Batch size is the number of training samples processed in one iteration. In production training, it is configured alongside learning rate, model architecture, and image size to balance efficiency, resource use, accuracy, and training time. section 8.1.2
When configuring batch size, account for:
- GPU memory: larger batches can exceed available memory. section 8.1.2
- Processing speed: increasing the batch size can make processing faster. section 8.1.2
- Optimization behavior: larger batches can reduce the number of gradient updates and may decrease performance. section 8.1.2
- Iteration and validation: change the batch size with learning rate or epoch settings when training fails to converge. section 8.1.7
Ultralytics training configuration
A pipeline can pass batch as a configurable value to the Ultralytics YOLO train method. The example uses batch=batch alongside a dataset configuration, image size, epoch count, model name, and artifact directory. section 8.1.2
Listing 8.3 Writing the data configuration to YAML
from ultralytics import YOLO
model = YOLO('yolov8n.pt') #1
results = model.train( #2
data='custom_data.yaml',
imgsz=640,
epochs=epochs,
batch=batch,
name='yolov8n_custom',
project=project_path) #3
Inspect the runnable YOLO training configuration, including the batch parameter and its surrounding production settings. section 8.1.2
