README.md

ChildHope_DeepLearning

Setup

Prerequisites

• Python 3.11
• Poetry (Python package manager)

Installation Steps

1. Clone the repository

   $ git clone https://github.com/your-repo/ChildHope_DeepLearning.git
   

2. Install poetry if not already installed

   • Unix/MacOS/WSL:

     curl -sSL https://install.python-poetry.org | python3 -
     

   • Windows:

     (Invoke-WebRequest -Uri https://install.python-poetry.org -UseBasicParsing).Content | py -
     

   New to Poetry? Check out the beginner's guide.

3. Package Dependencies

   $ poetry install
   

Project Structure

childhope/
├── models/                    # Neural network architectures
│   ├── baseline_lstm.py      # Reference LSTM implementation
│   ├── tide.py              # TiDE (Time-series Dense Encoder) implementation
│   └── your_model.py         # Add your new architectures here
├── data_processing/          # Data preprocessing utilities
│   ├── preprocessing.py      # Current preprocessing functions
│   ├── eda.py               # Exploratory Data Analysis utilities
│   └── your_processing.py    # Add your preprocessing methods here
└── common/                   # Shared utilities
    └── logger.py             # Logging configuration

experiments/                  # Different training approaches
├── baseline/                # Reference LSTM experiment
├── tide/                    # TiDE model experiments and visualization
└── your_experiment/         # Add your experiments here

tensorboard_logs/            # Experiment tracking and visualization
├── experiment_1_timestamp/  # Training/validation metrics, model behavior
└── experiment_2_timestamp/  # Learning rates, predictions, custom metrics

data/                        # Datasets (not tracked in git)

Contributing

Adding New Models

Place new model architectures in childhope/models/. The baseline LSTM provides a reference implementation for vital signs prediction.

Adding New Preprocessing Methods

Preprocessing methods go in childhope/data_processing/. Current preprocessing includes:

• Time series resampling
• Missing value handling
• Train/validation/test splitting

The data splitting strategy ensures patient independence:

• Patients are randomly split into train (80%) and test (20%) sets
• All time series from a patient are kept together in either train or test
• Validation set is created from the training patients (not test patients)
• This prevents data leakage and ensures model evaluation on completely unseen patients

Feel free to add new methods that might improve data quality or model performance.

Adding New Experiments

The experiments/ directory contains different training approaches. Each experiment should:

• Have its own directory
• Log results to the common TensorBoard directory for easy comparison:

  # Standard logging setup for all experiments
  experiment_name = "your_experiment_name"  # e.g., "lstm_attention", "transformer", etc.
  timestamp = datetime.datetime.now().strftime("%Y%m%d_%H%M%S")
  log_dir = os.path.join(os.path.dirname(__file__), "../../tensorboard_logs", experiment_name + "_" + timestamp)
  writer = SummaryWriter(log_dir)

• Use the following standard metrics for comparison:
  • Training loss
  • Validation loss (plotted together with training loss)
  • Learning rate
  • Any experiment-specific metrics

Take a look at the baseline LSTM experiment in experiments/baseline/ for reference implementation, logging setup, and overall structure.

Comparing Experiments in TensorBoard

To view and compare all experiments:

tensorboard --logdir=tensorboard_logs

This will display all experiments in the same TensorBoard interface. Here are some examples comparing two LSTM implementations - one with a single layer and another with two layers:

Training Loss Comparison

Track and compare training loss across different experiments to evaluate convergence rates and stability.

Validation Loss Comparison

Compare validation performance to assess generalization and identify potential overfitting.