WildTrain Detection Training Configuration

Detailed reference for detection training, evaluation, and sweep YAML configuration files.

Overview

WildTrain supports YOLO-based object detection via Ultralytics. Training configs define the model, data, training hyperparameters, and MLflow tracking.

Usage:

# Train
wildtrain train detector -c configs/detection/yolo_configs/yolo.yaml

# Evaluate
wildtrain evaluate detector -c configs/detection/yolo_configs/yolo_eval.yaml

# Hyperparameter sweep
wildtrain pipeline detection -c configs/detection/detection_sweep.yaml

---

Training Config (yolo.yaml)

The YOLO training config follows the Ultralytics configuration format with additional WildTrain fields.

Key Fields

Field	Type	Description
model	str	Path to YOLO model config or pretrained weights
data	str	Path to YOLO data YAML file
epochs	int	Number of training epochs
imgsz	int	Input image size
batch	int	Batch size
device	str/int	Device: 0 (GPU 0), cpu, etc.
lr0	float	Initial learning rate
lrf	float	Final learning rate factor
optimizer	str	Optimizer: SGD, Adam, AdamW
weight_decay	float	Weight decay for regularization

Data YAML Format

The data field points to a standard YOLO data config:

train: ./images/train
val: ./images/val
nc: 3
names: ['elephant', 'giraffe', 'zebra']

---

Evaluation Config (yolo_eval.yaml)

Used with wildtrain evaluate detector -c CONFIG.

Field	Type	Description
model	str	Path to trained YOLO weights
data	str	Path to YOLO data YAML
imgsz	int	Input image size
batch	int	Batch size
device	str/int	Device for evaluation

---

Hyperparameter Sweep Config

Used with wildtrain pipeline detection -c CONFIG for Optuna-based hyperparameter optimization.

Sweep Structure

# Base config used as template
base_config: configs/detection/yolo_configs/yolo.yaml

# Search space
parameters:
  model: null          # Model variants (optional)
  train:
    lr0: [0.0001, 0.001, 0.01]
    lrf: [0.01, 0.1, 0.5]
    batch: [8, 16, 32]
    epochs: [10, 20, 30]
    imgsz: [640, 1280]
    optimizer: [SGD, AdamW, Adam]
    weight_decay: [0.0001, 0.0005, 0.001]

Sweep Settings

Field	Type	Default	Description
base_config	str	—	Path to base training config
sweep_name	str	—	Name for the sweep experiment
n_trials	int	20	Number of Optuna trials
seed	int	42	Random seed
objective	str	map_50	Metric to optimize: precision, recall, f1_score, map, map_50, map_50_95, fitness
direction	str	maximize	Optimization direction: minimize or maximize
timeout	int/null	null	Maximum optimization time in seconds

Output Settings

Field	Type	Default	Description
output.directory	str	results/sweeps/{name}	Results output directory
output.save_results	bool	true	Save trial results
output.save_plots	bool	true	Save optimization plots
output.format	str	both	Output format: json, csv, or both
output.include_optimization_history	bool	true	Include full optimization history

---

Complete Training Example

# Model
model: yolo12n.pt
data: configs/detection/yolo_configs/data/wildlife.yaml

# Training
epochs: 100
imgsz: 800
batch: 16
device: 0

# Learning rate
lr0: 0.001
lrf: 0.01
optimizer: AdamW
weight_decay: 0.0005

# Augmentation (built into Ultralytics)
augment: true
mosaic: 1.0
mixup: 0.0

---

See also:

• WildTrain CLI Reference — train detector, evaluate detector commands
• YOLO Config Guide
• Model Training Tutorial