RT-DETR

Standalone install? If this session was not initialized by the TAO skill bank plugin, run the tao-setup skill first (host preflight, credentials, cross-skill discovery).

RT-DETR (Real-Time DEtection TRansformer) for 2D object detection. Designed for real-time inference with competitive accuracy. Supports distillation and quantization for deployment optimization.

Set model.pretrainedbackbonepath for backbone weights or train.pretrainedmodelpath for full model.

For TAO Deploy TensorRT actions (gentrtengine, TensorRT evaluate, and TensorRT inference), read references/tao-deploy-rtdetr.md first. Deploy spec templates live in this skill's references/ folder with the spectemplatedeploy_*.yaml prefix.

Dataclass Schemas

Generated TAO Core schemas are packaged in schemas/<action>.schema.json, with schemas/manifest.json listing available actions. Each generated schema also emits references/spectemplate<action>.yaml from the schema top-level default field. AutoML enablement is declared at the model layer in references/skillinfo.yaml via automlenabled. Runnable AutoML for an action requires schemas/<action>.schema.json and references/spectemplate<action>.yaml to exist and parse. Use the packaged selected-action schema for automldefaultparameters, automldisabledparameters, defaults, min/max bounds, enums, option weights, math conditions, dependencies, and popular parameters. Do not expect ~/tao-core at runtime; maintainers regenerate schemas/templates before packaging the skill bank.

Train Action Policy

This model is AutoML-enabled at the model layer. Before handling any train-stage request, read references/skillinfo.yaml and resolve the run override from either an explicit automlpolicy value or the user's workflow request. Use automlpolicy: on by default and only expose on / off in new launch prompts. Treat phrases like "turn off AutoML", "disable AutoML", "no HPO", or "plain training" as automlpolicy: off for this run only. When automlpolicy: on, automlenabled: true, and both schemas/train.schema.json and references/spectemplatetrain.yaml are packaged, route the train action through tao-skill-bank:tao-run-automl by default with this model's skilldir. Preserve workflow/application overrides for datasets, specs, output directories, GPU/platform settings, parent checkpoints, and automlpolicy. Use direct model training only when automl_policy: off or the packaged train schema/template is missing; in the missing-schema case, report that AutoML is enabled but not runnable for this model until schemas are generated.

Non-train actions such as evaluate, inference, export, and deploy flows stay in this model skill. The per-run automl_policy override does not change model metadata.

Supported Actions

The packaged RT-DETR PyT CLI supports train, distill, quantize, evaluate, export, inference, and defaultspecs. This model skill exposes train, distill, quantize, evaluate, export, and inference; resume/retrain is performed through train with train.resumetrainingcheckpointpath.

The parent PyT CLI does not expose gentrtengine. Use models/rtdetr/deploy for TensorRT engine generation, TensorRT evaluation, and TensorRT inference.

Training Requirements

• Dataset type: object_detection
• Formats: coco, coco_raw
• Monitoring metric: mAP50 for quick operational checks; val_mAP for COCO/paper-style benchmark comparisons.

Per-Action Dataset Requirements

Typical Spec Overrides

Data source overrides are mandatory for every action — the agent MUST construct data source paths from the Per-Action Dataset Requirements table above and include them in spec_overrides.

S3_TRAIN = "s3://bucket/data/train"
S3_EVAL = "s3://bucket/data/eval"
CHECKPOINT = "/results/{train_job_id}/results_dir/model_epoch_000.pth"
ONNX_FILE = "/results/{export_job_id}/results_dir/rtdetr.onnx"

train (mandatory data sources):

{
    "train.num_epochs": 10,
    "train.checkpoint_interval": 10,
    "train.validation_interval": 10,
    "train.num_gpus": 1,
    "train.gpu_ids": [0],
    "dataset.num_classes": "<num_classes> + 1",
    "dataset.train_data_sources": [{"image_dir": f"{S3_TRAIN}/images.tar.gz", "json_file": f"{S3_TRAIN}/annotations.json"}],
    "dataset.val_data_sources": {"image_dir": f"{S3_EVAL}/images.tar.gz", "json_file": f"{S3_EVAL}/annotations.json"},
}

resume train (mandatory checkpoint):

{
    "train.num_epochs": 11,
    "train.resume_training_checkpoint_path": CHECKPOINT,
    "dataset.num_classes": "<num_classes> + 1",
    "dataset.eval_class_ids": [1, 2, 3, 4],
    "dataset.train_data_sources": [{"image_dir": f"{S3_TRAIN}/images.tar.gz", "json_file": f"{S3_TRAIN}/annotations.json"}],
    "dataset.val_data_sources": {"image_dir": f"{S3_EVAL}/images.tar.gz", "json_file": f"{S3_EVAL}/annotations.json"},
}

evaluate (mandatory data sources and checkpoint):

{
    "dataset.num_classes": "<num_classes> + 1",
    "dataset.eval_class_ids": [1, 2, 3, 4],
    "dataset.test_data_sources": {"image_dir": f"{S3_EVAL}/images.tar.gz", "json_file": f"{S3_EVAL}/annotations.json"},
    "evaluate.checkpoint": CHECKPOINT,
}

export (mandatory checkpoint and output):

{
    "dataset.num_classes": "<num_classes> + 1",
    "export.checkpoint": CHECKPOINT,
    "export.onnx_file": ONNX_FILE,
    "export.input_height": 640,
    "export.input_width": 640,
}

quantize (mandatory data sources):

{
    "dataset.num_classes": "<num_classes> + 1",
    "quantize.layers": [
        {
            "module_name": "*",
            "weights": {
                "dtype": "float8_e4m3fn"
            },
            "activations": {
                "dtype": "float8_e4m3fn"
            }
        }
    ],
    "dataset.train_data_sources": [{"image_dir": f"{S3_TRAIN}/images.tar.gz", "json_file": f"{S3_TRAIN}/annotations.json"}],
    "dataset.val_data_sources": {"image_dir": f"{S3_EVAL}/images.tar.gz", "json_file": f"{S3_EVAL}/annotations.json"},
    "dataset.quant_calibration_data_sources": {"image_dir": f"{S3_TRAIN}/images.tar.gz", "json_file": f"{S3_TRAIN}/annotations.json"},
    "quantize.model_path": CHECKPOINT,
}

inference (mandatory data sources and checkpoint):

{
    "dataset.num_classes": "<num_classes> + 1",
    "dataset.infer_data_sources": {"image_dir": [f"{S3_EVAL}/images.tar.gz"], "classmap": f"{S3_EVAL}/label_map.txt"},
    "inference.checkpoint": CHECKPOINT,
}

distill (mandatory data sources and teacher checkpoint):

{
    "dataset.train_data_sources": [{"image_dir": f"{S3_TRAIN}/images.tar.gz", "json_file": f"{S3_TRAIN}/annotations.json"}],
    "dataset.val_data_sources": {"image_dir": f"{S3_EVAL}/images.tar.gz", "json_file": f"{S3_EVAL}/annotations.json"},
    "distill.pretrained_teacher_model_path": CHECKPOINT,
}

Eval Dataset

Optional. Provides validation mAP at each checkpoint if supplied.

Important Parameters

• dataset.num_classes: Number of classes. Default 80 (MSCOCO 80-class). Must match your dataset annotations.
• model.backbone: Default resnet_50. Supported: ResNet variants, ConvNeXt, FAN, EfficientViT. RT-DETR is optimized for real-time with lighter backbones.
• train.optim.lr: Learning rate. Default 1e-4 (lower than DINO's 2e-4). lr_backbone defaults to 1e-5.
• dataset.augmentation.trainspatialsize: Training input size. Default [640, 640]. Smaller than DINO's multi-scale (up to 1333). Key to RT-DETR's speed.
• model.numfeaturelevels: Default 3 (vs DINO's 4). returnintermindices is [1,2,3].
• train.enable_ema: Exponential moving average. Default False. Enable for potentially smoother convergence.
• dataset.remapmscococategory: Default False. Set True only for original MSCOCO dataset with 91-to-80 category ID remapping.

Multi-GPU / Multi-Node

Launch method: torchrun (LIGHTNINGEXCLUDEDNETWORK). The entrypoint runs torchrun --nnodes=N --nproc-per-node=M train.py, NOT plain python.

• When increasing train.numgpus, also set train.gpuids to the same

visible device range. For example, an 8-GPU single-node Slurm run must include both "train.numgpus": 8 and "train.gpuids": [0, 1, 2, 3, 4, 5, 6, 7].

• CUDAVISIBLEDEVICES is explicitly set (unlike Lightning-managed models which use TAOVISIBLEDEVICES)
• ddp with activation checkpointing: findunusedparameters=False
• ddp without: findunusedparameters=True
• fsdp supported, forces FP16

Multi-node env vars (set by orchestrator):

CRITICAL: NODE_RANK is copied to RANK if RANK is unset. This is required for torchrun multinode.

Export / TRT Defaults

• Export input: 640x640, opset 17
• TRT data types: FP32, FP16, INT8
• TRT workspace: 1024 MB
• TRT maxbatchsize: 4

Distillation

RT-DETR supports knowledge distillation with a teacher model. Requires distill action with distill.pretrainedteachermodel_path and a distillation binding configuration.

Use the packaged references/spectemplatedistill.yaml as the starting point. The validated default binding uses the RT-DETR distiller's explicit IOU feature path:

distill:
  bindings:
  - student_module_name: srcs
    teacher_module_name: srcs
    criterion: IOU
    weight: 1.0

Do not substitute DINO-style output names such as predlogits / predboxes, and do not bind arbitrary decoder heads unless you have verified the module returns captured feature lists. The RT-DETR distiller asserts that IOU bindings must use srcs or dsrcs.

Hardware

Minimum 1 GPU(s), recommended 2 GPU(s). 16GB+ (V100 or A100) VRAM per GPU. RT-DETR is more memory-efficient than DINO/GDINO due to smaller input size (640x640) and fewer feature levels. Trains well on single GPU for small-medium datasets.

Error Patterns

CUDA out of memory: Reduce batchsize. RT-DETR at 640x640 is lighter than DINO at 1333px, but batchsize > 8 may still OOM on 16GB GPUs.

numclasses mismatch: RT-DETR defaults to 80 (not 91 like DINO). Ensure dataset.numclasses matches your annotation categories.

CUDA index assert from category IDs: If COCO category IDs are one-based or otherwise not remapped to zero-based contiguous IDs, set dataset.numclasses to max(categoryid) + 1 and keep dataset.evalclassids aligned to the actual category IDs. For the packaged four-class S3 sample with IDs 1-4, use dataset.numclasses: 5 and dataset.evalclass_ids: [1, 2, 3, 4].

returnintermindices vs numfeaturelevels: Default is [1,2,3] with numfeaturelevels=3. Must be consistent if changed.

Export shape mismatch: Keep RT-DETR export and deploy consumer input size at the validated 640x640 default unless the model has been trained and checked for a different shape. The older packaged 960x544 template shape can fail during ONNX tracing with The size of tensor a (...) must match the size of tensor b (...) in hybrid_encoder.py positional embedding addition.

AutoML metric extraction: RT-DETR emits detection metrics in structured training status and logs. For COCO/paper-style benchmark comparisons, optimize valmAP with direction: maximize; for explicit AP50 workflows, optimize mAP50. Prefer resultsdir/train/status.json or AutoML result state before parsing raw logs. Do not optimize val_loss for default detection model invocations.

Checkpoint handoff: For evaluate/export/inference/quantize/distill/resume, use the checkpoint resolver on the best AutoML child job's resultsdir/train/ folder and select the action-appropriate modelepoch*.pth checkpoint. RT-DETR may also write a latest symlink, but that should only be used when a caller explicitly requests latest. Keep dataset.numclasses, dataset.evalclassids, model.numqueries, and model.numselect consistent with training.

Parent rtdetr gentrtengine rejected by the PyT CLI: In the validated 7.0.0 PyT container, rtdetr gentrtengine is not a valid parent-model subtask. Use the RT-DETR deploy workflow (references/tao-deploy-rtdetr.md) for TensorRT engine generation, TensorRT evaluation, and TensorRT inference.

Spec Param / Parent Model Inference

Model-specific inference mappings belong in this MD file, not in config.json. Generated runners should read this section and apply the mappings with SDK helpers before createjob(). This mirrors the old microservices inferparams.py flow.

Inference mappings from TAO Core rtdetr.config.json:

For parentmodel or parentmodelfolder, pass the upstream train/export/AutoML child job id as parentjob_id. The SDK lists the parent result folder, filters checkpoint artifacts, and returns the selected model file or folder. Do not add these mappings back to config.json and do not patch generated runner scripts to guess checkpoint paths.

Deployment

• [tao-deploy-rtdetr](references/tao-deploy-rtdetr.md)