nanyizjm/radio-b-npu
模型介绍文件和版本Pull Requests讨论分析
下载使用量0

NPU 标签证明

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/radio-b-npu
竞赛任务Track 1 模型适配
硬件元数据hardware: NPU
必需标签#+NPU
README 数据政策推理、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/radio-b-npu
原始模型或权重来源https://gitcode.com/hf_mirrors/nvidia/RADIO-B
竞赛赛道Track 1:模型适配
目标硬件昇腾 NPU
必需功能NPU 推理成功运行或明确记录阻塞原因
必需精度NPU 结果与 CPU/GPU 参考值相比,误差小于 1%
必需标签#+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 或截图证明已提供

精度证明要求

README 必须包含明确的 CPU/GPU 与 NPU 数值对比数据。关键验收目标为误差小于 1%。相应的结构化证明在可用时应保存至 results/accuracy_eval.json 和 logs/accuracy_eval.log。

#+NPU

radio-b 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。不依赖嵌入图像。

审核项直接结果
仓库radio-b-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 0.0017650524387136102% 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

2026-05-15 01:34:58,160 [INFO] Device: npu
2026-05-15 01:34:58,163 [INFO] Throughput: 58.48 img/s
"throughput_img_s": 58.477411943090374

NPU 推理指标

项目数值
证据在已检入的文本文件中未检测到

CPU/GPU 参考指标与 NPU 精度证据对比

来源指标数值
results/accuracy_eval.jsonsummary.relative_error.max_relative_error1
results/accuracy_eval.jsonsummary.relative_error.mean_relative_error0.014222432859241962
results/accuracy_eval.jsonsummary.relative_error.filtered_mean_relative_error0.0027351537719368935
results/accuracy_eval.jsonsummary.relative_error.max_absolute_error0.010772705078125
results/accuracy_eval.jsonsummary.relative_error.mean_absolute_error0.0012489656219258904
results/accuracy_eval.jsonsummary.cosine_similarity0.9999803332312517
results/accuracy_eval.jsonsummary.mre_passtrue
results/accuracy_eval.jsonsummary.cosine_passtrue
results/accuracy_eval.jsonpatch_features.relative_error.max_relative_error1
results/accuracy_eval.jsonpatch_features.relative_error.mean_relative_error0.013926004990935326

精度结论:通过 - 选定的可复现误差 0.0017650524387136102% 低于 1%。

性能证据

来源指标数值
results/performance_eval.jsondevicenpu
results/performance_eval.jsondtypefloat32
results/performance_eval.jsonwarmup3
results/performance_eval.jsonnum_runs10
results/performance_eval.jsonresults[0].batch_size1
results/performance_eval.jsonresults[0].mean_time_s0.016738491295836867
results/performance_eval.jsonresults[0].std_time_s0.000030622754096756846
results/performance_eval.jsonresults[0].p50_s0.016726584479329176
results/performance_eval.jsonresults[0].p95_s0.016793474441510626
results/performance_eval.jsonresults[0].throughput_img_s59.742540849468085

RADIO-B on Ascend NPU

1. 简介

本文档记录 RADIO-B 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。

RADIO-B 的当前适配任务类型为:图像识别 / 视觉特征提取。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。

相关获取地址:

  • 相关地址:https://gitcode.com/hf_mirrors/nvidia/RADIO-B
  • 相关地址:https://atomgit.com/nanyizjm/radio-b-npu.git
  • 相关地址:https://gitcode.com/nanyizjm/radio-b-npu
  • 适配代码仓库:https://gitcode.com/nanyizjm/radio-b-npu

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 5.10.0-182.0.0.95.r2220_156.hce2.aarch64 aarch64
Python3.11.14
NPU 型号Ascend910_9362
NPU 数量2
CANN/usr/local/Ascend/cann-8.5.1
PyTorch2.9.0+cpu
torch_npu2.9.0.post1+gitee7ba04
transformers4.57.6
timm1.0.27
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> --image_path <image.jpg> --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RADIO-B (radio_v2.5-b)
任务类型图像识别 / 视觉特征提取
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dfe34e0

5.2 推理性能

测试结果来源:results/performance_eval.json

指标结果
devicenpu
dtypefloat32
image_size768
num_runs10
warmup3

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
--image_path见脚本默认值输入样例路径
--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
--image_size见脚本默认值脚本参数,详见 python inference.py --help
--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> --image_path <image.jpg> --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_resultassets/accuracy_eval_result.png
env_checkassets/env_check.png
git_submit_resultassets/git_submit_result.png
inference_resultassets/inference_result.png
performance_eval_resultassets/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:2677 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: radio-b-npu
# Model: RADIO-B (radio_v2.5-b)
# 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            | 175.6       48                0    / 0             |
| 0     0                   | 0000:0A:00.0  | 0           0    / 0          3107 / 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: 3acd2b22dd8b13f543359c3768cd554d432a8f07

<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:415 bytes
  • 以下内容为 README 直接文本转写,不是外部路径引用。
{
  "os": "Linux 5.10.0-182.0.0.95.r2220_156.hce2.aarch64 aarch64",
  "python_version": "3.11.14",
  "architecture": "aarch64",
  "npu_model": "Ascend910_9362",
  "npu_count": 2,
  "cann_version": "/usr/local/Ascend/cann-8.5.1",
  "torch_version": "2.9.0+cpu",
  "torch_npu_version": "2.9.0.post1+gitee7ba04",
  "transformers_version": "4.57.6",
  "timm_version": "1.0.27",
  "numpy_version": "1.26.4"
}

logs/inference.log

  • 文件大小:1454 bytes
  • 以下内容为 README 直接文本转写,不是外部路径引用。
2026-05-15 01:34:56,252 [INFO] Loading RADIO-B from ./weights ...
2026-05-15 01:34:57,928 [INFO] Model loaded on npu, dtype=torch.float32
2026-05-15 01:34:57,929 [INFO] Parameters: 98,233,344 (98.2M)
2026-05-15 01:34:57,929 [INFO] Loading image: test_image_radio.png
2026-05-15 01:34:58,031 [INFO] Input shape: torch.Size([1, 3, 768, 768])
2026-05-15 01:34:58,031 [INFO] Warming up ...
2026-05-15 01:34:58,114 [INFO] Running inference ...
2026-05-15 01:34:58,160 [INFO] === RADIO-B Inference Results ===
2026-05-15 01:34:58,160 [INFO] Device: npu
2026-05-15 01:34:58,161 [INFO] Dtype: float32
2026-05-15 01:34:58,161 [INFO] Image size: 768x768
2026-05-15 01:34:58,161 [INFO] Summary shape: (1, 2304)
2026-05-15 01:34:58,161 [INFO] Patch features shape: (1, 2304, 768)
2026-05-15 01:34:58,161 [INFO] Summary L2 norm: 12.1684
2026-05-15 01:34:58,163 [INFO] Patch features mean L2 norm: 9.0184
2026-05-15 01:34:58,163 [INFO] Inference time: 0.0171s
2026-05-15 01:34:58,163 [INFO] Throughput: 58.48 img/s
2026-05-15 01:34:58,164 [INFO] Results: {
  "device": "npu",
  "dtype": "float32",
  "image_size": 768,
  "image_path": "test_image_radio.png",
  "summary_shape": [
    1,
    2304
  ],
  "patch_features_shape": [
    1,
    2304,
    768
  ],
  "summary_norm": 12.168350219726562,
  "patch_features_mean_norm": 9.01836109161377,
  "inference_time_s": 0.017100619996199384,
  "throughput_img_s": 58.477411943090374
}

logs/accuracy_eval.log

  • 文件大小:1133 bytes
  • 以下内容为 README 直接文本转写,不是外部路径引用。
2026-05-15 01:36:02,718 [INFO] Loading RADIO-B model ...
2026-05-15 01:36:03,090 [INFO] Loading image: test_image_radio.png
2026-05-15 01:36:03,108 [INFO] Running CPU inference ...
2026-05-15 01:36:15,787 [INFO] CPU summary shape: (1, 2304), features shape: (1, 2304, 768)
2026-05-15 01:36:15,787 [INFO] Running NPU inference ...
2026-05-15 01:36:17,261 [INFO] NPU summary shape: (1, 2304), features shape: (1, 2304, 768)
2026-05-15 01:36:17,261 [INFO] === Accuracy Comparison (NPU vs CPU) ===
2026-05-15 01:36:17,261 [INFO] Summary Filtered MRE: 0.2735%
2026-05-15 01:36:17,261 [INFO] Summary Cosine Similarity: 0.999980
2026-05-15 01:36:17,284 [INFO] Patch Features Filtered MRE: 0.1765%
2026-05-15 01:36:17,284 [INFO] Patch Features Cosine Similarity: 0.999980
2026-05-15 01:36:17,284 [INFO] Summary MRE < 1%: PASS
2026-05-15 01:36:17,284 [INFO] Summary Cosine > 0.9999: PASS
2026-05-15 01:36:17,284 [INFO] Features MRE < 1%: PASS
2026-05-15 01:36:17,284 [INFO] Features Cosine > 0.9999: PASS
2026-05-15 01:36:17,284 [INFO] Overall: PASS
2026-05-15 01:36:17,284 [INFO] Results saved to results/accuracy_eval.json

results/accuracy_eval.json

  • 文件大小:829 bytes
  • 以下内容为 README 直接文本转写,不是外部路径引用。
{
  "summary": {
    "relative_error": {
      "max_relative_error": 1.0,
      "mean_relative_error": 0.014222432859241962,
      "filtered_mean_relative_error": 0.0027351537719368935,
      "max_absolute_error": 0.010772705078125,
      "mean_absolute_error": 0.0012489656219258904
    },
    "cosine_similarity": 0.9999803332312517,
    "mre_pass": true,
    "cosine_pass": true
  },
  "patch_features": {
    "relative_error": {
      "max_relative_error": 1.0,
      "mean_relative_error": 0.013926004990935326,
      "filtered_mean_relative_error": 0.0017650524387136102,
      "max_absolute_error": 0.025739192962646484,
      "mean_absolute_error": 0.0015869825147092342
    },
    "cosine_similarity": 0.9999796290474484,
    "mre_pass": true,
    "cosine_pass": true
  },
  "overall": "PASS"
}

logs/performance_eval.log

  • 文件大小:1421 bytes
  • 以下内容为 README 直接文本转写,不是外部路径引用。
2026-05-15 01:36:46,295 [INFO] Loading RADIO-B from ./weights ...
2026-05-15 01:36:47,957 [INFO] Model loaded on npu, dtype=torch.float32
2026-05-15 01:36:47,957 [INFO] Image size: 768, Batch sizes: [1, 2, 4]
2026-05-15 01:36:47,957 [INFO] Warmup: 3, Num runs: 10
2026-05-15 01:36:47,957 [INFO]
--- Batch size: 1 ---
2026-05-15 01:36:48,359 [INFO]   Mean: 0.0167s, P50: 0.0167s, P95: 0.0168s
2026-05-15 01:36:48,359 [INFO]   Throughput: 59.74 img/s, NPU Memory: 381.6 MB
2026-05-15 01:36:48,359 [INFO]
--- Batch size: 2 ---
2026-05-15 01:36:48,797 [INFO]   Mean: 0.0325s, P50: 0.0325s, P95: 0.0326s
2026-05-15 01:36:48,797 [INFO]   Throughput: 61.46 img/s, NPU Memory: 388.3 MB
2026-05-15 01:36:48,797 [INFO]
--- Batch size: 4 ---
2026-05-15 01:36:49,664 [INFO]   Mean: 0.0664s, P50: 0.0665s, P95: 0.0665s
2026-05-15 01:36:49,664 [INFO]   Throughput: 60.20 img/s, NPU Memory: 401.8 MB
2026-05-15 01:36:49,664 [INFO]
=== Performance Summary ===
2026-05-15 01:36:49,664 [INFO]  Batch    Mean(s)     P50(s)     P95(s)      img/s    Mem(MB)
2026-05-15 01:36:49,664 [INFO]      1     0.0167     0.0167     0.0168      59.74      381.6
2026-05-15 01:36:49,664 [INFO]      2     0.0325     0.0325     0.0326      61.46      388.3
2026-05-15 01:36:49,664 [INFO]      4     0.0664     0.0665     0.0665      60.20      401.8
2026-05-15 01:36:49,664 [INFO] Results saved to results/performance_eval.json

results/performance_eval.json

  • 文件大小:1064 bytes
  • 以下内容为 README 直接文本转写,不是外部路径引用。
{
  "device": "npu",
  "dtype": "float32",
  "image_size": 768,
  "warmup": 3,
  "num_runs": 10,
  "results": [
    {
      "batch_size": 1,
      "image_size": 768,
      "mean_time_s": 0.016738491295836867,
      "std_time_s": 3.0622754096756846e-05,
      "p50_s": 0.016726584479329176,
      "p95_s": 0.016793474441510626,
      "throughput_img_s": 59.742540849468085,
      "npu_memory_mb": 381.55419921875
    },
    {
      "batch_size": 2,
      "image_size": 768,
      "mean_time_s": 0.03254341960127931,
      "std_time_s": 4.885195529129647e-05,
      "p50_s": 0.03252591499767732,
      "p95_s": 0.03263325610023458,
      "throughput_img_s": 61.45635660001072,
      "npu_memory_mb": 388.30419921875
    },
    {
      "batch_size": 4,
      "image_size": 768,
      "mean_time_s": 0.0664453989971662,
      "std_time_s": 4.763602290321936e-05,
      "p50_s": 0.06645253699389286,
      "p95_s": 0.06650142749858787,
      "throughput_img_s": 60.19980405521523,
      "npu_memory_mb": 401.80419921875
    }
  ]
}

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。