g
gcw_coj3XaOd/Ambiq_compressionkit-ppg-4x
Model Introduction
Files & versions
Pull Requests
Discussion
Analysis

compressionkit-ppg-4x - 昇腾 NPU 推理部署

1. 模型简介

模型名称: Ambiq/compressionkit-ppg-4x-v1.0

模型链接: HuggingFace

模型描述: PPG(光电容积描记)信号压缩编解码器,使用残差向量量化(RVQ)实现 4x 压缩比。专为边缘和可穿戴设备设计,支持在 Ascend NPU 上高效运行。

模型架构: Encoder2D_ds4 + Decoder2D_ds4(Conv2D 编码器 + 上采样 Conv2D 解码器 + RVQ Codebook)

参数规模: ~100K 参数(编码器 + 解码器)

输入规格:

  • 类型:float32 PPG 信号
  • 形状:[1, 1, 320, 1](batch=1, channels=1, width=320, height=1)
  • 采样率:64 Hz
  • 窗口:320 样本 = 5 秒

输出规格:

  • 类型:float32 重建信号
  • 形状:[1, 1, 320, 1]
  • 隐层表示:[1, 1, 80, 16](4x 压缩)

2. 环境依赖

依赖项版本要求说明
Python>= 3.10推荐 3.11
torch>= 2.1.0PyTorch 框架
torch_npu>= 2.1.0昇腾 NPU 后端
tflite-runtime>= 2.14.0TFLite 推理引擎
numpy>= 1.23.0数组计算
昇腾驱动CANN 8.0+推荐 CANN 8.5.1

硬件要求: Ascend 910 / 910B NPU 设备

安装命令:

pip install torch torch-npu tflite-runtime numpy

3. 启动服务

3.1 环境准备

# 检查 NPU 设备
npu-smi info

# 验证 torch_npu
python3 -c "import torch_npu; print(torch.npu.device_count(), torch.npu.get_device_name(0))"

3.2 模型下载

模型已包含在 checkpoint 目录中。手动下载:

# HuggingFace
huggingface-cli download Ambiq/compressionkit-ppg-4x-v1.0 --local-dir ./model_checkpoint

# 或使用 hf-mirror(国内加速)
export HF_ENDPOINT=https://hf-mirror.com
huggingface-cli download Ambiq/compressionkit-ppg-4x-v1.0 --local-dir ./model_checkpoint

3.3 运行推理

# 基础推理(float32 编码器)
python inference.py

# 指定编码器变体
python inference.py --encoder-variant int8
python inference.py --encoder-variant fp16
python inference.py --encoder-variant int16x8

# 指定数据路径
python inference.py --data-path ./path/to/ppg_data.npz

# 限制处理录音数
python inference.py --num-recordings 5

3.4 推理参数说明

参数类型默认值说明
--model-pathstr./model_checkpoint模型文件目录
--data-pathstrdemo_recordings.npzPPG 数据文件
--num-recordingsint10处理的录音数量
--encoder-variantstrfloat32编码器变体 (float32/int8/fp16/int16x8)

4. 推理成功日志

============================================================
  PPG Signal Compression Codec - Inference
  Model: Ambiq/compressionkit-ppg-4x-v1.0
============================================================

[Device]
  NPU available: 2 device(s)
  NPU device: Ascend910_9362

[Model Loading]
  Encoder input: [  1   1 320   1] output: [ 1  1 80 16]
  Decoder input: [ 1  1 80 16] output: [  1   1 320   1]
  Models loaded: OK

[Data Loading]
  Recordings: 10, Samples each: 1920, Sample rate: 64 Hz

============================================================
  Test 1: Single Recording Inference
============================================================

  Quality Metrics:
    MSE:              0.000077
    RMSE:             0.008758
    Cosine Similarity:0.999958
    PRD:              1.77%
    SNR:              35.02 dB

  Timing (6 chunks x 320 samples):
    Encode mean:      0.09 ms
    Decode mean:      0.36 ms
    Per-chunk total:  0.45 ms
    Total:            2.7 ms

============================================================
  Test 2: Batch Inference (10 recordings)
============================================================

  Recording 1: PRD: 1.77%  RMSE: 0.008758  Cosine: 0.999958  Time: 2.6ms
  Recording 2: PRD: 1.86%  RMSE: 0.009070  Cosine: 0.999967  Time: 2.6ms
  ...

  Batch Summary (n=10):
    PRD (%):     mean=1.79  median=1.89  P90=1.98
    RMSE:        mean=0.009767  median=0.008783  P90=0.010391
    Cosine Sim:  mean=0.999959  median=0.999958  P90=0.999965
    Latency(ms): mean=2.6  median=2.5  P90=2.6

5. 推理调用与测试样例

样例 1:单条 PPG 信号编解码

运行命令:

python inference.py --num-recordings 1

输出:

Input signal: 1920 samples (30.0s)
Reconstructed: 1920 samples

Quality Metrics:
  MSE:              0.000077
  RMSE:             0.008758
  Cosine Similarity:0.999958
  PRD:              1.77%
  SNR:              35.02 dB

Timing (6 chunks x 320 samples):
  Encode mean:      0.09 ms
  Decode mean:      0.36 ms
  Per-chunk total:  0.45 ms
  Total:            2.7 ms

样例 2:批量推理(10 条录音)

运行命令:

python inference.py --num-recordings 10

输出:

Recording 1:  PRD: 1.77%  RMSE: 0.008758  Cosine: 0.999958  Time: 2.6ms
Recording 2:  PRD: 1.86%  RMSE: 0.009070  Cosine: 0.999967  Time: 2.6ms
Recording 3:  PRD: 0.88%  RMSE: 0.017499  Cosine: 0.999965  Time: 2.5ms
...

Batch Summary (n=10):
  PRD (%):     mean=1.79  median=1.89  P90=1.98
  Cosine Sim:  mean=0.999959  median=0.999958  P90=0.999965
  Latency(ms): mean=2.6  median=2.5  P90=2.6

6. Agent 适配截图

6.1 Agent 完整适配工作流

Agent 适配流程

6.2 NPU 设备调用日志

NPU 设备调用

6.3 模型适配结果

模型适配结果


7. 精度评测

测试数据: BIDMC PPG and Respiration Dataset v1.0.0(demo_recordings.npz,10 条真实 PPG 录音)

评测指标:

指标结果README 声明值差异
PRD (%)1.79 (mean)1.14 (clean) / 3.44 (faithful)符合预期范围
Cosine Similarity0.999959 (mean)0.9959优于声明值
RMSE0.009767 (mean)0.0183优于声明值
SNR (dB)35.02 (single)—高质量重建
推理延迟 (ms/chunk)0.45—实时性能

评测命令:

python inference.py --num-recordings 10

8. NPU 配置说明

  • NPU 型号: Ascend910_9362
  • NPU 数量: 2 芯片
  • TFLite 后端: CPU + XNNPACK delegate(TFLite 本身使用 CPU 推理)
  • CANN 版本: 8.5.1
  • 推理方式: TFLite float32 推理,单条录音 6 个 chunk,总延迟 ~2.5ms

9. 已知问题

无。