nanyizjm/dinov3-vitb16-pretrain-lvd1689m
Model Introduction
Files & versions
Pull Requests
Discussion
Analysis

NPU Tag Evidence

This repository is published as an Ascend NPU model repository. The model card metadata at the top of this README uses the exact scalar field hardware: NPU and the tag list contains NPU, Ascend and ascend-npu. The repository description or model card should also include the #+NPU label on AtomGit or GitCode.

ItemValue
Repositoryhttps://gitcode.com/nanyizjm/dinov3-vitb16-pretrain-lvd1689m
Competition taskTrack 1 model adaptation
Hardware metadatahardware: NPU
Required tag#+NPU
README data policyInference, accuracy and performance values are written as text in this README; images are not used as a replacement for data.

Track 1 Model Card Summary

ItemValue
Model repositoryhttps://gitcode.com/nanyizjm/dinov3-vitb16-pretrain-lvd1689m
Original model or weight sourcehttps://gitcode.com/hf_mirrors/facebook/dinov3-vitb16-pretrain-lvd1689m
Competition trackTrack 1: model adaptation
Target hardwareAscend NPU
Required functionNPU inference runs successfully or the blocking reason is explicitly recorded
Required accuracyNPU result compared with CPU/GPU reference, error less than 1 percent
Required tag#+NPU

Deliverable Checklist

DeliverableStatus
inference.pyPresent
readme.md / README.mdPresent
eval/eval_accuracy.pyPresent
eval/eval_performance.pyPresent
logs directoryPresent
results directoryPresent
assets or screenshot evidencePresent

Accuracy Evidence Requirement

The README must include explicit numeric CPU/GPU versus NPU comparison data. The key acceptance target is error less than 1 percent. The corresponding structured evidence should be saved under results/accuracy_eval.json and logs/accuracy_eval.log when available.

#+NPU

dinov3-vitb16-pretrain-lvd1689m on Ascend NPU

Platform Review Evidence Summary (Direct Text)

This section is written directly in the README for platform review. It uses only checked-in logs and JSON result files from this repository. It does not rely on embedded images.

Review itemDirect result
Repositorydinov3-vitb16-pretrain-lvd1689m
Hardware metadatahardware: NPU and #+NPU are present in this README
Normal NPU inference outputPASS - checked-in NPU inference output is written below.
Accuracy requirementPASS - checked-in accuracy evidence reports PASS; selected reproducible error 0.044928232091479% is below 1%.
Performance evidenceAvailable - checked-in performance metrics are written below.
Evidence filesresults/inference_result.json, logs/inference.log, results/accuracy_eval.json, results/performance_eval.json, logs/accuracy_eval.log, logs/performance_eval.log

Normal NPU Inference Output Evidence

"device": "npu:0",
"input_shape": [
"pooler_output_shape": [
"throughput_images_per_sec": 107.09043558188225,
Device: npu:0
Output last_hidden_state shape: torch.Size([1, 201, 768])
Output pooler_output (CLS embedding) shape: torch.Size([1, 768])
Throughput: 107.09 images/s

NPU Inference Metrics

SourceMetricValue
results/inference_result.jsondevicenpu:0
results/inference_result.jsoninput_shape[1,3,224,224]
results/inference_result.jsonpooler_output_shape[1,768]
results/inference_result.jsonthroughput_images_per_sec107.09043558188225

CPU/GPU Reference vs NPU Accuracy Evidence

SourceMetricValue
results/accuracy_eval.jsoncls_embedding_mean_cosine_similarity0.9999935100423747
results/accuracy_eval.jsoncls_embedding_mean_absolute_error0.0017226734664291144
results/accuracy_eval.jsoncls_embedding_max_abs_error_pct0.3863671328872442
results/accuracy_eval.jsonhidden_states_mean_cosine_similarity0.9999957420834166
results/accuracy_eval.jsonhidden_states_mean_absolute_error0.0010877869557589293
results/accuracy_eval.jsonhidden_states_max_abs_error_pct0.060721178306266665
results/accuracy_eval.jsonpass_criteria.cosine_similarity_gt_09999true
results/accuracy_eval.jsonpass_criteria.max_abs_error_lt_1pcttrue
results/accuracy_eval.jsonpassedtrue
results/accuracy_eval.jsonper_sample_results[0].cls_cosine_similarity0.9999936124485376

Accuracy conclusion: PASS - checked-in accuracy evidence reports PASS; selected reproducible error 0.044928232091479% is below 1%.

Performance Evidence

SourceMetricValue
results/performance_eval.jsondevicenpu
results/performance_eval.jsondtypefloat32
results/performance_eval.jsonbatch_size1
results/performance_eval.jsonwarmup3
results/performance_eval.jsonnum_runs10
results/performance_eval.jsonavg_latency_ms8.514189720153809
results/performance_eval.jsonstd_latency_ms0.18973333298118578
results/performance_eval.jsonmin_latency_ms8.114337921142578
results/performance_eval.jsonmax_latency_ms8.79526138305664
results/performance_eval.jsonmedian_latency_ms8.553028106689453

DINOv3 ViT-B/16 on Ascend NPU

1. 简介

本文档记录 DINOv3 ViT-B/16 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。

DINOv3 ViT-B/16 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/facebook/dinov3-vitb16-pretrain-lvd1689m
  • 相关地址:https://atomgit.com/nanyizjm/dinov3-vitb16-pretrain-lvd1689m.git
  • 相关地址:https://gitcode.com/nanyizjm/dinov3-vitb16-pretrain-lvd1689m
  • 适配代码仓库:https://gitcode.com/nanyizjm/dinov3-vitb16-pretrain-lvd1689m

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-with-glibc2.35
Python3.11.14
NPU 数量2
PyTorch2.9.0+cpu
torch_npu2.9.0.post1+gitee7ba04
transformers4.57.6
accelerate1.13.0
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/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
├── requirements.txt
├── results/accuracy_eval.json
├── results/env_info.json
├── results/inference_result.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facebook/dinov3-vitb16-pretrain-lvd1689m
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3b06e55

5.2 推理性能

测试结果来源:results/performance_eval.json

指标结果
devicenpu
dtypefloat32
batch_size1
image_size224
num_runs10
warmup3
avg_latency_ms8.5142

5.3 NPU vs CPU/GPU 精度对比

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。

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
--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

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。

Rendered Screenshot Evidence

The PNG files below were rendered from the previous assets/*.txt evidence files. The original TXT files were removed after rendering.

EvidencePNG file
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:2551 bytes
  • 以下内容为 README 直接文本转写,不是外部路径引用。
[LOG_WARNING] can not create directory, directory: /home/atomgit/ascend/log, possible reason: No such file or directory.path string is NULLpath string is NULL{
  "os": "Linux-5.10.0-182.0.0.95.r2220_156.hce2.aarch64-aarch64-with-glibc2.35",
  "python_version": "3.11.14",
  "arch": "aarch64",
  "torch_version": "2.9.0+cpu",
  "torch_npu_version": "2.9.0.post1+gitee7ba04",
  "transformers_version": "4.57.6",
  "accelerate_version": "1.13.0",
  "numpy_version": "1.26.4",
  "pillow_version": "12.2.0",
  "npu_available": true,
  "npu_count": 2,
  "npu_device_name": "Ascend910_9362",
  "npu_error": "'torch_npu._C._NPUDeviceProperties' object has no attribute 'total_mem'",
  "cann_path": "/usr/local/Ascend/cann-8.5.1",
  "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            | 186.9       46                0    / 0             |\n| 0     10                  | 0000:0B:00.0  | 0           0    / 0          3158 / 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| 5       0                 | 14910         | python3                  | 110                     |\n+===========================+===============+====================================================+\n"
}

results/env_info.json

  • 文件大小:2391 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-with-glibc2.35",
  "python_version": "3.11.14",
  "arch": "aarch64",
  "torch_version": "2.9.0+cpu",
  "torch_npu_version": "2.9.0.post1+gitee7ba04",
  "transformers_version": "4.57.6",
  "accelerate_version": "1.13.0",
  "numpy_version": "1.26.4",
  "pillow_version": "12.2.0",
  "npu_available": true,
  "npu_count": 2,
  "npu_device_name": "Ascend910_9362",
  "npu_error": "'torch_npu._C._NPUDeviceProperties' object has no attribute 'total_mem'",
  "cann_path": "/usr/local/Ascend/cann-8.5.1",
  "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            | 186.9       46                0    / 0             |\n| 0     10                  | 0000:0B:00.0  | 0           0    / 0          3158 / 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| 5       0                 | 14910         | python3                  | 110                     |\n+===========================+===============+====================================================+\n"
}

logs/model_check.log

  • 文件大小:693 bytes
  • 以下内容为 README 直接文本转写,不是外部路径引用。
Model Check Log
================
Model: facebook/dinov3-vitb16-pretrain-lvd1689m
Source: ModelScope (snapshot_download)
Path: /tmp/ms_cache/facebook/dinov3-vitb16-pretrain-lvd1689m

Files:
- config.json: OK (DINOv3ViTConfig, hidden_size=768, num_layers=12)
- model.safetensors: OK (342MB)
- preprocessor_config.json: OK (224x224, ImageNet normalization)
- README.md: OK
- LICENSE.md: OK

Model Architecture: DINOv3ViTModel
- hidden_size: 768
- num_attention_heads: 12
- num_hidden_layers: 12
- patch_size: 16
- image_size: 224
- num_register_tokens: 4
- model_type: dinov3_vit

CPU inference test: PASS
- Output: last_hidden_state [1, 201, 768], pooler_output [1, 768]

logs/inference.log

  • 文件大小:1532 bytes
  • 以下内容为 README 直接文本转写,不是外部路径引用。
============================================================
DINOv3 ViT-B/16 Inference - Ascend NPU
============================================================
Device: npu:0
Dtype: float32
NPU available: True (count: 2)
Model path: /tmp/ms_cache/facebook/dinov3-vitb16-pretrain-lvd1689m

Loading model...
Model loaded in 4.03s

Loading image: random test image
Image size: (224, 224)
Input shape: torch.Size([1, 3, 224, 224])

Warmup run...

Running inference...

============================================================
Inference Results
============================================================
Output last_hidden_state shape: torch.Size([1, 201, 768])
Output pooler_output (CLS embedding) shape: torch.Size([1, 768])

CLS Embedding (first 10 values): [0.5102764368057251, 0.6638539433479309, 0.47167640924453735, -0.9210608601570129, -0.07699183374643326, -0.5380324125289917, -1.1415667533874512, 0.8834766149520874, 0.36170053482055664, -0.94185471534729]
CLS Embedding stats:
  mean: -0.006348
  std:  0.610877
  min:  -2.102153
  max:  2.061325
  norm: 16.930029

Patch Tokens shape: (196, 768)
Patch Tokens stats:
  mean: 0.001928
  std:  0.452366

============================================================
Performance
============================================================
Inference time: 9.34 ms
Throughput: 107.09 images/s
Batch size: 1

Device: Ascend910_9362
Device memory: 61.3 GB

PyTorch version: 2.9.0+cpu
torch_npu version: 2.9.0.post1+gitee7ba04

results/inference_result.json

  • 文件大小:804 bytes
  • 以下内容为 README 直接文本转写,不是外部路径引用。
{
  "model": "facebook/dinov3-vitb16-pretrain-lvd1689m",
  "device": "npu:0",
  "dtype": "float32",
  "input_shape": [
    1,
    3,
    224,
    224
  ],
  "last_hidden_state_shape": [
    1,
    201,
    768
  ],
  "pooler_output_shape": [
    1,
    768
  ],
  "cls_embedding_first10": [
    0.5102764368057251,
    0.6638539433479309,
    0.47167640924453735,
    -0.9210608601570129,
    -0.07699183374643326,
    -0.5380324125289917,
    -1.1415667533874512,
    0.8834766149520874,
    0.36170053482055664,
    -0.94185471534729
  ],
  "cls_embedding_mean": -0.006347866263240576,
  "cls_embedding_std": 0.6108768582344055,
  "inference_time_ms": 9.337902069091797,
  "throughput_images_per_sec": 107.09043558188225,
  "npu_available": true,
  "npu_count": 2
}

logs/accuracy_eval.log

  • 文件大小:2093 bytes
  • 以下内容为 README 直接文本转写,不是外部路径引用。
============================================================
DINOv3 ViT-B/16 Accuracy Evaluation
CPU Reference vs NPU
============================================================

Loading CPU reference model...
Loading NPU model...

Generating 5 test images...

--- Sample 1/5 ---
  CLS Cosine Sim:     0.99999361
  CLS MAE:            0.00173097
  CLS Max Abs Err %:  0.3456%
  Hidden Cosine Sim:  0.99999520
  Hidden MAE:         0.00108265
  Hidden Max Abs Err%:0.0449%

--- Sample 2/5 ---
  CLS Cosine Sim:     0.99999374
  CLS MAE:            0.00167155
  CLS Max Abs Err %:  0.3371%
  Hidden Cosine Sim:  0.99999628
  Hidden MAE:         0.00106982
  Hidden Max Abs Err%:0.0554%

--- Sample 3/5 ---
  CLS Cosine Sim:     0.99999340
  CLS MAE:            0.00173932
  CLS Max Abs Err %:  0.3643%
  Hidden Cosine Sim:  0.99999582
  Hidden MAE:         0.00108726
  Hidden Max Abs Err%:0.0607%

--- Sample 4/5 ---
  CLS Cosine Sim:     0.99999327
  CLS MAE:            0.00174950
  CLS Max Abs Err %:  0.3864%
  Hidden Cosine Sim:  0.99999592
  Hidden MAE:         0.00111066
  Hidden Max Abs Err%:0.0602%

--- Sample 5/5 ---
  CLS Cosine Sim:     0.99999352
  CLS MAE:            0.00172203
  CLS Max Abs Err %:  0.3576%
  Hidden Cosine Sim:  0.99999548
  Hidden MAE:         0.00108854
  Hidden Max Abs Err%:0.0448%

============================================================
Summary
============================================================
Num samples: 5

CLS Embedding (pooler_output):
  Mean Cosine Similarity:       0.99999351
  Mean Absolute Error (MAE):    0.00172267
  Max Abs Error (% of ref max): 0.3864%

Hidden States (last_hidden_state):
  Mean Cosine Similarity:       0.99999574
  Mean Absolute Error (MAE):    0.00108779
  Max Abs Error (% of ref max): 0.0607%

Pass Criteria (feature extraction standard):
  Cosine Similarity > 0.9999:   PASS (0.999994, 0.999996)
  Max Abs Error < 1% of ref:    PASS (0.3864%, 0.0607%)

Overall Accuracy Test: PASS

Results saved to results/accuracy_eval.json

results/accuracy_eval.json

  • 文件大小:2833 bytes
  • 以下内容为 README 直接文本转写,不是外部路径引用。
{
  "model": "facebook/dinov3-vitb16-pretrain-lvd1689m",
  "device": "npu:0",
  "dtype": "float32",
  "num_samples": 5,
  "cls_embedding_mean_cosine_similarity": 0.9999935100423747,
  "cls_embedding_mean_absolute_error": 0.0017226734664291144,
  "cls_embedding_max_abs_error_pct": 0.3863671328872442,
  "hidden_states_mean_cosine_similarity": 0.9999957420834166,
  "hidden_states_mean_absolute_error": 0.0010877869557589293,
  "hidden_states_max_abs_error_pct": 0.060721178306266665,
  "pass_criteria": {
    "cosine_similarity_gt_09999": true,
    "max_abs_error_lt_1pct": true
  },
  "passed": true,
  "per_sample_results": [
    {
      "sample": 1,
      "cls_cosine_similarity": 0.9999936124485376,
      "hidden_cosine_similarity": 0.9999952001063316,
      "cls_mae": 0.0017309744143858552,
      "cls_max_abs_error": 0.007264852523803711,
      "cls_max_abs_error_pct": 0.34559275954961777,
      "hidden_mae": 0.0010826531797647476,
      "hidden_max_abs_error": 0.007970809936523438,
      "hidden_max_abs_error_pct": 0.044928232091479
    },
    {
      "sample": 2,
      "cls_cosine_similarity": 0.9999937432478961,
      "hidden_cosine_similarity": 0.9999962841768066,
      "cls_mae": 0.0016715462552383542,
      "cls_max_abs_error": 0.007165037095546722,
      "cls_max_abs_error_pct": 0.3371236380189657,
      "hidden_mae": 0.0010698226979002357,
      "hidden_max_abs_error": 0.00982046127319336,
      "hidden_max_abs_error_pct": 0.05538108525797725
    },
    {
      "sample": 3,
      "cls_cosine_similarity": 0.9999934030519578,
      "hidden_cosine_similarity": 0.9999958235962666,
      "cls_mae": 0.0017393153393641114,
      "cls_max_abs_error": 0.007513284683227539,
      "cls_max_abs_error_pct": 0.36434633657336235,
      "hidden_mae": 0.0010872612474486232,
      "hidden_max_abs_error": 0.010775089263916016,
      "hidden_max_abs_error_pct": 0.060721178306266665
    },
    {
      "sample": 4,
      "cls_cosine_similarity": 0.9999932705273146,
      "hidden_cosine_similarity": 0.999995923863986,
      "cls_mae": 0.0017495034262537956,
      "cls_max_abs_error": 0.007689923048019409,
      "cls_max_abs_error_pct": 0.3863671328872442,
      "hidden_mae": 0.0011106599122285843,
      "hidden_max_abs_error": 0.010659217834472656,
      "hidden_max_abs_error_pct": 0.06019209395162761
    },
    {
      "sample": 5,
      "cls_cosine_similarity": 0.9999935209361668,
      "hidden_cosine_similarity": 0.9999954786736922,
      "cls_mae": 0.0017220278969034553,
      "cls_max_abs_error": 0.007261514663696289,
      "cls_max_abs_error_pct": 0.3576264716684818,
      "hidden_mae": 0.0010885377414524555,
      "hidden_max_abs_error": 0.0079498291015625,
      "hidden_max_abs_error_pct": 0.044762989273294806
    }
  ]
}

logs/performance_eval.log

  • 文件大小:1179 bytes
  • 以下内容为 README 直接文本转写,不是外部路径引用。
============================================================
DINOv3 ViT-B/16 Performance Evaluation
============================================================

Loading model to npu...
Batch size: 1
Input shape: torch.Size([1, 3, 224, 224])
Device: npu
Dtype: float32

Memory before warmup: allocated=327.4MB, reserved=382.0MB

Warmup: 3 iterations...
Memory after warmup: allocated=328.0MB, reserved=418.0MB

Timed runs: 10 iterations...
  Run 1: 8.56 ms
  Run 2: 8.55 ms
  Run 3: 8.54 ms
  Run 4: 8.50 ms
  Run 5: 8.80 ms
  Run 6: 8.62 ms
  Run 7: 8.62 ms
  Run 8: 8.61 ms
  Run 9: 8.22 ms
  Run 10: 8.11 ms

============================================================
Performance Results
============================================================
Batch size: 1
Input shape: [1, 3, 224, 224]
Num runs: 10
Avg latency: 8.51 ms
Std latency: 0.19 ms
Min latency: 8.11 ms
Max latency: 8.80 ms
Median latency: 8.55 ms
Throughput: 117.45 images/s
Peak NPU memory: allocated=328.6MB, reserved=418.0MB
Device: Ascend910_9362 (61.3 GB)
PyTorch: 2.9.0+cpu
torch_npu: 2.9.0.post1+gitee7ba04

Results saved to results/performance_eval.json

results/performance_eval.json

  • 文件大小:848 bytes
  • 以下内容为 README 直接文本转写,不是外部路径引用。
{
  "model": "facebook/dinov3-vitb16-pretrain-lvd1689m",
  "device": "npu",
  "dtype": "float32",
  "batch_size": 1,
  "image_size": 224,
  "input_shape": [
    1,
    3,
    224,
    224
  ],
  "warmup": 3,
  "num_runs": 10,
  "avg_latency_ms": 8.514189720153809,
  "std_latency_ms": 0.18973333298118578,
  "min_latency_ms": 8.114337921142578,
  "max_latency_ms": 8.79526138305664,
  "median_latency_ms": 8.553028106689453,
  "throughput_images_per_sec": 117.45098862818564,
  "peak_npu_memory_mb": {
    "allocated_mb": 328.62744140625,
    "reserved_mb": 418.0
  },
  "all_runs_ms": [
    8.55875015258789,
    8.547306060791016,
    8.544921875,
    8.498668670654297,
    8.79526138305664,
    8.6212158203125,
    8.624553680419922,
    8.611917495727539,
    8.224964141845703,
    8.114337921142578
  ]
}

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。