sconvlstm-uci-har-2

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.