YOLO1D - High-Performance Time-Series Detection

🚀 Overview

YOLO1D is a 1D convolutional neural network inspired by the YOLOv8 architecture, designed specifically for classification and detection on time-series data. It extends YOLO's strong detection capability to one-dimensional signals, accurately localizing and classifying events or patterns in complex sequences.

This project has undergone a major refactor, introducing a unified trainer, configuration management, automatic mixed precision (AMP) training, and advanced data augmentation to provide a high-performance, easy-to-use, and extensible time-series detection framework.

📁 Project Structure

yolo1d/
├── trainer_base.py          # Base trainer
├── trainer_amp.py           # AMP trainer
├── train.py                 # Unified training script
├── data_augmentation.py     # Data augmentation module
├── test_improvements.py     # Test scripts
├── yolo1d_model.py          # Model definition
├── yolo1d_loss.py           # Loss functions
├── dataset_generator.py     # Dataset generation
├── inference_yolo1d.py      # Inference script
├── config.yaml              # Configuration file
├── requirements.txt         # Dependencies
└── README.md                # Project documentation

🛠️ Quick Start

1. Environment Setup

# Create and activate a virtual environment
conda create --name yolo1d python=3.9 -y
conda activate yolo1d

# Install dependencies
pip install -r requirements.txt

2. Generate Dataset

python dataset_generator.py

3. Train the Model

We provide a unified training script that supports starting from a config file, resuming from checkpoints, and setting a custom run name.

# Train with the default config
python train.py

# Specify a config file
python train.py --config config.yaml

# Resume from a checkpoint
python train.py --resume best_model.pth

# Set a custom run name
python train.py --run-name my_experiment

4. Inference

python inference_yolo1d.py --model best_model.pth

🧠 Key Features and Improvements

1. Unified Trainer Architecture

Modular design with a base class (BaseTrainer) and derived classes (YOLO1DTrainer, YOLO1DAMPTrainer) for high code reuse and easy extensibility.

2. Config-Driven Experiments

All model, training, and data parameters are centrally managed via config.yaml, making experiments reproducible and configurable.

3. Enhanced Augmentation for Time-Series

Rich augmentations tailored for time-domain data, such as time warping, frequency masking, and Mixup. Use presets (light, medium, heavy) or customize parameters.

4. Automatic Mixed Precision (AMP)

Integrated AMP can speed up training by 1.5–2x and reduce memory usage by ~50% with minimal accuracy loss.

5. Early Stopping

Automatically monitors validation performance to stop training when improvements plateau, preventing overfitting and saving compute.

6. Detailed Monitoring and Logging

Use TensorBoard to track losses, mAP, learning rate, and more, with structured log outputs.

tensorboard --logdir runs

📝 Configuration

Core settings live in config.yaml.

# Model
model_size: 'n'          # model size (n, s, m, l, x)
num_classes: 2           # number of classes
input_channels: 1        # input channels
input_length: 1024       # input sequence length

# Data
dataset_path: 'sin_wave_dataset'
num_workers: 4

# Training
epochs: 100
batch_size: 16
learning_rate: 0.001
scheduler: 'cosine'      # LR scheduler ('cosine', 'onecycle', 'step')
patience: 10             # early stopping patience

# Loss hyperparameters
hyp:
  box: 1.0  # bbox loss weight
  cls: 3.0  # classification loss weight
  dfl: 0.8  # DFL loss weight

🎯 Model Performance

Model Size	Params	Inference Speed	Relative mAP
nano (n)	0.5M	Fastest	Baseline
small (s)	2M	Fast	+5%
medium (m)	5M	Medium	+10%
large (l)	10M	Slow	+15%
xlarge (x)	15M	Slowest	+20%

💡 Use Cases

• Anomaly Detection: Detect anomalies in industrial sensors, financial time series, etc.
• Signal Analysis: Identify and localize patterns in continuous signals (e.g., communications).
• Biomedical: Analyze ECG/EEG to detect arrhythmia or seizures.
• Speech Processing: Voice activity detection (VAD) or keyword spotting.
• Industrial Monitoring: Monitor equipment via vibration or acoustic signals.

🤝 Contributing

We welcome contributions of all kinds. Please follow the coding and commit guidelines.

Coding Style

• Use type annotations
• Add clear and complete docstrings
• Follow PEP 8

Commit Messages

feat: add new feature
fix: fix bug
docs: update documentation
style: code style adjustments
refactor: code refactor
test: add tests

📄 License

No License