Crate sensorlm 

sensorlm-rs

A complete Rust implementation of the SensorLM sensor-language foundation model (NeurIPS 2025) using the Burn deep-learning framework.

Backend selection

By default this crate uses the NdArray (CPU) backend. Enable the WGPU (GPU) backend by building with the wgpu feature:

cargo build --features wgpu

Architecture overview

                       ┌────────────────────────────────────────┐
  Raw wearable data    │             SensorLM (Two-Tower)       │
  (1440 × 34 × 1)      │                                        │
        │              │  ┌──────────────┐  SigLIP loss         │
        ▼              │  │ SensorEncoder│──────────────────┐   │
  Normalise / mask      │  │  (ViT-B/10/2)│                 │   │
        │              │  │  MAP pooling │                  │   │
        │              │  └──────────────┘  contrastive     │   │
        │              │         768-dim    alignment        │   │
        │              │                                     ▼   │
  Caption pipeline ──▶ │  ┌──────────────┐  ┌─────────────────┐│
  (stat/struct/semantic)│  │  TextEncoder │  │  temperature    ││
                        │  │  (Transformer│  │  + bias scalars ││
                        │  │   12 layers) │  └─────────────────┘│
                        │  └──────────────┘                      │
                        └────────────────────────────────────────┘

Modules

config

Hierarchical configuration structs for every subsystem.

constants

Physical constants, feature lists, and normalisation parameters.

data

Data pipeline: loading, preprocessing, download, and caption generation.

error

Crate-wide error types.

inference

Inference utilities.

loss

SigLIP sigmoid contrastive loss.

model

Model architecture for SensorLM.

quantization

Model quantisation.

training

Training infrastructure.

Type Aliases

CpuBackend

NdArray (CPU) backend for inference and testing.

CpuTrainBackend

Autodiff wrapper over NdArray for CPU training / unit tests.