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/rad-dino |
| 竞赛任务 | Track 1 模型适配 |
| 硬件元数据 | hardware: NPU |
| 必需标签 | #+NPU |
| README 数据策略 | 推理、精度和性能数值以文本形式写入本 README;不使用图片替代数据。 |
| 项目 | 数值 |
|---|---|
| 模型仓库 | https://gitcode.com/nanyizjm/rad-dino |
| 原始模型或权重来源 | https://ai.gitcode.com/hf_mirrors/microsoft/rad-dino |
| 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 参考结果相比,误差小于 1% |
| 必需标签 | #+NPU |
| 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 |
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
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 | rad-dino |
| 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 - 已签入的精度证据报告显示通过;具体记录值请见下表。 |
| 性能证据 | 可用 - 已签入的性能指标如下所示。 |
| 证据文件 | results/inference_result.json、logs/inference.log、results/accuracy_eval.json、results/performance_eval.json、logs/accuracy_eval.log、logs/performance_eval.log |
"device": "npu:0",
"input_shape": [
"pooler_output_shape": [
"throughput": 93.09710785075356,
Device: npu:0 | Dtype: float32 | NPU: True (2)
pooler_output shape: torch.Size([1, 768])
Throughput: 93.10 images/s| 来源 | 指标 | 值 |
|---|---|---|
results/inference_result.json | device | npu:0 |
results/inference_result.json | input_shape | [1,3,518,518] |
results/inference_result.json | pooler_output_shape | [1,768] |
results/inference_result.json | throughput | 93.09710785075356 |
| 来源 | 指标 | 值 |
|---|---|---|
results/accuracy_eval.json | cls_cosine_mean | 0.9999988374282698 |
results/accuracy_eval.json | hidden_cosine_mean | 0.9999995091640074 |
results/accuracy_eval.json | passed | true |
results/accuracy_eval.json | per_sample[0].cls_cosine | 0.9999987800348845 |
results/accuracy_eval.json | per_sample[0].hidden_cosine | 0.9999999999999688 |
results/accuracy_eval.json | per_sample[1].cls_cosine | 0.9999990556630898 |
results/accuracy_eval.json | per_sample[1].hidden_cosine | 0.9999998023004005 |
results/accuracy_eval.json | per_sample[2].cls_cosine | 0.9999985074633068 |
results/accuracy_eval.json | per_sample[2].hidden_cosine | 0.9999999999999688 |
results/accuracy_eval.json | per_sample[3].cls_cosine | 0.9999989697479544 |
精度结论:PASS - 已提交的精度验证报告显示 PASS;确切记录值请参见下表。
| 来源 | 指标 | 值 |
|---|---|---|
results/performance_eval.json | device | npu |
results/performance_eval.json | dtype | float32 |
results/performance_eval.json | batch_size | 1 |
results/performance_eval.json | warmup | 3 |
results/performance_eval.json | num_runs | 10 |
results/performance_eval.json | throughput | 94.57302045776879 |
results/performance_eval.json | peak_memory_mb | 334.2177734375 |
results/performance_eval.json | all_runs_ms | [10.584354400634766,10.561704635620117,10.576725006103516,10.590791702270508,10.581493377685547,10.563373565673828,10.571002960205078,10.558843612670898,10.5729 |
本文档记录 RAD-DINO 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。
RAD-DINO 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:
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。
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。
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。
| 组件 | 版本 / 说明 |
|---|---|
| 操作系统 | Linux-5.10.0-182.0.0.95.r2220_156.hce2.aarch64-aarch64-with-glibc2.35 |
| Python | 3.11.14 |
| NPU 数量 | 2 |
| PyTorch | 2.9.0+cpu |
| torch_npu | 2.9.0.post1+gitee7ba04 |
| transformers | 4.57.6 |
| 依赖安装 | pip install -r requirements.txt |
results/env_info.json 或 logs/env_check.log 为准)torch_npu,请先完成昇腾基础环境配置后再运行真实验证。.
├── .gitignore
├── assets/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
├── logs/accuracy_eval.log
├── logs/env_check.log
├── logs/inference.log
├── logs/model_check.log
├── logs/performance_eval.log
├── readme.md
├── requirements.txt
├── results/accuracy_eval.json
├── results/env_info.json
├── results/inference_result.json
└── results/performance_eval.json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pip install -r requirements.txt
python inference.py --model_path <model_path> --image_path <image.jpg> --device npupython eval/eval_accuracy.py --model_path <model_path> --device npu
python eval/eval_performance.py --model_path <model_path> --device npu| 指标 | 结果 |
|---|---|
| 模型名称 | rad-dino |
| 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 | 45f756a |
测试结果来源:results/performance_eval.json
| 指标 | 结果 |
|---|---|
device | npu |
dtype | float32 |
batch_size | 1 |
num_runs | 10 |
warmup | 3 |
avg_ms | 10.5738 |
throughput | 94.5730 |
结果来源:results/accuracy_eval.json
| 指标 | 结果 |
|---|---|
是否通过 | PASS |
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。
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。
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 |
--trust_remote_code | 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 |
python inference.py --help
python inference.py --model_path <model_path> --image_path <image.jpg> --device npu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。
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 |
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。
[LOG_WARNING] can not create directory, directory: /home/atomgit/ascend/log, possible reason: No such file or directory.path string is NULLpath string is NULLOK{
"os": "Linux-5.10.0-182.0.0.95.r2220_156.hce2.aarch64-aarch64-with-glibc2.35",
"python_version": "3.11.14",
"torch_version": "2.9.0+cpu",
"torch_npu_version": "2.9.0.post1+gitee7ba04",
"transformers_version": "4.57.6",
"npu_available": true,
"npu_count": 2,
"npu_device_name": "Ascend910_9362",
"ascend_toolkit_home": "/usr/local/Ascend/cann-8.5.1",
"soc_version": "ascend910_9391",
"npu_smi": "+------------------------------------------------------------------------------------------------+\n| npu-smi 25.5.2 Version: 25.5.2 |\n+---------------------------+---------------+----------------------------------------------------+\n| NPU Name | Health | Power(W) Temp(C) Hugepages-Usage(page)|\n| Chip Phy-ID | Bus-Id | AICore(%) Memory-Usage(MB) HBM-Usage(MB) |\n+===========================+===============+====================================================+\n| 5 Ascend910 | OK | 173.1 47 0 / 0 |\n| 0 10 | 0000:0B:00.0 | 0 0 / 0 3105 / 65536 |\n+------------------------------------------------------------------------------------------------+\n| 5 Ascend910 | OK | - 47 0 / 0 |\n| 1 11 | 0000:0A:00.0 | 0 0 / 0 2870 / 65536 |\n+===========================+===============+====================================================+\n+---------------------------+---------------+----------------------------------------------------+\n| NPU Chip | Process id | Process name | Process memory(MB) |\n+===========================+===============+====================================================+\n| No running processes found in NPU 5 |\n+===========================+===============+====================================================+\n"
}Model Check Log
================
Model: microsoft/rad-dino
Source: ModelScope
Architecture: Dinov2Model (DINOv2-based)
- hidden_size: 768, num_heads: 12, num_layers: 12
- patch_size: 14, image_size: 518
- Input: 518x518 RGB, mean=[0.5307,0.5307,0.5307], std=[0.2583,0.2583,0.2583]
- Output: last_hidden_state [1,1370,768], pooler_output [1,768]
CPU inference: PASS
NPU inference: PASS============================================================
RAD-DINO Inference - Ascend NPU
============================================================
Device: npu:0 | Dtype: float32 | NPU: True (2)
Model: /tmp/ms_cache/microsoft/rad-dino
Loading model...
Loaded in 1.45s
Image: random test (seed=42), 518x518
Processor: BitImageProcessor
resize: {'shortest_edge': 518}
crop: {'height': 518, 'width': 518}
mean: [0.5307, 0.5307, 0.5307]
std: [0.2583, 0.2583, 0.2583]
Input shape: torch.Size([1, 3, 518, 518])
============================================================
Results
============================================================
last_hidden_state shape: torch.Size([1, 1370, 768])
pooler_output shape: torch.Size([1, 768])
(1 CLS + 1369 patch tokens, hidden_dim=768)
CLS Embedding (first 10): [0.3168306052684784, -0.6419616937637329, -1.5727428197860718, -1.2535806894302368, -0.5395681262016296, -1.0882470607757568, -0.4416094422340393, 0.26741838455200195, -0.6182914972305298, 0.9658592343330383]
mean=0.031334 std=0.650056
min=-2.197603 max=2.005673
norm=18.035797
Patch Tokens shape: (1369, 768)
mean=0.034229 std=0.551296
Performance
Inference time: 10.74 ms
Throughput: 93.10 images/s
Device: Ascend910_9362 (61.3 GB)
PyTorch 2.9.0+cpu
torch_npu 2.9.0.post1+gitee7ba04{
"model": "microsoft/rad-dino",
"device": "npu:0",
"dtype": "float32",
"input_shape": [
1,
3,
518,
518
],
"last_hidden_state_shape": [
1,
1370,
768
],
"pooler_output_shape": [
1,
768
],
"cls_first10": [
0.3168306052684784,
-0.6419616937637329,
-1.5727428197860718,
-1.2535806894302368,
-0.5395681262016296,
-1.0882470607757568,
-0.4416094422340393,
0.26741838455200195,
-0.6182914972305298,
0.9658592343330383
],
"cls_mean": 0.03133395314216614,
"cls_std": 0.650056004524231,
"inference_ms": 10.741472244262695,
"throughput": 93.09710785075356,
"npu_available": true
}============================================================
RAD-DINO Accuracy: CPU vs NPU
============================================================
Loading CPU model...
Loading NPU model...
Testing 5 images (518x518)
Sample 1: CLS cos=0.999999 p99=0.0661% | Hid cos=1.000000 p99=0.0178%
Sample 2: CLS cos=0.999999 p99=0.0615% | Hid cos=1.000000 p99=0.0164%
Sample 3: CLS cos=0.999999 p99=0.0644% | Hid cos=1.000000 p99=0.0182%
Sample 4: CLS cos=0.999999 p99=0.0632% | Hid cos=0.999999 p99=0.0167%
Sample 5: CLS cos=0.999999 p99=0.0633% | Hid cos=0.999999 p99=0.0175%
============================================================
Summary
============================================================
CLS: Cosine=0.999999 P99Err=0.0661% MAE=0.0181%
Hid: Cosine=1.000000 P99Err=0.0182% MAE=0.0050%
Cos>0.9999: PASS
P99Err<1%: PASS
MAE<1%: PASS
Overall: PASS{
"model": "microsoft/rad-dino",
"device": "npu:0",
"dtype": "float32",
"num_samples": 5,
"cls_cosine_mean": 0.9999988374282698,
"hidden_cosine_mean": 0.9999995091640074,
"cls_p99_err_pct": 0.06613038145555634,
"hidden_p99_err_pct": 0.01824191949049823,
"cls_mae_pct": 0.018054616812150925,
"hidden_mae_pct": 0.004988936489098705,
"passed": true,
"per_sample": [
{
"sample": 1,
"cls_cosine": 0.9999987800348845,
"hidden_cosine": 0.9999999999999688,
"cls": {
"max_pct": 0.1272061374038458,
"p99_pct": 0.06613038145555634,
"mae_pct": 0.018424690642859787,
"mae_abs": 0.0007741052540950477
},
"hidden": {
"max_pct": 0.07381153991445899,
"p99_pct": 0.017760223430056992,
"mae_pct": 0.005080216578789987,
"mae_abs": 0.000586233043577522
}
},
{
"sample": 2,
"cls_cosine": 0.9999990556630898,
"hidden_cosine": 0.9999998023004005,
"cls": {
"max_pct": 0.12245668331161141,
"p99_pct": 0.06145578039711207,
"mae_pct": 0.017254924750886858,
"mae_abs": 0.0007324314792640507
},
"hidden": {
"max_pct": 0.05014257039874792,
"p99_pct": 0.016444956865431248,
"mae_pct": 0.004784430348081514,
"mae_abs": 0.000570478558074683
}
},
{
"sample": 3,
"cls_cosine": 0.9999985074633068,
"hidden_cosine": 0.9999999999999688,
"cls": {
"max_pct": 0.1262963400222361,
"p99_pct": 0.06439309744659101,
"mae_pct": 0.01873060828074813,
"mae_abs": 0.0007911864086054265
},
"hidden": {
"max_pct": 0.048601406160742044,
"p99_pct": 0.01824191949049823,
"mae_pct": 0.005227290603215806,
"mae_abs": 0.0005836078198626637
}
},
{
"sample": 4,
"cls_cosine": 0.9999989697479544,
"hidden_cosine": 0.9999989324947218,
"cls": {
"max_pct": 0.1247580861672759,
"p99_pct": 0.06318703062043086,
"mae_pct": 0.01777765282895416,
"mae_abs": 0.0007481399807147682
},
"hidden": {
"max_pct": 0.06481683230958879,
"p99_pct": 0.01674146481396491,
"mae_pct": 0.004826812073588371,
"mae_abs": 0.0005654863780364394
}
},
{
"sample": 5,
"cls_cosine": 0.9999988742321131,
"hidden_cosine": 0.9999988110249767,
"cls": {
"max_pct": 0.12703402899205685,
"p99_pct": 0.06333321044741012,
"mae_pct": 0.018085207557305694,
"mae_abs": 0.0007614145870320499
},
"hidden": {
"max_pct": 0.15527636278420687,
"p99_pct": 0.01749468010081684,
"mae_pct": 0.005025932841817848,
"mae_abs": 0.0005882361438125372
}
}
]
}============================================================
RAD-DINO Performance Evaluation
============================================================
Batch: 1 | Input: torch.Size([1, 3, 518, 518]) | Device: npu
Memory before warmup: 334.2 MB
Warmup: 3
Timed runs: 10
Run 1: 10.58 ms
Run 2: 10.56 ms
Run 3: 10.58 ms
Run 4: 10.59 ms
Run 5: 10.58 ms
Run 6: 10.56 ms
Run 7: 10.57 ms
Run 8: 10.56 ms
Run 9: 10.57 ms
Run 10: 10.58 ms
Device: Ascend910_9362 (61.3 GB)
Peak memory: 334.2 MB
============================================================
Results
============================================================
Avg: 10.57 ms Std: 0.01 ms
Min: 10.56 ms Max: 10.59 ms Median: 10.57 ms
Throughput: 94.57 images/s{
"model": "microsoft/rad-dino",
"device": "npu",
"dtype": "float32",
"batch_size": 1,
"input_shape": [
1,
3,
518,
518
],
"warmup": 3,
"num_runs": 10,
"avg_ms": 10.573840141296387,
"std_ms": 0.009832244607534313,
"min_ms": 10.558843612670898,
"max_ms": 10.590791702270508,
"median_ms": 10.574817657470703,
"throughput": 94.57302045776879,
"peak_memory_mb": 334.2177734375,
"all_runs_ms": [
10.584354400634766,
10.561704635620117,
10.576725006103516,
10.590791702270508,
10.581493377685547,
10.563373565673828,
10.571002960205078,
10.558843612670898,
10.57291030883789,
10.577201843261719
]
}license 元数据或 LICENSE 文件为准。