FlashHAR-X: Subject-Generalizable Human Activity Recognition

Trained model (not a Transformers checkpoint). Custom multiscale 1D-convolution + two-block attention model for classifying smartphone inertial signals.

• GitHub source: https://github.com/SURGEOPS99/FlashHAR_X
• Official subject-disjoint UCI HAR test accuracy: 0.9627 (96.27%)
• Official test macro-F1: 0.9631
• Task: 6-class human activity recognition.
• Dataset: UCI Human Activity Recognition Using Smartphones, DOI: 10.24432/C54S4K. The original dataset is CC BY 4.0. No dataset copies are included in this model repository.
• Training: official train subjects (7,352 windows); validation subject groups selected within training to fix epoch counts, then retrained using all official training subjects. Official held-out test: 2,947 windows from distinct subjects. Scores use the standard published split, not random window splits.
• Inputs: one 2.56-second sensor window containing 9 channels × 128 readings plus the original dataset's 561 engineered features, in the same order and normalization as training. Do not feed arbitrary accelerometer streams without replicating UCI preprocessing.
• Architecture: 3/7/15 multiscale convolution stem → positional encoding → two attention blocks → mean/max pooling → engineered-feature MLP fusion → 6-class classifier.
• Attention: trained using a custom, educational Triton FlashAttention-style forward and backward kernel. For portability, inference can use numerically equivalent PyTorch scaled_dot_product_attention with the SAME learned weights, including on CPU. This isn't a claim that the custom kernel runs faster than PyTorch SDPA.

Official test results

Model	Accuracy	Macro-F1
cnn_only	0.9684	0.9692
torch_sdpa	0.9630	0.9635
triton_flash	0.9627	0.9631
RBF-SVM (561 engineered features)	0.9549	0.9540

These results are produced by this project's artifacts/official_test_metrics.csv after completing the notebook, not copied from papers. Other research numbers are only comparable when inputs, split, and metric definitions match. The checkpoint chosen here is the custom-Triton-trained model, not necessarily the highest-scoring baseline.

Files

• model.safetensors: trained neural network parameters (no pickle).
• config.json: architecture details and class ordering.
• normalization.json: means/scales fitted on official train subjects only.
• official_test_metrics.csv: reported benchmark numbers.
• flashhar_model.py: network definition.

Reproduce predictions

Clone the associated GitHub project and install its inference dependencies, then:

python inference_hf.py --repo-id Saspatez/FlashHAR-X --dataset-dir data --index 0 --device cpu

The notebook must first have downloaded/extracted the official UCI HAR dataset under data/. The inference script reads a real official test window. Install torch, numpy, huggingface_hub and safetensors to run inference; running the custom Triton backend for training requires an appropriately configured CUDA GPU.

Important limitations

• Subject-independent testing on 30 volunteers does not demonstrate reliability for every user, device orientation, sensor vendor, mobility pattern, or unconstrained real-world scenario.
• The model uses UCI's precomputed 561 engineered features in addition to raw inertial signals. Real-time deployment requires reproducing those features, not just forwarding nine sensor channels.
• The notebook's GPU experiments should be independently verified; benchmark numbers and any performance claims belong to their recorded hardware and seed.
• This is a college research prototype, not a safety-critical or medical system.

References

Reyes-Ortiz et al., Human Activity Recognition Using Smartphones (2013). FlashAttention: Dao et al. (2022); FlashAttention-2: Dao (2023). Our Triton implementation is an educational reimplementation of tiled exact attention, not a novel attention algorithm.