This repo is to replicate results from paper: LSTM-CNN Architecture for Human Activity Recognition

Datasets folder structure

datasets
├── UCI_HAR
│   ├── README.txt
│   ├── activity_labels.txt
│   ├── features.txt
│   ├── features_info.txt
│   ├── segment.py
│   ├── test
│   │   ├── Inertial Signals
│   │   │   ├── body_acc_x_test.txt
│   │   │   ├── body_acc_y_test.txt
│   │   │   ├── body_acc_z_test.txt
│   │   │   ├── body_gyro_x_test.txt
│   │   │   ├── body_gyro_y_test.txt
│   │   │   ├── body_gyro_z_test.txt
│   │   │   ├── total_acc_x_test.txt
│   │   │   ├── total_acc_y_test.txt
│   │   │   └── total_acc_z_test.txt
│   │   ├── X_test.txt
│   │   ├── subject_test.txt
│   │   └── y_test.txt
│   ├── test.pkl
│   ├── test_raw.pkl
│   ├── train
│   │   ├── Inertial Signals
│   │   │   ├── body_acc_x_train.txt
│   │   │   ├── body_acc_y_train.txt
│   │   │   ├── body_acc_z_train.txt
│   │   │   ├── body_gyro_x_train.txt
│   │   │   ├── body_gyro_y_train.txt
│   │   │   ├── body_gyro_z_train.txt
│   │   │   ├── total_acc_x_train.txt
│   │   │   ├── total_acc_y_train.txt
│   │   │   └── total_acc_z_train.txt
│   │   ├── X_train.txt
│   │   ├── subject_train.txt
│   │   └── y_train.txt
│   ├── train.pkl
│   └── train_raw.pkl
└── WISDM_ar_v1.1
    ├── WISDM_ar_v1.1_raw.txt
    ├── WISDM_ar_v1.1_raw_about.txt
    ├── WISDM_ar_v1.1_trans_about.txt
    ├── WISDM_ar_v1.1_transformed.arff
    ├── readme.txt
    ├── segment.py
    ├── test.pkl
    └── train.pkl

For WISDM_ar_v1.1 dataset

• generate segments for train/test split: the generated segments are normalized to 0-1.

  cd datasets/WISDM_ar_v1.1
  python segment.py

• train: can achieve 95%-96% top-1 accuracy, similar to paper result.

  python train.py --batch_size 1000 --mode TESTTRAIN --init_lr .01

For UCI_HAR dataset

• generate segments for train/test split: the generated segments are not normalized, if you want normalized signals, uncomment relevant lines in segment.py.

  cd datasets/UCI_HAR
  python segment.py

• train: can achieve 92%-93% top-1 accuracy, worse than paper result.

  python train.py --batch_size 1000 --mode TESTTRAIN --init_lr .01

Gotchas

• batch norm layer is important
• gradients passed to lstm layers are very small, a 2-layer lstm is harder to learn than 1-layer lstm, and indeed 1-layer lstm achieves better result
• Adam optimizer is better than SGD

Tried to just use MLP, it can achieve 85% top-1 accuracy on WISDM_ar_v1.1.