README.md

---
license: mit
library_name: minerva-ml
tags:
- human-activity-recognition
- self-supervised-learning
- time-series
- sensor-data
- smartphone-har
datasets:
- daghar
metrics:
- accuracy
---

# Benchmarking Encoders and SSL for Smartphone-Based HAR

This repository hosts the checkpoints of the best models of the benchmark study: 
**"Benchmarking Encoders and Self-Supervised Learning for Smartphone-Based Human Activity Recognition"**, accepted for publication in **IEEE Access (2026)**.

## Project Resources
[![Project Website](https://img.shields.io/badge/Project-Website-blue?style=for-the-badge&logo=googlechrome&logoColor=white)](https://www.students.ic.unicamp.br/~ra271582/paper_encoders/webpage.html)
[![Paper](https://img.shields.io/badge/IEEE_Access-Paper-red?style=for-the-badge&logo=ieee&logoColor=white)](https://ieeexplore.ieee.org/xpl/RecentIssue.jsp?punumber=6287639)


## Model Description

This project provides a large-scale evaluation of **6 encoders** combined with **4 Self-Supervised Learning (SSL)** techniques (TF-C, TNC, LFR, and DIET). 

- **Developed by:** Hub of Artificial Intelligence and Cognitive Architectures (H.IAAC), University of Campinas.
- **Model Type:** Time-series Classification (Sensor-based).
- **Architecture:** Supports ResNet-SE-5, CNN-PFF, and others via the `minerva-ml` library.
- **SSL Paradigms:** TF-C, TNC, LFR, and DIET.

## How to Get Started

You can easily load these models using the `minerva-ml` framework. We provide the best ones


### LFR trained on motionsense and finetune refinement on motionsense - achieves 97.5% accuracy

[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/17SwN3no0r7m7v1K0g5Fa25iQUBvcIPm_?usp=sharing)


### TFC trained on motionsense and freeze refinement on UCI - for demonstration of tfc backbones

[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://drive.google.com/file/d/1qsQacs5cZNbOsAFlR5v1SOlQEzjnyyjx/view?usp=sharing)

New models are coming soon


### Prerequisites
```bash
pip install minerva-ml huggingface_hub
```

### Loading a Specific Checkpoint
```bash
from huggingface_hub import hf_hub_download
from minerva.models.nets.base import SimpleSupervisedModel
from minerva.models.nets.time_series.cnns import CNN_PF_Backbone
from minerva.models.ssl.tfc import TFC_Backbone
import torch

# 1. Download weights
checkpoint_path = hf_hub_download(
    repo_id="GustavoLuz-Projects/test_model_HAR",
    filename="best_ms_lfr_ts2vec_ft.ckpt" 
)

```

### Training Data
The models were trained/benchmarked using the DAGHAR datasets, standardized for 6-channel sensor input (Accelerometer and Gyroscope)
we used the [standardized view of the DAGHAR Dataset](https://zenodo.org/records/13987073), as introduced in the following paper:

```latex
Napoli, O., Duarte, D., Alves, P., Soto, D.H.P., de Oliveira, H.E., Rocha, A., Boccato, L. and Borin, E., 2024. 
A benchmark for domain adaptation and generalization in smartphone-based human activity recognition. 
Scientific Data, 11(1), p.1192.
```

### If you use these models or the benchmark results, please cite:
```latex
@article{daluz2026benchmarking,
  title={Benchmarking Encoders and Self-Supervised Learning for Smartphone-Based Human Activity Recognition},
  author={da Luz, Gustavo P. C. P. and Soto, Darlinne H. P. and Napoli, Otávio O. and Rocha, Anderson and Boccato, Levy and Borin, Edson},
  journal={IEEE Access},
  year={2026},
  publisher={IEEE}
}
```