DailyAndSportsActivities

Fig. 8 Five original windows from training subject 1, segment 1. Only torso acceleration x/y/z is shown (3 of 45 channels), without normalization. Each panel uses its own vertical scale.

Source and evaluation

Official UCI dataset (CC BY 4.0).

This builder defines train as subjects 1–6 (6,840 windows) and test as subjects 7–8 (2,280 windows). UCI does not provide this train/test split. Keep subjects separate when creating validation folds.

Each sample contains series (time-major float32, shape (125, 45)), label (zero-based activity index), and subject_id (original source ID). Sampling rate is 25 Hz. No extra normalization or windowing is performed. Class and channel order are exposed as CLASS_NAMES and CHANNEL_NAMES. The original segment_id is preserved. IDs are metadata, not classifier features.

Usage

import numpy as np
from stable_datasets.timeseries import DailyAndSportsActivities

dataset = DailyAndSportsActivities()
sample = dataset["train"][0]
signal = np.asarray(sample["series"], dtype=np.float32)

The first call downloads the official archive and prepares both cached splits. Later calls reuse the processed cache. Nested sequences can remain Python lists under formatting; convert explicitly with NumPy or torch.as_tensor.

Citation

@misc{daily_and_sports_activities_256,
  author = {Barshan, Billur and Altun, Kerem},
  title = {Daily and Sports Activities},
  year = {2010},
  publisher = {UCI Machine Learning Repository},
  doi = {10.24432/C5C59F},
  url = {https://doi.org/10.24432/C5C59F}
}