SConvLSTM for UCI Human Activity Recognition
This repository contains the training logs and checkpoints for a SConvLSTM model trained on the UCI Human Activity Recognition (HAR) dataset.
Model Description
The model is a hybrid Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) network (SConvLSTM).
- Input: 9 inertial signals (body acc x/y/z, body gyro x/y/z, total acc x/y/z).
- Architecture: 3x 1D Conv layers (feature extraction) -> 2x LSTM layers (temporal modeling) -> Fully Connected layer.
- Task: Multi-class classification (6 activities: Walking, Walking Upstairs, Walking Downstairs, Sitting, Standing, Laying).
Training
The model was trained using PyTorch Lightning.
- Optimizer: AdamW
- Loss: CrossEntropyLoss
- Hyperparameters:
- Batch size: 64
- Epochs: 20
- Learning Rates: Sweep [0.1, 0.01, ..., 1e-6]
Results
Check the TensorBoard logs in this repository for training and validation performance across different learning rates.
