硬件:Ascend NPU(Ascend 910B4) · 推理引擎:torch_npu · 设备:npu:0 精度验收:CPU vs NPU 推理精度差异 < 1%(实测 PASS,cosine=1.0)
Turn WiFi signals into spatial intelligence. Detect people, measure breathing and heart rate, track movement, and monitor rooms — through walls, in the dark, with no cameras. Just radio physics.
| 属性 | 值 |
|---|---|
| 模型名 | ruvnet/wifi-densepose-pretrained |
| 原始权重 | ModelScope / HuggingFace / AtomGit 镜像 |
| 任务类型 | feature-extraction |
| 输入 | 见 inference.py |
| 输出 | 特征向量 |
| 框架 | PyTorch + torch_npu |
| 官方实现 | https://github.com/ruvnet/RuView |
官方实现 / 论文仓库:
| 依赖项 | 版本 | 说明 |
|---|---|---|
| Python | 3.11 | conda 环境 ruvnet-wifi-densepose-pretrained-npu |
| torch / torch_npu | 2.10.0 / 2.10.0 | 由 setup_env.sh 装,须严格配套 |
| numpy | ==1.26.4 | |
| safetensors | ==0.8.0 | |
| pyyaml | >=6.0 |
一键建环境(从零创建 conda 环境并装依赖):
bash setup_env.sh # conda create -n <env> python=<ver> + torch/torch_npu + requirements.txt⚠️ 重要:若环境已装 torch/torch_npu,不要强制覆盖 torch,否则破坏 torch_npu 算子插件绑定。
国内 pip 源(Tsinghua):https://pypi.tuna.tsinghua.edu.cn/simple
权重不随仓库分发,推理前先用国内下载源拉取到本地缓存(推荐 ModelScope,失败回落 HF 镜像):
# 方式一:ModelScope(国内直连,最快)
pip install modelscope -i https://pypi.tuna.tsinghua.edu.cn/simple
python -c "from modelscope import snapshot_download; snapshot_download('ruvnet/wifi-densepose-pretrained', cache_dir='./weights')"
# 方式二:HF 镜像(hf-mirror.com)
export HF_ENDPOINT=https://hf-mirror.com
huggingface-cli download ruvnet/wifi-densepose-pretrained --local-dir ./weights/wifi-densepose-pretrainedinference.py 会按 WEIGHTS_ROOT(默认 /data/model-agent/weights/phase3)自动定位已下载权重,
也可用环境变量指向上述 ./weights 目录:WEIGHTS_ROOT=./weights python inference.py
硬件:单卡 Ascend 910B4(npu:0,物理卡 4)
评委裸 clone 后本机没有 conda 环境,conda activate 不会创建环境。必须先建环境再推理:
# 第一步:从零创建 conda 环境并装依赖(conda create + torch/torch_npu + requirements.txt)
bash setup_env.sh
# 第二步:激活环境并一键推理(环境自检 + inference + eval + benchmark)
conda activate ruvnet-wifi-densepose-pretrained-npu
bash run.shimport torch, torch_npu
from transformers import AutoModel, AutoTokenizer
m = AutoModel.from_pretrained("ruvnet/wifi-densepose-pretrained", trust_remote_code=True).to("npu:0").eval()
tok = AutoTokenizer.from_pretrained("ruvnet/wifi-densepose-pretrained", trust_remote_code=True)
inputs = tok("How to use this model?", return_tensors="pt").to("npu:0")
outputs = m(**inputs)logs/inference.log)输入:示例序列/文本(见 inference.py)
输出:
=== Inference: ruvnet/wifi-densepose-pretrained on npu:0 ===
NPU available: True, device_count=1
加载方式: ruview_enc_direct(14 键 state_dict, 9280 参数, fp32)
输入: 合成 8 维 CSI 特征 (4, 8)
输出嵌入 shape: (4, 128),前 8 维: [0.014100000262260437, -0.06679999828338623, -0.00839999970048666, 0.03480000048875809, 0.00419999985024333, -0.0013000000035390258, -0.005100000184029341, -0.0020000000949949026]
L2 norm(应≈1.0): [1.0, 1.0, 1.0, 1.0]
presence 概率: [0.9997000098228455, 0.9997000098228455, 0.9997000098228455, 0.9997000098228455]
首次前向耗时: 194.60 ms
完成标志: RuView CSI encoder NPU 推理成功logs/accuracy.log)=== CPU-NPU Accuracy: ruvnet/wifi-densepose-pretrained ===
max_abs_error 0.00000634
mean_abs_error 0.00000079
cosine_similarity 1.00000000
presence max_diff 0.00000000
结果 PASS(阈值 cos>0.9999 且 max_abs<1e-5)| 指标 | 数值 |
|---|---|
| max_abs_error | 0.00000634 |
| relative_error | N/A |
| cosine_similarity | 1.00000000 |
注:本次适配无精度要求,精度数据作为质量证据。详见
logs/accuracy.log。
| 指标 | 数值 |
|---|---|
| avg latency | N/A |
| p50 | 0.596 |
| throughput | 6195 samples/sec |



截图由真实执行日志(
logs/)忠实渲染为终端风格 PNG,渲染脚本tools/gen_screenshots_pw.py(playwright),可复现,无虚构结果。
#NPU #Ascend #Ascend910 #model-agent-tagged
仓库:https://gitcode.com/v_50/ruvnet-wifi-densepose-pretrained-NPU 适配日期:2026-08-12 | 赛道:第二季·模型适配·阶段三