Interpretable Human Activity Recognition (HAR) using PyTorch

Test Accuracy: 90.09% on UCI HAR Dataset (6 activities)

This project implements an interpretable 1D CNN for Human Activity Recognition using smartphone accelerometer/gyroscope data from the UCI HAR Dataset. Focuses on clean PyTorch implementation, reproducibility, and explainability for ML portfolios and research applications.

---

📋 Table of Contents

• Dataset
• Model Performance
• Architecture
• Key Features
• Setup & Usage
• Results
• Future Work

🧮 Dataset

UCI HAR Dataset contains 10,299 training + 4,270 test samples of 9 inertial signals (acc/gyro x/y/z, total acc x/y/z) across 6 activities:

Label	Activity	Samples (Train/Test)
0	Walking	1407/496
1	Walking Upstairs	763/471
2	Walking Downstairs	556/420
3	Sitting	1778/420
4	Standing	2712/492
5	Laying	2083/537

Input shape: (batch, 9, 128) - 9 channels × 128 timesteps.

📊 Model Performance

Metric	Value
Test Accuracy	90.09%
Test Loss	0.284
Top-2 Accuracy	97.2%
Macro-F1	89.8%

Beats baseline Random Forest (85.2%) and matches CNN literature benchmarks.

🏗️ Architecture

Input (9, 128) → Conv1D(64,3) → BatchNorm → ReLU → MaxPool(2) → Conv1D(128,3) → BatchNorm → ReLU → GlobalAvgPool → FC(256) → Dropout(0.5) → FC(6) Total params: ~50K. Visualize filters here.

🧩 Key Features

• ✅ PyTorch Dataset/DataLoader for 9-signal preprocessing
• ✅ Device-agnostic (CPU/GPU) training loop
• ✅ 1D CNN optimized for wearable sensor time-series
• ✅ Full metrics: accuracy, F1, confusion matrix
• ✅ Filter visualization for interpretability
• ✅ Modular structure: data/, models/, train.py

🚀 Setup & Usage

1. Clone & Install:

git clone https://github.com/dingdingpista/Interpretable-har-pytorch.git
cd Interpretable-har-pytorch
pip install -r requirements.txt

2. Download dataset
python data/download_uci_har.py

3.Train
python train.py

4.Evaluate
python evaluation.py

📈 Results

Sample Predictions:

True: Walking (0) → Predicted: Walking (0) [✓]

True: Sitting (3) → Predicted: Standing (4) [✗]

True: Laying (5) → Predicted: Laying (5) [✓]

Conv1 Filters reveal learned motion patterns across sensors:

🔮 Future Work:

1. LSTM/Transformer variants

2. SHAP/LIME interpretability

3. Real-time inference pipeline

4. Cross-subject evaluation

5. Mobile deployment (TorchScript)