Advanced Training Parameters¶

YOLOmatic exposes 70+ YOLO training parameters through the Configure Model → Advanced Settings wizard step. This reference documents every parameter with its default value, valid range, and practical guidance.

Parameters are organized into eight categories matching the wizard's interface.

Core Parameters¶

Parameter	Default	Range	Description
`epochs`	`300`	1–10000	Total training iterations (passes over the full dataset). 100 is sufficient for small fine-tuning; 300 is standard; 500+ for complex datasets where accuracy is still improving.
`patience`	`50`	0–500	Early stopping: halts training if no improvement for this many epochs. Increase to 100 for noisy datasets. Set to 0 to disable early stopping.
`batch`	`-1`	-1–1024	Images per training step. `-1` enables Auto-Batch, which finds the largest size that fits in VRAM. Only set manually if Auto-Batch causes issues.
`imgsz`	`640`	32–2048	Input resolution in pixels. Use 320–416 for mobile/speed, 640 for standard, 1280 for small-object detection. Larger sizes use significantly more VRAM.
`time`	`None`	0–1000	Maximum training duration in hours. Useful on cloud platforms billed by the hour. Training saves and stops when the limit is reached.

Hardware Parameters¶

Parameter	Default	Range	Description
`device`	`"0"`	`0`, `0,1`, `cpu`, `mps`, `cuda`, `npu`	Compute device. `0` = first NVIDIA GPU. `cpu` = CPU inference (slow). `mps` = Apple Silicon. `0,1` = multi-GPU.
`workers`	`8`	0–64	CPU dataloader threads. Increase if GPU utilization is below 80%; decrease if high CPU usage or DataLoader warnings appear.
`cache`	`False`	`False`, `True`/`"ram"`	Dataset caching. `"ram"` caches images in memory (fast, requires sufficient RAM). Persistent disk caching is disabled because it writes uncompressed `.npy` files beside source images. `False` reads images from storage each epoch.

Optimizer Parameters¶

Parameter	Default	Range	Description
`optimizer`	`"auto"`	auto, SGD, Adam, AdamW, Adamax, NAdam, RAdam, RMSProp, MuSGD	Learning algorithm. `auto` lets YOLO pick based on model size. `SGD` is reliable for large datasets. `AdamW` converges faster on smaller datasets. `MuSGD` is the YOLO26-era hybrid optimizer.
`lr0`	`0.01`	0.000001–1.0	Initial learning rate. 0.01 is the standard starting point. Reduce to 0.001 if loss is erratic; raise to 0.02 if progress is too slow.
`lrf`	`0.01`	0.0001–1.0	Final learning rate as a fraction of `lr0`. 0.01 means training finishes at 1% of the starting rate.
`momentum`	`0.937`	0.0–1.0	SGD momentum (gradient inertia). Heavily tuned for YOLO — do not change unless you are doing research.
`weight_decay`	`0.0005`	0.0–0.1	L2 regularization penalty on large weights. Increase to 0.001 if the model is overfitting on small datasets.
`warmup_epochs`	`3.0`	0.0–50.0	Number of warmup epochs at the start of training where the learning rate ramps up slowly. Prevents early instability. Increase to 5.0 for very large models.
`warmup_momentum`	`0.8`	0.0–1.0	Starting momentum value during the warmup phase; grows to `momentum` over `warmup_epochs`.
`cos_lr`	`False`	bool	Use a cosine annealing schedule instead of linear decay. Often produces slightly better final accuracy by annealing more smoothly.

Augmentation Parameters¶

These parameters control on-the-fly training augmentation applied by the Ultralytics dataloader. For offline augmentation, see Augmentation guide.

Parameter	Default	Range	Description
`hsv_h`	`0.015`	0.0–1.0	Hue shift range. Keep small (0.015). Large values change object color semantics.
`hsv_s`	`0.7`	0.0–1.0	Saturation shift range. 0.7 simulates varied camera quality and lighting.
`hsv_v`	`0.4`	0.0–1.0	Brightness/value shift range. 0.4 handles sunny vs. overcast lighting variation.
`degrees`	`0.0`	-180–180	Random rotation in degrees. 0.0 for upright objects (cars, people); 15–30 for tilted cameras; 180 for orientation-independent objects (cells, minerals).
`translate`	`0.1`	0.0–1.0	Random image translation as a fraction of image size. 0.1 = 10% shift in x/y.
`scale`	`0.5`	0.0–1.0	Random zoom range. 0.5 means scale between 0.5× and 1.5×. Critical for detecting objects at varied distances.
`shear`	`0.0`	-180–180	Random shear (slant) in degrees. Use for extreme camera angles only.
`perspective`	`0.0`	0.0–1.0	3D perspective distortion. Use 0.0001–0.001 for drone or security footage.
`flipud`	`0.0`	0.0–1.0	Vertical flip probability. 0.0 for normal cameras; 0.5 for satellite or top-down imagery.
`fliplr`	`0.5`	0.0–1.0	Horizontal flip probability. 0.5 is standard. Set to 0.0 when left/right orientation is semantically meaningful (text, directional signs).
`mosaic`	`1.0`	0.0–1.0	Probability of stitching four training images into one. A core YOLO augmentation technique — leave at 1.0 for most runs.
`mixup`	`0.0`	0.0–1.0	Probability of overlaying two images. Use 0.1 for large, complex datasets. Keep at 0.0 for small datasets.
`copy_paste`	`0.0`	0.0–1.0	Copy-paste segmentation augmentation — pastes object masks from one image into another. Requires segmentation task.
`auto_augment`	`"randaugment"`	`""`, `randaugment`, `autoaugment`, `augmix`	Automatic augmentation policy. `randaugment` is the modern default. Set to `""` to disable.
`erasing`	`0.4`	0.0–1.0	Random erasing probability — blacks out random image regions. Teaches the model to recognize partially occluded objects.
`close_mosaic`	`10`	0–100	Number of final epochs that run without mosaic augmentation. Allows the model to fine-tune on clean images before the run ends.

Loss Weight Parameters¶

Parameter	Default	Range	Description
`box`	`7.5`	0.0–20.0	Weight for bounding box coordinate loss. Increase if boxes are slightly off while classification is correct.
`cls`	`0.5`	0.0–10.0	Weight for classification loss. Increase if box locations are correct but class labels are wrong.
`dfl`	`1.5`	0.0–10.0	Distribution Focal Loss weight for bounding-box edge precision. Rarely needs adjustment.
`nbs`	`64`	1–1024	Nominal batch size for loss scaling normalization. Do not change unless you are an ML researcher adjusting loss scaling behavior.

Advanced Training Control¶

Parameter	Default	Range	Description
`amp`	`True`	bool	Automatic Mixed Precision (FP16). Reduces VRAM use and speeds training. Disable only if you see `NaN` loss values.
`pretrained`	`True`	bool	Start from COCO-pretrained weights. Disable only when your domain is radically different from natural images (thermal, medical, ultrasound).
`deterministic`	`False`	bool	Force deterministic algorithms for full reproducibility. Slows training; use for scientific paper experiments.
`seed`	`0`	0–999999	Random seed for reproducible augmentation order and initialization.
`rect`	`False`	bool	Rectangular training — groups images by aspect ratio to reduce padding. Can speed up training on wide-format datasets.
`fraction`	`1.0`	0.01–1.0	Fraction of the dataset to use. Set to 0.1 for a quick smoke test before committing to a full run.
`multi_scale`	`0.0`	0.0–1.0	Change image size slightly each batch to teach scale invariance. Set 0.1 for datasets with objects at highly varied sizes.
`resume`	`False`	bool	Resume training from `last.pt`. Use when training was interrupted and you want to continue from the last saved state.
`freeze`	`None`	0–100	Number of early layers to freeze during fine-tuning. Try `10` for small datasets that are visually similar to the source domain.
`single_cls`	`False`	bool	Treat all objects as one class. Use when you only need to locate objects without classifying them.
`save_period`	`-1`	-1–1000	Save a checkpoint every N epochs. `-1` saves only best and last. Set to 50 to enable periodic checkpoints for crash recovery.
`profile`	`False`	bool	Profile model speed during training. For deployment research; not needed for normal accuracy tuning.

Segmentation Parameters¶

These parameters apply only to segmentation tasks (when the model variant ends in -seg).

Parameter	Default	Range	Description
`overlap_mask`	`True`	bool	Allow object masks to overlap. Should be `True` for natural scenes where objects occlude each other.
`mask_ratio`	`4`	1–16	Mask resolution relative to image size. Lower values (1–2) produce higher-quality masks but use more VRAM. 4 is the recommended balance.

Validation Parameters¶

Parameter	Default	Range	Description
`val`	`True`	bool	Run validation after every epoch. Leave enabled to track accuracy during training.
`conf`	`None`	0.0–1.0	Confidence threshold for validation detections. `None` uses Ultralytics defaults. Use `0.001` to measure maximum potential recall; use `0.25+` for real-world deployment thresholds.
`max_det`	`300`	1–10000	Maximum detections per image during validation. Increase for dense detection tasks (crowd counting, aerial imagery).
`plots`	`True`	bool	Generate training curve and confusion matrix plots after training. Leave enabled for post-training analysis.

Output Parameters¶

Parameter	Default	Description
`save`	`True`	Write checkpoints and run artifacts. Disable only for throwaway diagnostic runs.
`project`	`None`	Parent directory for run output. Defaults to the trainer's standard `runs/` folder.
`name`	`None`	Name of the run subdirectory. YOLOmatic generates a timestamped name when left blank.
`exist_ok`	`False`	Allow reusing an existing `project/name` directory. Keep `False` to avoid mixing run artifacts.
`verbose`	`True`	Print detailed trainer logs. Disable for quieter output in scripted environments.

RF-DETR Training Parameters¶

RF-DETR uses a separate trainer with different parameter names. The key parameters exposed through YOLOmatic's RF-DETR configuration wizard are:

Parameter	Typical Values	Description
`epochs`	50–100	Training epochs
`batch_size`	4–8	Transformer models need more VRAM per image than CNN-based models
`lr`	`1e-4`	Learning rate
`grad_accum_steps`	4	Gradient accumulation steps for effective larger batch sizes
`resolution`	384–880	Input resolution; RF-DETR models have model-specific defaults

See RF-DETR guide for the full RF-DETR workflow.

Related pages: Configuration, YOLO Guide, Augmentation.