FireYOLO

Object detection training and deployment pipeline built on top of TinyissimoYOLO, using a customized Ultralytics codebase. Designed for training lightweight YOLO models that can run on resource-constrained platforms (MCUs, smart glasses, etc.).

---

Setup

Requirements: Python >= 3.10, PyTorch >= 1.7, CUDA (recommended).

python3.10 -m venv venv
source venv/bin/activate
pip install -r requirements.txt

---

Repository Structure

FireYOLO/
├── a_train_export.py            # Main script: train, export to ONNX, and validate
├── inference.py                 # Inference script (Ultralytics + ONNX Runtime flow)
├── nms.py                       # Custom NMS utilities
├── requirements.txt             # Python dependencies
├── sweeps/                      # W&B hyperparameter sweep configs
│   └── sweep_models_all_classes.yaml
├── ultralytics/                 # Local (modified) Ultralytics codebase
│   └── cfg/
│       ├── models/tinyissimo/   # Tinyissimo model definitions
│       └── datasets/            # Dataset YAMLs (COCO, VOC, etc.)
├── results_save/                # Saved training/export runs
└── wandb/                       # Weights & Biases run logs and artifacts

---

Training

Standard Training (a_train_export.py)

This is the main entry point. It trains a TinyissimoYOLO model, exports to ONNX, and runs validation — all in one script. Results are logged to Weights & Biases.

If the selected dataset is not already available locally, the script will download it automatically.

python a_train_export.py --model tinyissimo-v8-n --data coco

Arguments:

Argument	Default	Description
--model	tinyissimo-v8-n	Model architecture. Maps to a YAML file in ultralytics/cfg/models/tinyissimo/.
--data	coco	Dataset config name (without .yaml). Maps to ultralytics/cfg/datasets/.

Available models: tinyissimo-v8-b, tinyissimo-v8-n, tinyissimo-v8-s, tinyissimo-v8-m, tinyissimo-v8-l, tinyissimo-v8-x

These correspond to different width/depth scaling factors (from smallest b to largest x). Batch size is automatically adjusted based on model size to fit in ~24 GB VRAM.

Training details:

• Optimizer: SGD
• Image size: 128x128
• Epochs: 1000
• Output: results/<model>_<data>/ (weights, metrics, ONNX export)

---

Evaluation

Validation is run automatically at the end of training by a_train_export.py, both for the trained PyTorch model and for the exported ONNX model.

---

Inference

Run inference on a single image using a trained model (supports both .pt and .onnx):

python inference.py

Edit model_path_onnx and image_path inside the script to point to your model and image. The script runs both Ultralytics and custom NMS pipelines and visualizes detections.

---

Model Configuration

Model architectures live in ultralytics/cfg/models/tinyissimo/. Each YAML defines backbone and head layers, plus scaling factors:

nc: 80        # Number of classes — change this to match your dataset
scale: 'n'    # Scale variant (b, n, s, m, l, x)

To use a custom number of classes, edit the nc field in the model YAML or in the dataset YAML.

---

Dataset Configuration

Dataset YAML files in ultralytics/cfg/datasets/ define paths to images/labels and class names. Available configs include COCO, VOC, and others. To add a custom dataset, create a new YAML following the same format.

---

W&B Sweeps

To run a grid sweep across all models and datasets:

wandb sweep sweeps/sweep_models_all_classes.yaml
wandb agent <sweep_id>

This trains every combination of model size (b through x) on both VOC and COCO.

---

Acknowledgments

Based on TinyissimoYOLO by ETH Zurich, built on top of Ultralytics.