Awesome Efficient HAR

This repository is under active development.

A curated list of resources for efficient, edge, and wearable Human Activity Recognition (HAR).

Focus areas: wearable and smartphone sensor data (IMU, multimodal), compact models for time-series HAR, and on-device deployment with reliable benchmarking.

Legend: [📄 paper] [💻 code] [📦 dataset] [⚡ efficient] [📱 on-device] [🧪 benchmark] [🧠 distillation] [🧩 attention] [🔧 quantization] [🪓 pruning] [🧰 tooling]

This README is generated from data/entries.yaml via scripts/generate_readme.py.

Contents

1. Datasets (wearable/IMU/multimodal)
2. Lightweight architectures for time-series
3. Attention modules for sensor/time-series
4. Knowledge distillation for HAR/time-series
5. Quantization/pruning/compression
6. On-device benchmarking and tooling
7. Reproducible benchmarks/leaderboards
8. Deployment patterns (windowing/streaming/personalization)
9. Efficiency reporting checklist for HAR papers
10. Edge HAR starter packs

Datasets (wearable/IMU/multimodal)

• KU-HAR (2020) - Open HAR dataset for heterogeneous smartphone sensing with 18 daily activities and standardized splits. [📦 dataset] [🧪 benchmark]
• MotionSense (2019) - Motion data from iPhone sensors for activity and user context tasks. [📦 dataset] [💻 code]
• SHL Dataset (Sussex-Huawei Locomotion) (2018) - Large-scale smartphone and wearable locomotion benchmark with official Sussex challenge resources. [📦 dataset] [🧪 benchmark]
• ExtraSensory (2017) - Multimodal in-the-wild behavior dataset with mobile and wearable sensors, mirrored via Kaggle and paper resources. [📦 dataset]
• SisFall (2017) - Fall and movement dataset with elderly participants, released with the SisFall benchmark paper. [📦 dataset]
• UniMiB SHAR (2017) - Smartphone accelerometer ADL and fall dataset introduced with the UniMiB SHAR benchmark paper. [📦 dataset]
• MobiAct (2016) - Smartphone inertial dataset for daily activities and falls from the MobiAct benchmark collection. [📦 dataset]
• RealWorld HAR (2016) - Smartphone and smartwatch placement diversity in real-world conditions. [📦 dataset]
• Heterogeneity HAR (HHAR) (2015) - Device and user heterogeneity benchmark across phones and watches. [📦 dataset] [🧪 benchmark]
• mHealth (2014) - Body-worn sensor dataset for activity monitoring. [📦 dataset]
• DSADS (Daily and Sports Activities) (2013) - Body-worn motion sensors across daily and sport actions. [📦 dataset]
• OPPORTUNITY Activity Recognition (2012) - Multimodal ambient and wearable sensor dataset. [📦 dataset] [🧪 benchmark]
• PAMAP2 Physical Activity Monitoring (2012) - Multi-sensor IMU and heart-rate recordings. [📦 dataset] [🧪 benchmark]
• UCI HAR Dataset (2012) - Smartphone inertial HAR benchmark with subject-wise protocol. [📦 dataset] [🧪 benchmark]
• USC-HAD (2012) - Wearable sensor activities with multiple subjects and repetitions. [📦 dataset]
• WISDM (2011) - Phone/watch accelerometer HAR data with classic activity labels. [📦 dataset]

Lightweight architectures for time-series

• Hydra (2022) - Competing convolutional kernels for fast and accurate time-series classification in low-latency settings. [📄 paper] [⚡ efficient] [🧪 benchmark]
• MultiROCKET (2021) - Extends ROCKET with multiple pooling operators and transformations for faster, accurate time-series classification. [📄 paper] [⚡ efficient] [🧪 benchmark] [💻 code]
• Benchmarking TinyML Systems: Challenges and Direction (2020) - Surveys TinyML benchmarking challenges, including hardware variability, metrics, and reproducibility constraints for edge deployments. [📄 paper] [⚡ efficient] [📱 on-device] [🧪 benchmark]
• MiniROCKET (2020) - Faster deterministic ROCKET variant with strong speed-accuracy tradeoff. [📄 paper] [⚡ efficient] [🧪 benchmark]
• EfficientNet (2019) - Compound scaling principles adaptable to time-series CNNs. [📄 paper] [⚡ efficient]
• InceptionTime (2019) - Competitive time-series architecture often used as HAR baseline. [📄 paper] [🧪 benchmark]
• MobileNetV3 (2019) - Hardware-aware search and lightweight attention for edge latency. [📄 paper] [⚡ efficient] [📱 on-device]
• ROCKET (2019) - Fast random convolution features for time-series classification. [📄 paper] [⚡ efficient] [🧪 benchmark]
• MobileNetV2 (2018) - Inverted residuals and linear bottlenecks for compact sensor models. [📄 paper] [⚡ efficient] [📱 on-device]
• ShuffleNet V2 (2018) - Practical efficiency guidelines for real hardware throughput. [📄 paper] [⚡ efficient] [📱 on-device]
• TCN Sequence Modeling (2018) - Strong and efficient sequence baseline for sensor windows. [📄 paper] [⚡ efficient]
• DeepSense (2017) - Unified deep architecture for mobile sensing signals and tasks. [📄 paper] [📱 on-device]
• MobileNets (2017) - Depthwise separable convolutions widely reused for mobile HAR backbones. [📄 paper] [⚡ efficient] [📱 on-device]
• ShuffleNet (2017) - Pointwise group convolution and channel shuffle for low FLOPs. [📄 paper] [⚡ efficient]
• DeepConvLSTM for HAR (2016) - Early strong wearable HAR model combining temporal conv and recurrence. [📄 paper] [🧪 benchmark]
• SqueezeNet (2016) - Fire modules for very small parameter footprint. [📄 paper] [⚡ efficient]

Attention modules for sensor/time-series

• PatchTST (2022) - Patching strategy for efficient transformer-style time-series representation. [📄 paper] [🧩 attention]
• Informer (2020) - ProbSparse attention for long time-series encoding. [📄 paper] [🧩 attention] [⚡ efficient]
• Linformer (2020) - Linear-complexity attention approximation for longer windows. [📄 paper] [🧩 attention] [⚡ efficient]
• Performer (2020) - FAVOR+ linear attention with kernel feature maps. [📄 paper] [🧩 attention] [⚡ efficient]
• ECA-Net (2019) - Efficient channel attention without heavy dimensionality reduction. [📄 paper] [🧩 attention] [⚡ efficient]
• CBAM (2018) - Lightweight channel and spatial attention plug-in module. [📄 paper] [🧩 attention]
• Attention Is All You Need (2017) - Transformer self-attention foundation for sequence modeling. [📄 paper] [🧩 attention]
• Squeeze-and-Excitation Networks (2017) - Channel reweighting block often adapted to 1D sensor CNNs. [📄 paper] [🧩 attention] [⚡ efficient]

Knowledge distillation for HAR/time-series

• DeiT Distillation through Attention (2020) - Distillation-token strategy transferable to sensor transformers. [📄 paper] [🧠 distillation] [🧩 attention]
• KD-Lib (2020) - Open-source PyTorch distillation framework for rapid experiments. [💻 code] [🧠 distillation] [🧰 tooling]
• MobileBERT (2020) - Distilled and compressed transformer optimized for mobile. [📄 paper] [🧠 distillation] [📱 on-device]
• DistilBERT (2019) - Practical distillation recipe for lighter transformer deployment. [📄 paper] [🧠 distillation] [⚡ efficient]
• TinyBERT (2019) - Multi-stage transformer distillation with layer-wise constraints. [📄 paper] [🧠 distillation] [⚡ efficient]
• Born-Again Neural Networks (2018) - Self-distillation via sequential teacher-student training. [📄 paper] [🧠 distillation]
• Paying More Attention to Attention (2016) - Attention transfer losses for compact students. [📄 paper] [🧠 distillation] [🧩 attention]
• Distilling the Knowledge in a Neural Network (2015) - Foundational teacher-student distillation objective. [📄 paper] [🧠 distillation]
• FitNets (2014) - Hint-based intermediate feature matching. [📄 paper] [🧠 distillation]

Quantization/pruning/compression

• AIMET (AI Model Efficiency Toolkit) (2021) - Compression toolkit with quantization and pruning recipes. [💻 code] [🧰 tooling] [🔧 quantization] [🪓 pruning]
• Movement Pruning (2020) - Structured sparsification for transfer and compression. [📄 paper] [🪓 pruning]
• ONNX Runtime Quantization (2020) - Dynamic and static quantization for portable runtime stacks. [🧰 tooling] [🔧 quantization]
• PyTorch Quantization (2019) - Eager, FX, and PT2 quantization flows for production models. [🧰 tooling] [🔧 quantization]
• TensorFlow Lite Post-Training Integer Quantization (2019) - Full-int8 export path for edge inference. [🧰 tooling] [🔧 quantization] [📱 on-device]
• TensorFlow Model Optimization Toolkit (2019) - Pruning and quantization-aware training utilities. [🧰 tooling] [🔧 quantization] [🪓 pruning]
• The Lottery Ticket Hypothesis (2018) - Sparse subnetworks for compact retraining. [📄 paper] [🪓 pruning]
• Deep Compression (2015) - Classical prune-quantize-Huffman compression pipeline. [📄 paper] [🪓 pruning] [🔧 quantization]
• TensorRT (2014) - Deployment-oriented precision and kernel optimization. [🧰 tooling] [🔧 quantization] [📱 on-device]

On-device benchmarking and tooling

• LiteRT for Microcontrollers (2024) - Official LiteRT microcontroller runtime guide for deploying tiny models on embedded targets. [🧰 tooling] [📱 on-device]
• PyTorch ExecuTorch (2023) - On-device inference runtime and tooling for edge deployment. [🧰 tooling] [📱 on-device]
• Edge Impulse Deployment (2021) - Deployment guides for exporting and running edge ML models across embedded and mobile runtimes. [🧰 tooling] [📱 on-device]
• MLPerf Tiny (2021) - Standardized tiny and edge benchmarking suite including HAR-relevant tasks. [🧪 benchmark] [📱 on-device]
• TensorFlow Lite Model Analyzer (2021) - Inspect model memory and op-level deployment constraints. [🧰 tooling] [📱 on-device]
• TensorFlow Lite Benchmark Tool (2019) - CLI profiling for latency and memory on target hardware. [🧰 tooling] [📱 on-device] [🧪 benchmark]
• TensorFlow Lite for Microcontrollers (tflite-micro) (2019) - Reference embedded inference runtime and kernels for tiny on-device ML deployments. [🧰 tooling] [📱 on-device] [💻 code]
• Apache TVM (2018) - End-to-end model compilation stack with microTVM support for constrained edge targets. [🧰 tooling] [📱 on-device]
• CMSIS-NN (2018) - Optimized neural network kernels for Arm Cortex-M CPUs to accelerate edge inference. [🧰 tooling] [📱 on-device] [💻 code] [⚡ efficient]
• Perfetto (2018) - System-level tracing for CPU scheduling and thermal effects. [🧰 tooling] [📱 on-device] [🧪 benchmark]
• Android Battery Historian (2015) - Power usage inspection to support energy proxy reporting. [🧰 tooling] [📱 on-device]
• Android ADB (2008) - Essential deployment and repeatable benchmark automation entrypoint. [🧰 tooling] [📱 on-device]

Reproducible benchmarks/leaderboards

• Hugging Face Collection - Human Activity Recognition (HAR) (2025) - Curated HAR papers and models reference after Papers with Code task pages became unavailable. [🧪 benchmark]
• sktime Benchmarking (2022) - Reproducible experiment tooling for time-series models. [💻 code] [🧰 tooling] [🧪 benchmark]
• Monash Time Series Forecasting Repository (2021) - Protocol and split hygiene references transferable to sensor tasks. [📦 dataset] [🧪 benchmark]
• UCR Time Series Classification Archive (2018) - Standardized time-series benchmark collection for sanity checks. [📦 dataset] [🧪 benchmark]

Deployment patterns (windowing/streaming/personalization)

Windowing policy: report window length, stride, overlap, and label-assignment rule; avoid hidden overlap leakage between train and test. Streaming inference: prefer causal feature extraction and stateful models for low-latency online HAR. Subject split discipline: always separate users across train/val/test when claiming generalization. Personalization track: report both cold-start (no user fine-tune) and adaptation (few-shot or calibration) metrics. Fallback behavior: define unknown and transition states plus confidence thresholds for real-world deployment. Battery-accuracy tradeoff: jointly report duty cycle, sampling rate, and latency.

Efficiency reporting checklist for HAR papers

Report these metrics together: parameter count, MACs/FLOPs for the stated input window, end-to-end latency on target hardware, peak memory (RAM and model size), and an energy proxy (power draw, battery drain rate, or joules per inference).

Common pitfalls to document and avoid: window leakage from overlap across data splits, subject split mistakes from random splitting instead of subject-wise splitting, and personalization evaluation without a clear adaptation budget or protocol.

Edge HAR starter packs

• Starter pack: Distillation-ready sensor training loop (2026) - Recipe-backed starter path for teacher-student training experiments. [🧰 tooling] [🧠 distillation]
• Starter pack: Reproducible benchmark harness (2026) - Recipe-backed starter path for repeatable Android edge benchmark runs. [🧰 tooling] [🧪 benchmark]
• Starter pack: TFLite Android HAR baseline (2026) - Recipe-backed starter path for TFLite int8 export and mobile deployment setup. [🧰 tooling] [📱 on-device]

MIT license. See https://github.com/sumeyye-agac/awesome-efficient-har/blob/main/LICENSE.