Monster Jam Detection

Project Overview

This repository provides code and configuration for training a YOLO object detection model to identify die-cast toy Monster Jam trucks using a Roboflow dataset and MLflow for experiment tracking.

What is YOLO?

YOLO (You Only Look Once) is a cutting-edge object detection algorithm known for its incredible speed and efficiency. It examines an entire image in a single pass to identify all objects, making it ideal for real-time applications. 🚀

---

The Dataset

This project is trained on a custom dataset of my kid's die-cast toy Monster Jam trucks, which he lovingly leaves scattered all over our home. The images are managed on Roboflow and include augmentations—like rotations—to help the model recognize the toy trucks from any angle. The dataset is also mirrored on Hugging Face. 🤗

---

Project Goal

The primary goal is to train a model that can identify the various die-cast toy Monster Jam trucks with at least 90% accuracy.

Getting Started

Clone the repository:

git clone https://github.com/cbrownstein-lambda/monster-jam-detection.git
cd monster-jam-detection

Prerequisites

• Python 3.8+
• Docker (for MLflow server)
• Roboflow account and API key
• GPU recommended for training
• Recommended: Use a Python virtual environment (venv) to avoid package conflicts.
• Required Python modules (install with pip):
  • mlflow
  • torch
  • ultralytics
  • roboflow
  • pynvml
  • psutil

Set up a Python virtual environment

python3 -m venv .venv
source .venv/bin/activate

Install all required modules:

pip install -U mlflow torch ultralytics roboflow pynvml psutil

1. Start MLflow Tracking Server

MLflow is used to track experiments and results. The project includes a mlflow.compose.yaml for easy setup with Docker Compose.

Note for Lambda On-Demand Cloud users: The ubuntu user may need to be added to the docker group to run Docker commands without sudo:

sudo usermod -aG docker ubuntu
# Log out and back in for group changes to take effect

Alternatively, you can prefix Docker commands with sudo:

sudo docker compose -f mlflow.compose.yaml up -d

# Start MLflow server (from project root)
docker compose -f mlflow.compose.yaml up -d

• MLflow UI will be available at http://127.0.0.1:5000
• Data is stored in a Docker volume (mlflow_backend)

2. Download the Dataset from Roboflow

Use download_data.py to fetch the dataset. You need a Roboflow API key. You can pass it as an argument or set it as an environment variable.

# Set your Roboflow API key (recommended)
export ROBOFLOW_API_KEY=your_api_key_here

# Download the dataset (default settings)
python download_data.py

# Custom options:
python download_data.py --api_key your_api_key --workspace cody-brownstein --project monster-jam-detection --version_number 12 --download_format yolov8

Defaults:

• Workspace: cody-brownstein
• Project: monster-jam-detection
• Version: 12
• Format: yolov8

The dataset will be downloaded to a folder printed in the output.

‼️ Note: ‼️ Downloaded datasets are expected to be placed in the datasets/ directory. Make sure your data config path (e.g., datasets/Monster-Jam-Detection-12/data.yaml) matches the location of your downloaded dataset.

You can also download the datasets directly from Roboflow Universe.

3. Train the YOLO Model

Use train_yolo.py to start training. MLflow will automatically log parameters and results.

Important: Ensure MLflow logging is enabled in YOLO settings. In Ultralytics YOLO, set mlflow=True in your training command or configuration if it's not already enabled. See the Ultralytics MLflow integration docs for details.

# Train with default settings
python train_yolo.py

# Custom options:
python train_yolo.py \
  --tracking_uri http://127.0.0.1:5000 \
  --experiment_name "Monster Jam Detection" \
  --data_config datasets/Monster-Jam-Detection-12/data.yaml \
  --model_name yolo11l.pt \
  --epochs 100 \
  --img_size 640 \
  --batch_size 64

Defaults:

• Tracking URI: http://127.0.0.1:5000
• Experiment Name: Monster Jam Detection
• Data Config: datasets/Monster-Jam-Detection-12/data.yaml
• Model Name: yolo11l.pt
• Epochs: 100
• Image Size: 640
• Batch Size: 64

Tip: System metrics logging (CPU, RAM, GPU, etc.) is enabled by default in the training script. If you want to control this manually, set the environment variable before running:

export MLFLOW_ENABLE_SYSTEM_METRICS_LOGGING=true  # Enable
export MLFLOW_ENABLE_SYSTEM_METRICS_LOGGING=false # Disable

4. View Results

Open http://127.0.0.1:5000 in your browser to view experiment runs, metrics, and artifacts.

Example Artifacts and MLflow Screenshots

Below are example artifacts produced by YOLO11 training and example MLflow UI screenshots to help visualize and evaluate your model's performance:

YOLO Training Artifacts

• val_batch0_labels

  

  Shows ground truth labels for a batch of validation images.

• val_batch0_pred

  

  Shows predicted bounding boxes and classes for the same batch of validation images.

• confusion_matrix

  

  Displays the confusion matrix summarizing model accuracy and misclassifications across all classes.

• results

  

  Shows training and validation metrics (e.g., loss, mAP) over epochs.

MLflow UI Screenshots

• MLflow Chart View (Model Metrics for Two Runs)

  

  Screenshot of MLflow chart view showing model metrics for two runs.

• MLflow Comparison View (Two Runs)

  

  Screenshot of MLflow comparison view for the same two runs.

Environment Variables

You can set the following environment variables to override defaults:

• ROBOFLOW_API_KEY
• ROBOFLOW_WORKSPACE
• ROBOFLOW_PROJECT
• ROBOFLOW_VERSION_NUMBER
• ROBOFLOW_DOWNLOAD_FORMAT
• MLFLOW_EXPERIMENT_NAME

---

For questions or issues, please open an issue on GitHub.