

2.1 NPU 推理适配

仓库提供 inference.py 作为统一推理入口，运行时通过 --device npu 或脚本默认设备在昇腾 NPU 上执行推理。推理代码保留 model.eval()、无梯度推理、输入输出摘要、耗时统计和日志保存逻辑，便于复现与核验。

2.2 精度与性能评测

仓库保留精度评测与性能评测材料。精度验证以 CPU/GPU 参考输出与 NPU 输出进行对比，目标为误差小于 1%；性能验证记录延迟、吞吐、batch size、输入尺寸/长度、dtype、NPU 内存等信息。所有结果以 logs/ 与 results/ 中的真实运行文件为准。

2.3 证据文本化与提交整理

自验证截图中的关键内容已转写为 README 文本证据，避免仅依赖图片展示。仓库 README、日志、JSON 结果和附件材料均用于 AtomGit/GitCode 公开提交，README 顶部已声明 hardware: NPU 与 #+NPU 标签。

3. 环境要求

组件	版本 / 说明
操作系统	Linux 5.10.0 aarch64
Python	3.11.14
NPU 型号	Ascend910_9362
NPU 数量	2
CANN	8.5.1
PyTorch	2.9.0+cpu
torch_npu	2.9.0.post1
transformers	4.57.6
依赖安装	`pip install -r requirements.txt`

NPU：Ascend NPU（具体型号以 results/env_info.json 或 logs/env_check.log 为准）
Python：3.8+，推荐使用比赛 / 适配容器中的 Python 版本
说明：如本地环境缺少 NPU、CANN 或 torch_npu，请先完成昇腾基础环境配置后再运行真实验证。

4. 快速开始

4.1 目录结构

.
├── .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/__init__.py
├── 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/performance_eval.json

4.2 权重准备

本仓库不提交大体积模型权重；请按原模型发布页、ModelScope、GitCode 或 HuggingFace 镜像下载后通过参数传入。

推荐约定：

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

4.3 NPU 推理

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

4.4 精度与性能评测

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

5. 验证结果

5.1 模型信息

指标	结果
模型名称	`hubert-base-960h-itw-deepfake`
任务类型	模型推理适配
推理设备	Ascend NPU
推理框架	PyTorch / torch_npu 或仓库脚本声明的推理框架
仓库分支	`main`
当前提交	`7bdf09a`

5.2 推理性能

测试结果来源：results/performance_eval.json

指标	结果
`device`	npu:0
`dtype`	float32
`num_runs`	10

5.3 NPU vs CPU/GPU 精度对比

结果来源：results/accuracy_eval.json

指标	结果
结果	下方“结果数据直接文本”已写入实际日志/JSON内容

结论：README 仅记录仓库中已有的真实评测数据；若某项指标未在 JSON/日志中出现，请以对应日志文件为准，不在文档中补造数值。

5.4 精度性能评测脚本

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

关键日志和结构化 JSON 已在下方“结果数据直接文本”中直接写入；原始文件路径仅用于复核。

6. 推理脚本说明

inference.py 支持的参数以脚本自身 --help 输出为准。当前 README 从脚本中提取到的主要参数如下：

参数	默认值	说明
`--model_path`	见脚本默认值	模型权重或模型目录路径
`--audio_path`	见脚本默认值	脚本参数，详见 python inference.py --help
`--sample_rate`	见脚本默认值	脚本参数，详见 python inference.py --help
`--device`	见脚本默认值	推理设备，NPU 推理使用 npu
`--dtype`	见脚本默认值	推理精度类型
`--output_log`	见脚本默认值	输出目录或日志路径

手动调用示例

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

7. 自验证文本证据

以下内容来自仓库已有 README 证据段、运行日志或结果文件。图片文件如保留在 assets/ 中，仅作为附件材料；README 中直接写入可检索的文本证据。

渲染截图证据

以下 PNG 文件由之前的 assets/*.txt 证据文件渲染生成。渲染完成后，原始 TXT 文件已被移除。

证据	PNG 文件
accuracy_eval_result	`assets/accuracy_eval_result.png`
env_check	`assets/env_check.png`
git_submit_result	`assets/git_submit_result.png`
inference_result	`assets/inference_result.png`
performance_eval_result	`assets/performance_eval_result.png`

9. 结果数据直接文本

本节将仓库中已提交的评测 JSON、推理日志、环境日志和性能日志直接写入 README。原始文件路径仅用于标识数据来源，主要数值和输出内容已在下面以文本形式完整展开。

logs/env_check.log

文件大小：2697 bytes
以下内容为 README 直接文本转写，不是外部路径引用。

# Environment Check Log
# Repository: hubert-base-deepfake-npu
# Model: hubert-base-960h-itw-deepfake
# Date: 2026-05-16 07:03:22

## System Info
Linux pod-8e032c81b34d489191e775768926f3b6 5.10.0-182.0.0.95.r2220_156.hce2.aarch64 #1 SMP Sat Sep 14 02:34:54 UTC 2024 aarch64 aarch64 aarch64 GNU/Linux

## Python
Python 3.11.14
pip 26.0.1 from /usr/local/python3.11.14/lib/python3.11/site-packages/pip (python 3.11)

## NPU Info
+------------------------------------------------------------------------------------------------+
| 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)        |
+===========================+===============+====================================================+
| 0     Ascend910           | OK            | 174.3       48                0    / 0             |
| 0     0                   | 0000:0A:00.0  | 0           0    / 0          3106 / 65536         |
+------------------------------------------------------------------------------------------------+
| 0     Ascend910           | OK            | -           48                0    / 0             |
| 1     1                   | 0000:0B:00.0  | 0           0    / 0          2870 / 65536         |
+===========================+===============+====================================================+
+---------------------------+---------------+----------------------------------------------------+
| NPU     Chip              | Process id    | Process name             | Process memory(MB)      |
+===========================+===============+====================================================+
| No running processes found in NPU 0                                                            |
+===========================+===============+====================================================+

## CANN Version
8.5.1

## PyTorch
2.9.0+cpu

## torch_npu
2.9.0.post1+gitee7ba04

## transformers
4.57.6

## Git Info
Branch: main
Commit: 82a40ea9dd44047862972bc3d78eeed06fb3ba4a

<redacted sensitive line>
ASCEND_TOOLKIT_HOME=/usr/local/Ascend/cann-8.5.1
PYTHONPATH=/usr/local/Ascend/cann-8.5.1/python/site-packages:/usr/local/Ascend/cann-8.5.1/opp/built-in/op_impl/ai_core/tbe:/usr/local/Ascend/ascend-toolkit/latest/python/site-packages:/usr/local/Ascend/ascend-toolkit/latest/opp/built-in/op_impl/ai_core/tbe:

results/env_info.json

文件大小：352 bytes
以下内容为 README 直接文本转写，不是外部路径引用。

{
  "os": "Linux 5.10.0 aarch64",
  "python_version": "3.11.14",
  "architecture": "aarch64",
  "npu_model": "Ascend910_9362",
  "npu_count": 2,
  "npu_smi_version": "25.5.2",
  "cann_version": "8.5.1",
  "torch_version": "2.9.0+cpu",
  "torch_npu_version": "2.9.0.post1",
  "transformers_version": "4.57.6",
  "numpy_version": "1.26.4"
}

logs/inference.log

文件大小：389 bytes
以下内容为 README 直接文本转写，不是外部路径引用。

{"model": "hubert-base-960h-itw-deepfake", "model_path": "./weights", "audio_path": "./test_audio_5s.wav", "audio_duration_s": 5.0, "sample_rate": 16000, "device": "npu:0", "dtype": "torch.float32", "prediction": "bona-fide", "prediction_id": 0, "logits": [3.186036, -3.157881], "probabilities": [0.998246, 0.001754], "inference_time_s": 0.2457, "model_load_time_s": 1.41, "rtf": 0.0491}

logs/accuracy_eval.log

文件大小：972 bytes
以下内容为 README 直接文本转写，不是外部路径引用。

{"model": "hubert-base-960h-itw-deepfake", "audio_path": "./test_audio_5s.wav", "audio_duration_s": 5.0, "dtype": "float32", "evaluations": [{"comparison": "NPU vs CPU", "logits": {"cpu_logits": [3.1859610080718994, -3.1582372188568115], "npu_logits": [3.1860363483428955, -3.157881021499634], "max_relative_error": 5.639497976517305e-05, "mean_relative_error": 3.410931458347477e-05, "filtered_mean_relative_error": 3.410931458347477e-05, "normalized_mae": 4.80996495753061e-05, "cosine_similarity": 1.0000001192092896}, "prediction": {"cpu_prediction": "bona-fide", "npu_prediction": "bona-fide", "match": true}, "pass_criteria": {"filtered_mre_lt_1pct": true, "cosine_similarity_gt_09999": true, "prediction_match": true, "overall_pass": true}, "hidden_states": {"max_relative_error": 1.0, "mean_relative_error": 0.01068372093141079, "filtered_mean_relative_error": 0.0009429390775039792, "cosine_similarity": 0.9999975562095642}}], "summary": {"overall_pass": true}}

results/accuracy_eval.json

文件大小：1317 bytes
以下内容为 README 直接文本转写，不是外部路径引用。

{
  "model": "hubert-base-960h-itw-deepfake",
  "audio_path": "./test_audio_5s.wav",
  "audio_duration_s": 5.0,
  "dtype": "float32",
  "evaluations": [
    {
      "comparison": "NPU vs CPU",
      "logits": {
        "cpu_logits": [
          3.1859610080718994,
          -3.1582372188568115
        ],
        "npu_logits": [
          3.1860363483428955,
          -3.157881021499634
        ],
        "max_relative_error": 5.639497976517305e-05,
        "mean_relative_error": 3.410931458347477e-05,
        "filtered_mean_relative_error": 3.410931458347477e-05,
        "normalized_mae": 4.80996495753061e-05,
        "cosine_similarity": 1.0000001192092896
      },
      "prediction": {
        "cpu_prediction": "bona-fide",
        "npu_prediction": "bona-fide",
        "match": true
      },
      "pass_criteria": {
        "filtered_mre_lt_1pct": true,
        "cosine_similarity_gt_09999": true,
        "prediction_match": true,
        "overall_pass": true
      },
      "hidden_states": {
        "max_relative_error": 1.0,
        "mean_relative_error": 0.01068372093141079,
        "filtered_mean_relative_error": 0.0009429390775039792,
        "cosine_similarity": 0.9999975562095642
      }
    }
  ],
  "summary": {
    "overall_pass": true
  }
}

logs/performance_eval.log

文件大小：1587 bytes
以下内容为 README 直接文本转写，不是外部路径引用。

{"model": "hubert-base-960h-itw-deepfake", "device": "npu:0", "dtype": "float32", "warmup_runs": 3, "num_runs": 10, "benchmarks": [{"audio_duration_s": 1.0, "avg_latency_s": 0.0067, "std_latency_s": 0.0, "min_latency_s": 0.0066, "max_latency_s": 0.0067, "p50_latency_s": 0.0067, "p95_latency_s": 0.0067, "rtf": 0.0067, "throughput_inferences_per_s": 150.06, "npu_memory": {"allocated_mb": 361.17, "reserved_mb": 480.0}}, {"audio_duration_s": 3.0, "avg_latency_s": 0.0064, "std_latency_s": 0.0001, "min_latency_s": 0.0063, "max_latency_s": 0.0067, "p50_latency_s": 0.0063, "p95_latency_s": 0.0067, "rtf": 0.0021, "throughput_inferences_per_s": 155.69, "npu_memory": {"allocated_mb": 361.29, "reserved_mb": 516.0}}, {"audio_duration_s": 5.0, "avg_latency_s": 0.0064, "std_latency_s": 0.0, "min_latency_s": 0.0063, "max_latency_s": 0.0064, "p50_latency_s": 0.0064, "p95_latency_s": 0.0064, "rtf": 0.0013, "throughput_inferences_per_s": 156.88, "npu_memory": {"allocated_mb": 361.41, "reserved_mb": 570.0}}, {"audio_duration_s": 10.0, "avg_latency_s": 0.0068, "std_latency_s": 0.0, "min_latency_s": 0.0067, "max_latency_s": 0.0068, "p50_latency_s": 0.0068, "p95_latency_s": 0.0068, "rtf": 0.0007, "throughput_inferences_per_s": 147.96, "npu_memory": {"allocated_mb": 361.71, "reserved_mb": 800.0}}, {"audio_duration_s": 30.0, "avg_latency_s": 0.0162, "std_latency_s": 0.0, "min_latency_s": 0.0162, "max_latency_s": 0.0163, "p50_latency_s": 0.0162, "p95_latency_s": 0.0163, "rtf": 0.0005, "throughput_inferences_per_s": 61.71, "npu_memory": {"allocated_mb": 363.1, "reserved_mb": 1478.0}}]}

results/performance_eval.json

文件大小：2156 bytes
以下内容为 README 直接文本转写，不是外部路径引用。

{
  "model": "hubert-base-960h-itw-deepfake",
  "device": "npu:0",
  "dtype": "float32",
  "warmup_runs": 3,
  "num_runs": 10,
  "benchmarks": [
    {
      "audio_duration_s": 1.0,
      "avg_latency_s": 0.0067,
      "std_latency_s": 0.0,
      "min_latency_s": 0.0066,
      "max_latency_s": 0.0067,
      "p50_latency_s": 0.0067,
      "p95_latency_s": 0.0067,
      "rtf": 0.0067,
      "throughput_inferences_per_s": 150.06,
      "npu_memory": {
        "allocated_mb": 361.17,
        "reserved_mb": 480.0
      }
    },
    {
      "audio_duration_s": 3.0,
      "avg_latency_s": 0.0064,
      "std_latency_s": 0.0001,
      "min_latency_s": 0.0063,
      "max_latency_s": 0.0067,
      "p50_latency_s": 0.0063,
      "p95_latency_s": 0.0067,
      "rtf": 0.0021,
      "throughput_inferences_per_s": 155.69,
      "npu_memory": {
        "allocated_mb": 361.29,
        "reserved_mb": 516.0
      }
    },
    {
      "audio_duration_s": 5.0,
      "avg_latency_s": 0.0064,
      "std_latency_s": 0.0,
      "min_latency_s": 0.0063,
      "max_latency_s": 0.0064,
      "p50_latency_s": 0.0064,
      "p95_latency_s": 0.0064,
      "rtf": 0.0013,
      "throughput_inferences_per_s": 156.88,
      "npu_memory": {
        "allocated_mb": 361.41,
        "reserved_mb": 570.0
      }
    },
    {
      "audio_duration_s": 10.0,
      "avg_latency_s": 0.0068,
      "std_latency_s": 0.0,
      "min_latency_s": 0.0067,
      "max_latency_s": 0.0068,
      "p50_latency_s": 0.0068,
      "p95_latency_s": 0.0068,
      "rtf": 0.0007,
      "throughput_inferences_per_s": 147.96,
      "npu_memory": {
        "allocated_mb": 361.71,
        "reserved_mb": 800.0
      }
    },
    {
      "audio_duration_s": 30.0,
      "avg_latency_s": 0.0162,
      "std_latency_s": 0.0,
      "min_latency_s": 0.0162,
      "max_latency_s": 0.0163,
      "p50_latency_s": 0.0162,
      "p95_latency_s": 0.0163,
      "rtf": 0.0005,
      "throughput_inferences_per_s": 61.71,
      "npu_memory": {
        "allocated_mb": 363.1,
        "reserved_mb": 1478.0
      }
    }
  ]
}

8. 许可证与声明

适配代码许可证以本仓库 license 元数据或 LICENSE 文件为准。
原始模型权重许可证以模型发布方为准。
本仓库不应提交私钥、token、API key、缓存目录或大体积权重文件。
文档中的运行结果来自仓库现有日志和 JSON 结果文件；未验证的数值不会在 README 中虚构。