Hyperparameters

Detailed description of all training hyperparameters defined in hyperparameters.yaml.

Training parameters

Parameter	Value	Ultralytics default	Description
epochs	1000	100	Maximum number of full passes through the training dataset. Each epoch processes all images once.
imgsz	640	640	Input image resolution in pixels (images are resized to 640x640). Larger = more detail but slower.
batch	-1	16	AutoBatch: automatically determines the largest batch size that fits in GPU memory (~60% VRAM utilization). Lets each model train optimally for the available hardware.
patience	50	100	Early stopping: if mAP doesn't improve for 50 consecutive epochs, training stops. Prevents wasting time when the model has converged.

Optimizer

Parameter	Value	Ultralytics default	Description
optimizer	AdamW	auto	Optimizer algorithm. AdamW adds proper weight decay to Adam, preventing overfitting better than standard Adam.
lr0	0.001	0.01	Initial learning rate. Controls step size during gradient descent. Too high = unstable (NaN loss), too low = slow convergence. Large models and yolo12 medium use 0.0005 override to prevent NaN loss from attention layer sensitivity.
cos_lr	True	False	Cosine annealing: learning rate follows a cosine curve from lr0 down to near zero. Starts fast, then fine-tunes gradually at the end.
weight_decay	0.0005	0.0005	L2 regularization penalty. Discourages large weights, reducing overfitting. Applied correctly by AdamW (decoupled from gradient).

Augmentation — Geometric

Parameter	Value	Ultralytics default	Description
mosaic	1.0	1.0	100% probability. Combines 4 random training images into one by placing them in quadrants. Forces the model to learn objects at different positions and scales, especially effective for small object detection.
close_mosaic	50	10	Disables mosaic for the last 50 epochs. Lets the model fine-tune on clean, unmodified images before finishing — improves final accuracy.
mixup	0.15	0.0	15% probability. Blends two images together with transparency (alpha compositing). Creates soft transitions between classes, encouraging smoother decision boundaries.
copy_paste	0.5	0.0	50% probability. Cuts segmentation masks of objects from one image and pastes them onto another. Very effective for instance segmentation but computationally expensive — operates on individual masks.
flipud	0.5	0.0	50% probability of vertical flip (upside down). Useful when objects can appear in any vertical orientation.
fliplr	0.5	0.5	50% probability of horizontal flip (mirror). Standard augmentation — most objects look equally valid when mirrored.
scale	0.6	0.5	Random scaling factor +/-60% (image zoom between 0.4x and 1.6x). Teaches the model to recognize objects at different sizes.
degrees	25.0	0.0	Random rotation up to +/-25 degrees. Helps with objects that aren't always axis-aligned.
shear	5	0.0	Random shear up to +/-5 degrees. Simulates perspective distortion by slanting the image.
perspective	0.0005	0.0	Very slight random perspective transformation. Simulates camera angle variation. Small value keeps it subtle.

Augmentation — Color

Parameter	Value	Ultralytics default	Description
hsv_h	0.015	0.015	Random hue shift +/-1.5% of the color wheel. Small value — colors shift slightly but objects remain recognizable.
hsv_s	0.7	0.7	Random saturation change +/-70%. Large range — images can become very vivid or nearly grayscale. Teaches robustness to lighting/camera variation.
hsv_v	0.4	0.4	Random brightness (value) change +/-40%. Simulates different exposure levels — shadows, bright lights.

Transfer learning

Parameter	Value	Ultralytics default	Description
pretrained	False	True	Default: train from random weights (scratch). The training script overrides this to True for transfer learning runs, loading COCO-pretrained weights as the starting point.

Bold values indicate parameters that differ from Ultralytics defaults. 13 out of 22 parameters are customized.