#+NPU
本模型仓库明确声明了所需的 NPU 模型卡片标签。
| 项目 | 数值 |
|---|---|
| 硬件元数据 | hardware: NPU |
| 所需标签 | #+NPU |
| 模型卡片标签 | NPU, Ascend, ascend-npu |
| 竞赛类别 | $category |
| 仓库 | $repo |
本文档记录 $name 在华为昇腾 NPU 环境下的赛道一模型适配、推理验证、精度验证、性能验证与提交材料整理。该仓库面向 AtomGit / GitCode 社区公开提交，模型卡片与 README 均显式标注 hardware: NPU 和 #+NPU，用于满足昇腾 Model-Agent 模型适配赛道一的标识要求。
| 项目 | 内容 |
|---|---|
| 模型 / 仓库 | $repo |
| 任务类型 | 图像识别 / 视觉特征提取 |
| 赛道 | 赛道一：模型适配 |
| 目标硬件 | 昇腾 NPU |
| 提交标签 | #+NPU |
| 精度要求 | 与 CPU / GPU 参考结果误差 < 1% |
| 结果呈现 | README 直接写入文本化证据，截图仅作为辅助材料，不替代数据表与日志摘录 |
| 交付项 | 路径 | 状态 |
|---|---|---|
| 推理脚本 | $(System.Collections.Hashtable.path) | 已提供 |
| 部署文档 | $(System.Collections.Hashtable.path) | 已提供 |
| 精度评测源码 | $(System.Collections.Hashtable.path) | 已提供 |
| 性能评测源码 | $(System.Collections.Hashtable.path) | 已提供 |
| 运行日志目录 | $(System.Collections.Hashtable.path) | 已提供 |
| 结构化结果目录 | $(System.Collections.Hashtable.path) | 已提供 |
| 自验证截图或文本化证据目录 | $(System.Collections.Hashtable.path) | 已提供 |
| 依赖说明 | $(System.Collections.Hashtable.path) | 已提供 |
| 文件 | 状态 | 大小 |
|---|---|---|
| $p | 已提供 | 1088 bytes |
| $p | 已提供 | 3996 bytes |
| $p | 已提供 | 998 bytes |
| $p | 已提供 | 3996 bytes |
| $p | 已提供 | 998 bytes |
说明：本 README 后续章节中的推理输出、精度数据和性能数据均以文本形式展开；如果同时存在 assets/ 截图，截图只用于人工复核，不作为唯一证据。
python inference.py --help
python inference.py --device npu
python eval/eval_accuracy.py --device npu
python eval/eval_performance.py --device npu本部分直接写入 README 以供平台审核。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-L-Ascend |
| 硬件元数据 | 本 README 中包含 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；选定的可复现误差 0.006439671851694584% 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 |
"input_shape": [
"global_feature_shape": [
"patch_feature_shape": [
"avg_latency_s": 0.048869642795762044,
"throughput_images_per_s": 20.462600968442512,
"device": "npu:0",
"latency_per_run_ms": [| 来源 | 指标 | 数值 |
|---|---|---|
logs/inference.log | input_shape | [1,3,768,768] |
logs/inference.log | global_feature_shape | [1,3072] |
logs/inference.log | patch_feature_shape | [1,2304,1024] |
logs/inference.log | avg_latency_s | 0.048869642795762044 |
logs/inference.log | throughput_images_per_s | 20.462600968442512 |
logs/inference.log | device | npu:0 |
logs/inference.log | device_name | Ascend910_9362 |
logs/inference.log | latency_per_run_ms | [48.7495280103758,48.797048977576196,49.10935298539698,48.78758901031688,49.09615300130099,48.76466799760237,49.11362298298627,48.67468698648736,48.921160981990 |
| 来源 | 指标 | 数值 |
|---|---|---|
results/accuracy_eval.json | pass_count | 3 |
results/accuracy_eval.json | overall_passed | true |
results/accuracy_eval.json | threshold | cosine_similarity > 0.999 AND l2_relative_error < 1% |
results/accuracy_eval.json | samples[0].passed | true |
results/accuracy_eval.json | samples[0].global_feature.cosine_similarity | 0.9999738931655884 |
results/accuracy_eval.json | samples[0].global_feature.l2_relative_error | 0.00793202593922615 |
results/accuracy_eval.json | samples[0].global_feature.max_absolute_error | 0.011475682258605957 |
results/accuracy_eval.json | samples[0].global_feature.max_element_relative_error | 12.828276634216309 |
results/accuracy_eval.json | samples[0].global_feature.mean_element_relative_error | 0.034381087869405746 |
results/accuracy_eval.json | samples[0].patch_feature.cosine_similarity | 0.9999776482582092 |
精度结论：PASS - 已提交的精度验证报告显示 PASS；选定的可复现误差 0.006439671851694584% 低于 1%。
| 来源 | 指标 | 数值 |
|---|---|---|
results/performance_eval.json | device | Ascend910_9362 |
results/performance_eval.json | dtype | float32 |
results/performance_eval.json | batch_size | 1 |
results/performance_eval.json | warmup_runs | 5 |
results/performance_eval.json | num_runs | 20 |
results/performance_eval.json | avg_latency_ms | 48.98 |
results/performance_eval.json | min_latency_ms | 48.71 |
results/performance_eval.json | max_latency_ms | 49.82 |
results/performance_eval.json | p50_latency_ms | 48.98 |
results/performance_eval.json | p90_latency_ms | 49.19 |
本文档记录 RADIO-L 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-L 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 |
| NPU 数量 | 2 |
| 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
├── 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
├── assets/test_image.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/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 | RADIO-L |
| 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 | master |
| 当前提交 | 4862473 |
测试结果来源：results/performance_eval.json
| 指标 | 结果 |
|---|---|
device | Ascend910_9362 |
dtype | float32 |
batch_size | 1 |
image_size | 768 |
num_runs | 20 |
avg_latency_ms | 48.9800 |
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。
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 |
--image_size | 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 |
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 |
|---|---|
| 精度评估结果 | assets/accuracy_eval_result.png |
| 环境检查 | assets/env_check.png |
| Git 提交结果 | assets/git_submit_result.png |
| 推理结果 | assets/inference_result.png |
| 性能评估结果 | 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。
+------------------------------------------------------------------------------------------------+
| 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) |
+===========================+===============+====================================================+
| 6 Ascend910 | OK | 167.5 48 0 / 0 |
| 0 12 | 0000:0A:00.0 | 0 0 / 0 3106 / 65536 |
+------------------------------------------------------------------------------------------------+
| 6 Ascend910 | OK | - 47 0 / 0 |
| 1 13 | 0000:0B:00.0 | 0 0 / 0 2870 / 65536 |
+===========================+===============+====================================================+
+---------------------------+---------------+----------------------------------------------------+
| NPU Chip | Process id | Process name | Process memory(MB) |
+===========================+===============+====================================================+
| No running processes found in NPU 6 |
+===========================+===============+====================================================+
{
"model": "RADIO-L (radio_v2.5-l)",
"model_path": "/opt/atomgit/models/RADIO-L",
"image_path": "assets/test_image.png",
"image_size": 768,
"input_shape": [
1,
3,
768,
768
],
"global_feature_shape": [
1,
3072
],
"patch_feature_shape": [
1,
2304,
1024
],
"global_feature_dtype": "torch.float32",
"patch_feature_dtype": "torch.float32",
"global_feature_mean": -0.011490246281027794,
"global_feature_std": 0.25086286664009094,
"patch_feature_mean": -0.0022878902964293957,
"patch_feature_std": 0.42785727977752686,
"avg_latency_s": 0.048869642795762044,
"throughput_images_per_s": 20.462600968442512,
"num_runs": 10,
"device": "npu:0",
"device_name": "Ascend910_9362",
"npu_count": 2,
"dtype": "float32",
"latency_per_run_ms": [
48.7495280103758,
48.797048977576196,
49.10935298539698,
48.78758901031688,
49.09615300130099,
48.76466799760237,
49.11362298298627,
48.67468698648736,
48.9211609819904,
48.68261702358723
]
}{
"os": "Linux-5.10.0-182.0.0.95.r2220_156.hce2.aarch64-aarch64-with-glibc2.35",
"python": "3.11.14",
"hostname": "pod-8e032c81b34d489191e775768926f3b6",
"arch": "aarch64",
"torch": "2.9.0+cpu",
"torch_npu": "2.9.0.post1+gitee7ba04",
"transformers": "5.8.1",
"accelerate": "1.13.0",
"timm": "1.0.27",
"npu_available": true,
"npu_count": 2,
"npu_name": "Ascend910_9362",
"ascend_home": "/usr/local/Ascend/cann-8.5.1"
}=== Model Check ===
Model: RADIO-L (radio_v2.5-l)
Source: https://gitcode.com/hf_mirrors/nvidia/RADIO-L
Weight file: model.safetensors (1.28 GB)
Config: config.json
Preprocessor: preprocessor_config.json
Architecture: ViT-Large (vit_large_patch16_224)
Patch size: 16
Preferred resolution: 768x768
Max resolution: 2048
Version: radio_v2.5-l
Output: summary (B, 3072), features (B, T, 1024)
Loading method: AutoModel.from_pretrained with trust_remote_code=True
Processor: CLIPImageProcessor
=== Files ===
/opt/atomgit/models/RADIO-L/.gitattributes
/opt/atomgit/models/RADIO-L/LICENSE
/opt/atomgit/models/RADIO-L/README.md
/opt/atomgit/models/RADIO-L/adaptor_base.py
/opt/atomgit/models/RADIO-L/adaptor_generic.py
/opt/atomgit/models/RADIO-L/adaptor_mlp.py
/opt/atomgit/models/RADIO-L/adaptor_registry.py
/opt/atomgit/models/RADIO-L/cls_token.py
/opt/atomgit/models/RADIO-L/common.py
/opt/atomgit/models/RADIO-L/config.json
/opt/atomgit/models/RADIO-L/enable_cpe_support.py
/opt/atomgit/models/RADIO-L/enable_spectral_reparam.py
/opt/atomgit/models/RADIO-L/eradio_model.py
/opt/atomgit/models/RADIO-L/extra_timm_models.py
/opt/atomgit/models/RADIO-L/hf_model.py
/opt/atomgit/models/RADIO-L/input_conditioner.py
/opt/atomgit/models/RADIO-L/model.safetensors
/opt/atomgit/models/RADIO-L/open_clip_adaptor.py
/opt/atomgit/models/RADIO-L/preprocessor_config.json
/opt/atomgit/models/RADIO-L/radio_model.py
/opt/atomgit/models/RADIO-L/vit_patch_generator.py
/opt/atomgit/models/RADIO-L/vitdet.py
Model check OK{
"model": "RADIO-L (radio_v2.5-l)",
"model_path": "/opt/atomgit/models/RADIO-L",
"image_path": "assets/test_image.png",
"image_size": 768,
"input_shape": [
1,
3,
768,
768
],
"global_feature_shape": [
1,
3072
],
"patch_feature_shape": [
1,
2304,
1024
],
"global_feature_dtype": "torch.float32",
"patch_feature_dtype": "torch.float32",
"global_feature_mean": -0.011490246281027794,
"global_feature_std": 0.25086286664009094,
"patch_feature_mean": -0.0022878902964293957,
"patch_feature_std": 0.42785727977752686,
"avg_latency_s": 0.048869642795762044,
"throughput_images_per_s": 20.462600968442512,
"num_runs": 10,
"device": "npu:0",
"device_name": "Ascend910_9362",
"npu_count": 2,
"dtype": "float32",
"latency_per_run_ms": [
48.7495280103758,
48.797048977576196,
49.10935298539698,
48.78758901031688,
49.09615300130099,
48.76466799760237,
49.11362298298627,
48.67468698648736,
48.9211609819904,
48.68261702358723
]
}{
"model": "RADIO-L (radio_v2.5-l)",
"model_path": "/opt/atomgit/models/RADIO-L",
"dtype": "float32",
"num_samples": 3,
"pass_count": 3,
"overall_passed": true,
"threshold": "cosine_similarity > 0.999 AND l2_relative_error < 1%",
"samples": [
{
"sample": 1,
"passed": true,
"global_feature": {
"name": "global_feature",
"cosine_similarity": 0.9999738931655884,
"l2_relative_error": 0.00793202593922615,
"max_absolute_error": 0.011475682258605957,
"max_element_relative_error": 12.828276634216309,
"mean_element_relative_error": 0.034381087869405746,
"mse": 3.941072009183699e-06,
"mae": 0.0015377149684354663,
"ref_mean": -0.011432674713432789,
"test_mean": -0.01149024534970522,
"ref_std": 0.2500169575214386,
"test_std": 0.2508220374584198
},
"patch_feature": {
"name": "patch_feature",
"cosine_similarity": 0.9999776482582092,
"l2_relative_error": 0.006439671851694584,
"max_absolute_error": 0.022428035736083984,
"max_element_relative_error": 62.05610275268555,
"mean_element_relative_error": 0.03944610431790352,
"mse": 7.577109499834478e-06,
"mae": 0.002171494998037815,
"ref_mean": -0.002250277902930975,
"test_mean": -0.0022878905292600393,
"ref_std": 0.4274482727050781,
"test_std": 0.4278571903705597
}
},
{
"sample": 2,
"passed": true,
"global_feature": {
"name": "global_feature",
"cosine_similarity": 0.9999738931655884,
"l2_relative_error": 0.00793202593922615,
"max_absolute_error": 0.011475682258605957,
"max_element_relative_error": 12.828276634216309,
"mean_element_relative_error": 0.034381087869405746,
"mse": 3.941072009183699e-06,
"mae": 0.0015377149684354663,
"ref_mean": -0.011432674713432789,
"test_mean": -0.01149024534970522,
"ref_std": 0.2500169575214386,
"test_std": 0.2508220374584198
},
"patch_feature": {
"name": "patch_feature",
"cosine_similarity": 0.9999776482582092,
"l2_relative_error": 0.006439671851694584,
"max_absolute_error": 0.022428035736083984,
"max_element_relative_error": 62.05610275268555,
"mean_element_relative_error": 0.03944610431790352,
"mse": 7.577109499834478e-06,
"mae": 0.002171494998037815,
"ref_mean": -0.002250277902930975,
"test_mean": -0.0022878905292600393,
"ref_std": 0.4274482727050781,
"test_std": 0.4278571903705597
}
},
{
"sample": 3,
"passed": true,
"global_feature": {
"name": "global_feature",
"cosine_similarity": 0.9999738931655884,
"l2_relative_error": 0.00793202593922615,
"max_absolute_error": 0.011475682258605957,
"max_element_relative_error": 12.828276634216309,
"mean_element_relative_error": 0.034381087869405746,
"mse": 3.941072009183699e-06,
"mae": 0.0015377149684354663,
"ref_mean": -0.011432674713432789,
"test_mean": -0.01149024534970522,
"ref_std": 0.2500169575214386,
"test_std": 0.2508220374584198
},
"patch_feature": {
"name": "patch_feature",
"cosine_similarity": 0.9999776482582092,
"l2_relative_error": 0.006439671851694584,
"max_absolute_error": 0.022428035736083984,
"max_element_relative_error": 62.05610275268555,
"mean_element_relative_error": 0.03944610431790352,
"mse": 7.577109499834478e-06,
"mae": 0.002171494998037815,
"ref_mean": -0.002250277902930975,
"test_mean": -0.0022878905292600393,
"ref_std": 0.4274482727050781,
"test_std": 0.4278571903705597
}
}
]
}{
"model": "RADIO-L (radio_v2.5-l)",
"model_path": "/opt/atomgit/models/RADIO-L",
"dtype": "float32",
"num_samples": 3,
"pass_count": 3,
"overall_passed": true,
"threshold": "cosine_similarity > 0.999 AND l2_relative_error < 1%",
"samples": [
{
"sample": 1,
"passed": true,
"global_feature": {
"name": "global_feature",
"cosine_similarity": 0.9999738931655884,
"l2_relative_error": 0.00793202593922615,
"max_absolute_error": 0.011475682258605957,
"max_element_relative_error": 12.828276634216309,
"mean_element_relative_error": 0.034381087869405746,
"mse": 3.941072009183699e-06,
"mae": 0.0015377149684354663,
"ref_mean": -0.011432674713432789,
"test_mean": -0.01149024534970522,
"ref_std": 0.2500169575214386,
"test_std": 0.2508220374584198
},
"patch_feature": {
"name": "patch_feature",
"cosine_similarity": 0.9999776482582092,
"l2_relative_error": 0.006439671851694584,
"max_absolute_error": 0.022428035736083984,
"max_element_relative_error": 62.05610275268555,
"mean_element_relative_error": 0.03944610431790352,
"mse": 7.577109499834478e-06,
"mae": 0.002171494998037815,
"ref_mean": -0.002250277902930975,
"test_mean": -0.0022878905292600393,
"ref_std": 0.4274482727050781,
"test_std": 0.4278571903705597
}
},
{
"sample": 2,
"passed": true,
"global_feature": {
"name": "global_feature",
"cosine_similarity": 0.9999738931655884,
"l2_relative_error": 0.00793202593922615,
"max_absolute_error": 0.011475682258605957,
"max_element_relative_error": 12.828276634216309,
"mean_element_relative_error": 0.034381087869405746,
"mse": 3.941072009183699e-06,
"mae": 0.0015377149684354663,
"ref_mean": -0.011432674713432789,
"test_mean": -0.01149024534970522,
"ref_std": 0.2500169575214386,
"test_std": 0.2508220374584198
},
"patch_feature": {
"name": "patch_feature",
"cosine_similarity": 0.9999776482582092,
"l2_relative_error": 0.006439671851694584,
"max_absolute_error": 0.022428035736083984,
"max_element_relative_error": 62.05610275268555,
"mean_element_relative_error": 0.03944610431790352,
"mse": 7.577109499834478e-06,
"mae": 0.002171494998037815,
"ref_mean": -0.002250277902930975,
"test_mean": -0.0022878905292600393,
"ref_std": 0.4274482727050781,
"test_std": 0.4278571903705597
}
},
{
"sample": 3,
"passed": true,
"global_feature": {
"name": "global_feature",
"cosine_similarity": 0.9999738931655884,
"l2_relative_error": 0.00793202593922615,
"max_absolute_error": 0.011475682258605957,
"max_element_relative_error": 12.828276634216309,
"mean_element_relative_error": 0.034381087869405746,
"mse": 3.941072009183699e-06,
"mae": 0.0015377149684354663,
"ref_mean": -0.011432674713432789,
"test_mean": -0.01149024534970522,
"ref_std": 0.2500169575214386,
"test_std": 0.2508220374584198
},
"patch_feature": {
"name": "patch_feature",
"cosine_similarity": 0.9999776482582092,
"l2_relative_error": 0.006439671851694584,
"max_absolute_error": 0.022428035736083984,
"max_element_relative_error": 62.05610275268555,
"mean_element_relative_error": 0.03944610431790352,
"mse": 7.577109499834478e-06,
"mae": 0.002171494998037815,
"ref_mean": -0.002250277902930975,
"test_mean": -0.0022878905292600393,
"ref_std": 0.4274482727050781,
"test_std": 0.4278571903705597
}
}
]
}{
"model": "RADIO-L (radio_v2.5-l)",
"device": "Ascend910_9362",
"dtype": "float32",
"batch_size": 1,
"image_size": 768,
"input_shape": [
1,
3,
768,
768
],
"output_summary_shape": [
1,
3072
],
"output_features_shape": [
1,
2304,
1024
],
"warmup_runs": 5,
"num_runs": 20,
"avg_latency_ms": 48.98,
"min_latency_ms": 48.71,
"max_latency_ms": 49.82,
"p50_latency_ms": 48.98,
"p90_latency_ms": 49.19,
"p99_latency_ms": 49.82,
"throughput_images_per_s": 20.41,
"memory_before": {
"allocated_mb": 1228.14,
"reserved_mb": 1426.0
},
"memory_after": {
"allocated_mb": 1246.23,
"reserved_mb": 1706.0
},
"latency_per_run_ms": [
48.8,
48.9,
49.19,
48.81,
49.1,
48.83,
49.03,
48.87,
49.1,
48.81,
49.18,
48.81,
49.06,
49.82,
48.98,
48.88,
49.02,
48.79,
48.98,
48.71
]
}{
"model": "RADIO-L (radio_v2.5-l)",
"device": "Ascend910_9362",
"dtype": "float32",
"batch_size": 1,
"image_size": 768,
"input_shape": [
1,
3,
768,
768
],
"output_summary_shape": [
1,
3072
],
"output_features_shape": [
1,
2304,
1024
],
"warmup_runs": 5,
"num_runs": 20,
"avg_latency_ms": 48.98,
"min_latency_ms": 48.71,
"max_latency_ms": 49.82,
"p50_latency_ms": 48.98,
"p90_latency_ms": 49.19,
"p99_latency_ms": 49.82,
"throughput_images_per_s": 20.41,
"memory_before": {
"allocated_mb": 1228.14,
"reserved_mb": 1426.0
},
"memory_after": {
"allocated_mb": 1246.23,
"reserved_mb": 1706.0
},
"latency_per_run_ms": [
48.8,
48.9,
49.19,
48.81,
49.1,
48.83,
49.03,
48.87,
49.1,
48.81,
49.18,
48.81,
49.06,
49.82,
48.98,
48.88,
49.02,
48.79,
48.98,
48.71
]
}license 元数据或 LICENSE 文件为准。