API¶

The package is deliberately small and source-oriented. Install it in editable mode before importing:

python -m pip install -e .

Package¶

Module	Main responsibility
`imu_activity_pipeline.config`	Paths, split names, channels, windows, classes, and training/decoder defaults
`signal_file_reader`	Read per-user tab-separated sensor streams
`sensor_data_processing`	Filtering, feature preparation, window construction, and labels
`neural_network_models`	Losses and PyTorch detector/classifier definitions
`train`, `train_parallel`	Sequential and parallel training workflows
`inference`	Model loading, multi-scale prediction, decoding, and segment generation
`evaluate`	Same-class segment matching and metrics
`prediction_writer`	Write output rows to Excel

Version¶

import imu_activity_pipeline

print(imu_activity_pipeline.__version__)

Current package version:

0.1.0

Signals¶

DataReader reads every .txt file in a directory and returns a dictionary keyed by file stem.

from imu_activity_pipeline.signal_file_reader import DataReader

reader = DataReader("data/signals/external_test")
sessions = reader.read_data()

for user_id, frame in sessions.items():
    print(user_id, frame.shape)

Each frame should contain the canonical timestamp and six IMU channels described in Data.

Inference¶

from imu_activity_pipeline.inference import run_inference

segments = run_inference(
    data_dir="data/signals/internal_eval",
    output_file="predictions_internal_eval.xlsx",
)

The returned segment rows and workbook use:

user_id, category, start, end

Output¶

from imu_activity_pipeline.prediction_writer import DataOutput

rows = [
    ["HNU00001", "跑步", 1760000000000, 1760000600000],
]

DataOutput(
    rows,
    output_file="predictions_external_test.xlsx",
).save_predictions()

Classifier¶

import torch

from imu_activity_pipeline.neural_network_models import CombinedModel

model = CombinedModel(
    input_channels=6,
    num_classes=6,
    window_size=300,
)

x = torch.randn(2, 300, 6)
logits = model(x)
print(logits.shape)  # torch.Size([2, 6])

The practical model combines three convolution kernel sizes with a bidirectional LSTM and fused classification head.

Configuration¶

from imu_activity_pipeline import config

print(config.SPLIT_NAMES)
print(config.WINDOW_CONFIGS)
print(config.ACTIVITIES)

Path settings such as HLS_HAR_DATA_ROOT and HLS_HAR_MODEL_DIR are read when config is imported. Set environment variables before launching Python.

Compatibility¶

The scripts at repository root preserve source-checkout commands while delegating to the package:

run_inference.py
train.py
train_parallel.py
train_single_model.py
evaluate.py

For reproducible automation, prefer python -m imu_activity_pipeline.<module> when a package module exposes a CLI, and record the package/versioned commit.