Classic object detection 2.7.0¶
For standard object detection. The model list shows s / m / l / x / sod; SOD uses the actual small-object model configuration, not simply a smaller input size for an ordinary model.
Train from one COCO folder¶
Select a COCO folder and click Check dataset to verify classes and training/validation image counts. The software generates data configuration within the experiment; manually editing the built-in coco_detection file is unnecessary.
Unannotated OK images participate as negative background samples. Without a separate validation set, all training samples are retained: no automatic split, validation or mAP calculation occurs. Loss and learning rate still appear; this is not a training failure.
The advanced Use YAML dataset configuration option retains the model file's dataset settings; users must ensure correct paths and classes. Relative paths are resolved from the classic detection working directory, not automatically from an arbitrary external YAML's directory.
Height and width: three modes¶
| Size mode | Training | Validation, testing and export |
|---|---|---|
| Use model size settings | Preserve the reference configuration, including multi-scale training | Preserve the reference configuration |
| Fixed size | Use the specified height × width and disable random size selection | Use the same height × width |
| Customize validation and export size only | Preserve the original training size strategy | Use the specified height × width |
Enter height and width separately; both must be 96–8192 and multiples of 32. Custom sizes directly resize images without preserving the original aspect ratio.
For height 320 and width 640, select the third mode to reproduce changing only eval_height=320 and eval_width=640. Training may still randomly select square sizes from the original configuration; it does not become 320×640. Select the second mode if training must also use this fixed rectangle.
If random augmentation is also disabled, training preprocessing changes accordingly; the size mode does not re-enable disabled augmentation.
Validation metrics¶
With an independent validation set, evaluation runs at the model configuration's validation interval; metrics are not guaranteed every epoch. Charts show actual detection evaluation results such as mAP50-95, mAP50 and mAP75. Classification accuracy is not a general detection metric on this page.
No validation set, validation not yet due, and validation with no valid results are different states, shown separately in the charts. Missing values are not filled with 0; an actually computed 0 is retained.
Automatic export¶
After training, a saved checkpoint is selected automatically and the official inference export is run. After success, export/export_det in the experiment usually contains:
model.pdmodel
model.pdiparams
model.pdiparams.info
infer_cfg.yml
The export-completed log gives the actual path. infer_cfg.yml stores effective preprocessing and classes; custom dimensions are reflected in exported inputs and preprocessing. Export runs on CPU even when training used a GPU.
For other optimizer, loader and initialization settings, see classic common parameters; all controls are in the detection parameter table.
Evaluate on training set¶
Unchecked by default. In COCO mode, selecting this option evaluates all training images and annotations in place of a separate validation set, without automatic splitting or changes to the original data. Training data is used even when a separate validation set is available. This option is disabled when using YAML dataset settings.
AP/mAP on the training set reflects only the fit to that data and does not measure generalization to new images. Evaluation follows the interval in the model configuration. Raw logs retain the best AP information; charts and historical metrics identify the evaluation data source.