Modern common parameters and augmentation¶
The four tasks share basic parameters, but not every augmentation applies to every task. Actual controls and defaults are listed in the parameter reference.
Basic parameters¶
| Parameter | Function and usage |
|---|---|
| Model weights | Select preset weights matching the task. Classification, detection, segmentation and OBB weights are not freely interchangeable. |
| imgsz | A single input size. Detection/segmentation/OBB resize and pad according to engine rules; classification additionally uses its own short-edge resize and crop. |
| epochs | Maximum epochs, not a guaranteed number of completed epochs. |
| patience | Consecutive epochs without validation metric improvement before early stopping; 0 disables it. Reaching epochs normally is not early stopping. |
| batch | Images per batch; increasing this generally uses more GPU memory and does not guarantee proportional speedup. |
| workers | Number of data loader processes; 0 loads in the main process. This is neither the training thread count nor the GPU count. |
| device | auto lets the engine choose; gpu maps to GPU 0; cpu trains on the CPU. |
| seed | Random seed for controlling reproducible randomness where supported. |
| amp | GPU automatic mixed precision; this does not force every tensor to half precision. |
The size field parses a single number. Non-digit text falls back to default size (640 for detection/segmentation/OBB, 960 for classification), rather than parsing 640×480 or lists. Check the logs; this field does not set rectangular dimensions.
Optimizer¶
optimizer=auto selects the optimizer, initial learning rate and momentum again, and sets warmup_bias_lr to 0. Select an explicit optimizer when you need manual learning-rate or momentum settings.
Modern classification has lr0 and lrf input fields; detection, segmentation and OBB pages have no separate learning rate field. Do not look for controls from tutorials for other software. Classification lrf is the final learning rate as a multiplier of the initial rate.
SGD-type optimizers use momentum; Adam-type optimizers use beta1. weight_decay controls weight decay. warmup_epochs is the warmup duration and accepts fractions; detection, segmentation and OBB also expose warmup starting momentum and bias learning rate.
Position and geometry augmentation¶
degrees is the maximum absolute rotation angle; flipud and fliplr are flip probabilities; translate is a shift relative to image size; scale is random scale variation. These are training augmentations, not inference postprocessing settings.
mosaic is the mosaic probability, also subject to the engine's late-training disabling rule. The UI has no separate control for the disabling epoch count; the current probability does not guarantee activation in every final epoch.
Software-adapted augmentation¶
For standard detection, copy_paste probabilistically copies a complete object within the same image to a non-overlapping position. It is skipped if the source overlaps another object or no safe destination exists. Polygon instance segmentation uses the native mask copy-paste fraction.
For detection and segmentation, erasing erases only background outside object boxes, leaving their interiors and existing annotations intact. It differs from classification random erasing that may cover objects arbitrarily.
The OBB page has no copy_paste or erasing controls. The classification page also has none of these background augmentations.
Loss weights¶
For detection, segmentation and OBB, box, cls and dfl are loss coefficients, not prediction confidence or class counts. OBB angle is an angle loss coefficient, unrelated to rotation augmentation degrees. Passing a coefficient does not create a loss term that the model architecture does not use.
Output location¶
Detection, segmentation and OBB results are stored in runs under the dataset directory; classification results use a sibling folder named after the dataset with the _runs suffix. The actual experiment path is the one recorded in the log, the chart's Open experiment directory action and training history.
The software automatically exports ONNX with simplify=False, without automatically installing onnxslim for graph simplification. If export fails, inspect the log; existing checkpoints may still be retained.