Skip to main content
RF-DETR has reorganised its internal package layout. Backward-compatibility shim modules are provided so that existing code continues to work, but each shim emits a DeprecationWarning on import.
All deprecated modules emit a DeprecationWarning on import and will be removed in v1.7. Update your imports before upgrading past that version.

​
Module renames

The table below maps every deprecated module path to its canonical replacement.
Deprecated moduleCanonical replacement
rfdetr.util.coco_classesrfdetr.assets.coco_classes
rfdetr.util.miscrfdetr.utilities
rfdetr.util.loggerrfdetr.utilities.logger
rfdetr.util.box_opsrfdetr.utilities.box_ops
rfdetr.util.filesrfdetr.utilities.files
rfdetr.util.packagerfdetr.utilities.package
rfdetr.util.get_param_dictsrfdetr.training.param_groups
rfdetr.util.drop_schedulerrfdetr.training.drop_schedule
rfdetr.util.visualizerfdetr.visualize.data
rfdetr.deployrfdetr.export
rfdetr.models.segmentation_headrfdetr.models.heads.segmentation

​
Migration examples

# Before (deprecated)
from rfdetr.util.coco_classes import COCO_CLASSES

# Before (deprecated)
from rfdetr.util.misc import get_rank, get_world_size, is_main_process, save_on_master

# Before (deprecated)
from rfdetr.util.logger import get_logger

# Before (deprecated)
from rfdetr.util.box_ops import box_cxcywh_to_xyxy, generalized_box_iou

# Before (deprecated)
from rfdetr.util.get_param_dicts import get_param_dict

# Before (deprecated)
from rfdetr.util.drop_scheduler import drop_scheduler

# Before (deprecated)
from rfdetr.util.visualize import save_gt_predictions_visualization

# Before (deprecated)
from rfdetr.deploy import export_onnx

# Before (deprecated)
from rfdetr.models.segmentation_head import SegmentationHead

​
TrainConfig deprecated fields

The following fields on TrainConfig are deprecated as of the current release. Each field now belongs on ModelConfig — setting it on TrainConfig emits a DeprecationWarning.
These fields will be removed in v1.9 after a full release cycle. Move them to ModelConfig before upgrading.
Deprecated fieldReasonUse instead
TrainConfig.group_detrArchitecture decisionSet on ModelConfig
TrainConfig.ia_bce_lossLoss type tied to architecture familySet on ModelConfig
TrainConfig.segmentation_headArchitecture flagSet on ModelConfig
TrainConfig.num_selectPostprocessor count is an architecture decisionSet on ModelConfig; SegmentationTrainConfig users should remove the override entirely

​
Version history

VersionDateNotable changes
1.6.12026-03-25RFDETR.export(simplify=..., force=...) deprecated (no-op, removed in v1.8); PTL checkpoint fixes
1.6.02026-03-20rfdetr.deploy.* and rfdetr.util.* deprecated with redirects; PyTorch Lightning training; multi-GPU DDP; rfdetr[metrics] → rfdetr[loggers] rename
1.5.22026-03-04Peak GPU memory in progress bars; segmentation evaluation metric fixes
1.5.12026-02-27Nested Albumentations container support; torchvision-native transform pipeline
1.5.02026-02-23Custom training augmentations via aug_config; ClearML and MLflow loggers; live training progress bar

Build docs developers (and LLMs) love

Get started for freeTalk to us