Skip to content

Latest commit

 

History

History
196 lines (149 loc) · 5.83 KB

README.md

File metadata and controls

196 lines (149 loc) · 5.83 KB

Digital Scale - Quick Start Guide

This folder contains a streamlined version of the Digital Scale BMI estimation system for quick setup and testing. It focuses on running forward passes (inference) with pre-trained models.

📁 Contents

  • model.py - DenseNet model definitions (SEDenseNet121 and SEDenseNet201)
  • dataset.py - Dataset handling and preprocessing utilities
  • predict_bmi.py - Main inference script for BMI prediction
  • requirements.txt - Python dependencies
  • .gitignore - Git ignore rules for weights and data

🚀 Quick Setup

1. Install Dependencies

pip install -r requirements.txt

2. Download Required Files

⚠️ Important: This folder does not include model weights or sample data. You need to download them separately:

Option A: Download from Authors (Recommended) Contact the authors to obtain the complete dataset and model weights:

Option B: Use Your Own Data If you have your own images, structure them as follows:

data/
├── visual-body-to-bmi.csv  # CSV with columns: individual_id, bmi, height_in, weight_lb, is_female, image_path
└── individual_folders/     # Folders named by individual_id
    ├── image1.jpg
    ├── image2.jpg
    └── ...

3. Organize Files

Place the downloaded files in the following structure:

get_started/
├── weights/
│   └── best_model.ckpt     # Model checkpoint file
├── data/
│   ├── visual-body-to-bmi.csv
│   ├── individual_folder_1/
│   ├── individual_folder_2/
│   └── ...
├── model.py
├── dataset.py
├── predict_bmi.py
└── requirements.txt

🔍 Usage

Basic Usage

Run BMI prediction on sample data:

python predict_bmi.py

Advanced Usage

Customize the prediction with various options:

# Use DenseNet201 instead of DenseNet121
python predict_bmi.py --model_type densenet201

# Specify custom paths
python predict_bmi.py --model_path weights/my_model.ckpt --data_dir my_data/

# Force CPU usage
python predict_bmi.py --device cpu

# Use larger batch size (if you have enough GPU memory)
python predict_bmi.py --batch_size 8

Command Line Options

Option Default Description
--model_path weights/best_model.ckpt Path to model checkpoint
--model_type densenet121 Model architecture (densenet121 or densenet201)
--data_dir data Directory containing dataset
--batch_size 1 Batch size for inference
--device auto Computation device (auto, cuda, or cpu)

📊 Expected Output

When running successfully, you should see output like:

Using device: cuda
Loading dataset...
Found 10 valid images for testing
Loading model...
Loaded densenet121 model from weights/best_model.ckpt

Running BMI predictions...
--------------------------------------------------------------------------------
ID: 1afea7     | Predicted: 38.06 | Actual: 38.51 | Error: 0.45
ID: 1afea7     | Predicted: 27.22 | Actual: 28.50 | Error: 1.28
ID: 1adchl     | Predicted: 32.70 | Actual: 32.32 | Error: 0.38
...
--------------------------------------------------------------------------------
Summary Statistics (n=10):
Average Error: 0.49
Max Error: 1.28
Min Error: 0.14

Completed predictions for 10 images.

🛠️ Troubleshooting

Common Issues

1. "Dataset CSV not found"

Dataset CSV not found at data/visual-body-to-bmi.csv
Please download the sample data from the authors:
Contact: rmanichand@ethz.ch or planger@ethz.ch

Solution: Download the dataset from the authors or check your data directory structure.

2. "Model checkpoint not found"

Model checkpoint not found at weights/best_model.ckpt
Please download the model weights from the authors:
Contact: rmanichand@ethz.ch or planger@ethz.ch

Solution: Download the model weights from the authors or check your weights directory.

3. "No valid images found"

No valid images found. Please check your data directory structure.
Expected structure:
  data/
    visual-body-to-bmi.csv
    individual_id_folders/
      *.jpg or *.png files

Solution: Ensure your data follows the expected directory structure.

4. CUDA Out of Memory

  • Use --device cpu to run on CPU
  • Reduce --batch_size to 1 (default)
  • Close other GPU-intensive applications

🔬 Model Details

SEDenseNet121

  • Architecture: DenseNet121 with Squeeze-and-Excitation blocks
  • Input Size: 224x224 RGB images
  • Output: Single BMI value (regression)
  • Parameters: ~7M parameters
  • Performance: MAPE ~7.9% on WayBED dataset

SEDenseNet201

  • Architecture: DenseNet201 with Squeeze-and-Excitation blocks
  • Input Size: 224x224 RGB images
  • Output: Single BMI value (regression)
  • Parameters: ~18M parameters
  • Performance: Potentially better accuracy but slower inference

🌟 Next Steps

After getting familiar with this quick start:

  1. Explore the full repository for training scripts, evaluation tools, and mobile deployment
  2. Check the main README for comprehensive documentation
  3. Try the CLAID mobile app for real-time BMI estimation
  4. Experiment with your own datasets following the data structure guidelines

📄 License

This project is part of the Digital Scale research. Please refer to the main repository for license information.

📧 Support

For questions, issues, or to request model weights and sample data:

Note: This quick start guide is designed to get you running BMI prediction quickly. For training new models, comprehensive evaluation, or mobile deployment, please refer to the main repository documentation.