凤
weixin_44093752/WiFi-DensePose-Aarongid-Ascend-NPU
Model Introduction
Files & versions
Pull Requests
Discussion
Analysis

aarongid/wifi-densepose-pretrained Ascend NPU Adaptation

hardware: NPU device: Ascend910 npu:0 model: aarongid/wifi-densepose-pretrained revision: ae4c938c250124457d0eecf97cf5f1b647cb2962 license: MIT

This delivery runs the official WiFi sensing encoder on Ascend NPU. The official Hugging Face page describes an 8 to 64 to 128 encoder with 9,280 parameters and a 128-dimensional normalized embedding.

Official source verification

Google search and the official Hugging Face page were checked on 2026-08-22. The source page is https://huggingface.co/aarongid/wifi-densepose-pretrained and the official project link is https://github.com/ruvnet/RuView.

The valid official checkpoint used here is csi-embed-v2.safetensors. It was loaded with strict=True. The same revision also exposes model.safetensors, but that file failed safetensors header parsing after download, so it was not silently substituted or reported as loaded.

Run

Install the dependencies in requirements.txt in an Ascend environment, source the CANN environment, then run:

python inference.py --model-dir . --device npu:0 --output validation/npu-run.json

The input is a deterministic batch of four standardized 8-dimensional CSI feature vectors. The output is a finite float32 tensor with shape [4, 128]. This is an NPU smoke test only; no full downstream sensing quality gate was run.

Compatibility and validation

No model weights were changed. The official PyTorch encoder and checkpoint were used as-is. The delivery includes the strict state-dict load, deterministic CPU/NPU smoke comparison, requirements, lock metadata, and evidence.

Evidence

Agent workflow

NPU device call

Model result