Tao Train Rtdetr

NVIDIAOfficialCreated by NVIDIA Corporation

Skill
Tool integration
Data and AI
Apache-2.0
Medium risk

From the author

RT-DETR (Real-Time DEtection TRansformer) for 2D object detection. Designed for real-time inference with competitive accuracy and supports distillation and quantization for deployment optimization. Use when training, evaluating, distilling, quantizing, exporting, or running inference for a TAO RT-DETR model. Trigger phrases include "train RT-DETR", "real-time DETR", "low-latency object detection", "RT-DETR distillation / quantization".

— the description by NVIDIA Corporation

AutoGineer summary

Written by AutoGineer from the skill’s own files, not by the author

The RT-DETR skill provides a complete pipeline for the Real-Time Detection Transformer model. It supports the full lifecycle of model development, including training, knowledge distillation, quantization for deployment, evaluation, and export. Designed for low-latency performance, it integrates seamlessly with TAO Toolkit workflows and includes support for AutoML to streamline hyperparameter tuning.

What it does

Trains RT-DETR models for 2D object detection tasks.
Supports knowledge distillation to improve model efficiency.
Enables model quantization for optimized deployment.
Provides tools for model evaluation and inference.
Facilitates model export to ONNX format for further deployment.
Automates hyperparameter tuning via integrated AutoML policies.

When to use it

When you need a high-accuracy, low-latency object detection model.
When you are working within the NVIDIA TAO ecosystem.
When you need to optimize a model for production using distillation or quantization.
When you want to automate the training process using AutoML.

Advantages

Optimized for real-time inference performance.
Supports advanced model compression techniques like distillation and quantization.
Compatible with standard COCO dataset formats.
Flexible configuration for multi-GPU and multi-node training environments.

Usage examples

“Train an RT-DETR model on my custom dataset.”: The agent initiates the training process using the specified dataset paths and default RT-DETR configurations.
“Quantize my trained RT-DETR model for deployment.”: The agent applies quantization layers to the model checkpoint and prepares it for optimized inference.
“Run distillation for my RT-DETR model using a teacher model.”: The agent configures the distillation process, binding the student and teacher models to improve performance.
“Export my RT-DETR model to ONNX.”: The agent exports the trained model checkpoint into an ONNX file suitable for TensorRT engine generation.

Install

Download the skill as a ZIP:
https://autogineer.com/agent-skills/nvidia/tao-train-rtdetr/download

Unzip it into the cross-client skills folder of your project:

mkdir -p .agents/skills && unzip tao-train-rtdetr.zip -d .agents/skills

Documentation by NVIDIA Corporation

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.pretrained_backbone_path for backbone weights or train.pretrained_model_path for full model.

For TAO Deploy TensorRT actions (gen_trt_engine, TensorRT evaluate, and TensorRT inference), read references/tao-deploy-rtdetr.md first. Deploy spec templates live in this skill's references/ folder with the spec_template_deploy_*.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/spec_template_<action>.yaml from the schema top-level default field. AutoML enablement is declared at the model layer in references/skill_info.yaml via automl_enabled. Runnable AutoML for an action requires schemas/<action>.schema.json and references/spec_template_<action>.yaml to exist and parse. Use the packaged selected-action schema for automl_default_parameters, automl_disabled_parameters, 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/skill_info.yaml and resolve the run override from either an explicit automl_policy value or the user's workflow request. Use automl_policy: 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 automl_policy: off for this run only. When automl_policy: on, automl_enabled: true, and both schemas/train.schema.json and references/spec_template_train.yaml are packaged, route the train action through tao-skill-bank:tao-run-automl by default with this model's skill_dir. Preserve workflow/application overrides for datasets, specs, output directories, GPU/platform settings, parent checkpoints, and automl_policy. 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 default_specs. This model skill exposes train, distill, quantize, evaluate, export, and inference; resume/retrain is performed through train with train.resume_training_checkpoint_path.

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

Training Requirements

Dataset type: object_detection
Formats: coco, coco_raw
AutoML training metrics: val_mAP50 for quick operational checks or val_mAP for COCO/paper-style benchmark comparisons; both are maximized. These are the exact structured train-status KPI names.
Standalone evaluation metrics: test_mAP50 and test_mAP. Do not use the evaluator's test_* names for AutoML trial ranking.

Per-Action Dataset Requirements

Action	Spec Key	Source	Files	List?
distill	dataset.train_data_sources	train_datasets	image_dir: images.tar.gz, json_file: annotations.json	Yes
distill	dataset.val_data_sources	eval_dataset	image_dir: images.tar.gz, json_file: annotations.json	No
evaluate	dataset.test_data_sources	eval_dataset	image_dir: images.tar.gz, json_file: annotations.json	No
inference	dataset.infer_data_sources	inference_dataset	image_dir: images.tar.gz, classmap: label_map.txt	Yes
quantize	dataset.train_data_sources	train_datasets	image_dir: images.tar.gz, json_file: annotations.json	Yes
quantize	dataset.val_data_sources	eval_dataset	image_dir: images.tar.gz, json_file: annotations.json	No
quantize	dataset.quant_calibration_data_sources	train_datasets	image_dir: images.tar.gz, json_file: annotations.json	No
train	dataset.train_data_sources	train_datasets	image_dir: images.tar.gz, json_file: annotations.json	Yes
train	dataset.val_data_sources	eval_dataset	image_dir: images.tar.gz, json_file: annotations.json	No

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.train_spatial_size: Training input size. Default [640, 640]. Smaller than DINO's multi-scale (up to 1333). Key to RT-DETR's speed.
model.num_feature_levels: Default 3 (vs DINO's 4). return_interm_indices is [1,2,3].
train.enable_ema: Exponential moving average. Default False. Enable for potentially smoother convergence.
dataset.remap_mscoco_category: Default False. Set True only for original MSCOCO dataset with 91-to-80 category ID remapping.

Multi-GPU / Multi-Node

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

Spec Key	Description	Default
`train.num_gpus`	Number of GPUs per node	1
`train.gpu_ids`	GPU device indices	[0]
`train.num_nodes`	Number of nodes	1
`train.distributed_strategy`	`ddp` or `fsdp`	`ddp`

When increasing train.num_gpus, also set train.gpu_ids to the same visible device range. For example, an 8-GPU single-node Slurm run must include both "train.num_gpus": 8 and "train.gpu_ids": [0, 1, 2, 3, 4, 5, 6, 7].
CUDA_VISIBLE_DEVICES is explicitly set (unlike Lightning-managed models which use TAO_VISIBLE_DEVICES)
ddp with activation checkpointing: find_unused_parameters=False
ddp without: find_unused_parameters=True
fsdp supported, forces FP16

Multi-node env vars (set by orchestrator):

Variable	Purpose
`WORLD_SIZE`	Number of nodes (triggers multinode mode)
`NODE_RANK`	This node's rank (0-indexed)
`MASTER_ADDR`	Rank-0 node IP
`MASTER_PORT`	Rank-0 port (default 29500)
`NUM_GPU_PER_NODE`	GPUs per node (default: all visible)

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 max_batch_size: 4

Distillation

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

Use the packaged references/spec_template_distill.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 pred_logits / pred_boxes, 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 batch_size. RT-DETR at 640x640 is lighter than DINO at 1333px, but batch_size > 8 may still OOM on 16GB GPUs.

num_classes mismatch: RT-DETR defaults to 80 (not 91 like DINO). Ensure dataset.num_classes 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.num_classes to max(category_id) + 1 and keep dataset.eval_class_ids aligned to the actual category IDs. For the packaged four-class S3 sample with IDs 1-4, use dataset.num_classes: 5 and dataset.eval_class_ids: [1, 2, 3, 4].

return_interm_indices vs num_feature_levels: Default is [1,2,3] with num_feature_levels=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 val_mAP with direction: maximize; for explicit AP50 workflows, optimize val_mAP50. Standalone evaluation emits the corresponding test_mAP and test_mAP50 keys. Prefer results_dir/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 results_dir/train/ folder and select the action-appropriate model_epoch_*.pth checkpoint. RT-DETR may also write a latest symlink, but that should only be used when a caller explicitly requests latest. Keep dataset.num_classes, dataset.eval_class_ids, model.num_queries, and model.num_select consistent with training.

Parent rtdetr gen_trt_engine rejected by the PyT CLI: In the validated 7.0.0 PyT container, rtdetr gen_trt_engine 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 create_job(). This mirrors the old microservices infer_params.py flow.

Inference mappings from TAO Core rtdetr.config.json:

Action	Spec Field	Inference Function	Meaning
distill	`distill.pretrained_teacher_model_path`	`parent_model`	model file inferred from the parent job results folder
distill	`encryption_key`	`key`	encryption key
distill	`results_dir`	`output_dir`	current job results directory
evaluate	`encryption_key`	`key`	encryption key
evaluate	`evaluate.checkpoint`	`parent_model`	model file inferred from the parent job results folder
evaluate	`evaluate.trt_engine`	`parent_model`	model file inferred from the parent job results folder
evaluate	`results_dir`	`output_dir`	current job results directory
export	`encryption_key`	`key`	encryption key
export	`export.checkpoint`	`parent_model`	model file inferred from the parent job results folder
export	`export.onnx_file`	`create_onnx_file`	output ONNX path
export	`results_dir`	`output_dir`	current job results directory
inference	`encryption_key`	`key`	encryption key
inference	`inference.checkpoint`	`parent_model`	model file inferred from the parent job results folder
inference	`inference.trt_engine`	`parent_model`	model file inferred from the parent job results folder
inference	`results_dir`	`output_dir`	current job results directory
quantize	`encryption_key`	`key`	encryption key
quantize	`quantize.model_path`	`parent_model`	model file inferred from the parent job results folder
quantize	`results_dir`	`output_dir`	current job results directory
train	`encryption_key`	`key`	encryption key
train	`model.pretrained_backbone_path`	`ptm_if_no_resume_model`	PTM when no resume checkpoint exists
train	`results_dir`	`output_dir`	current job results directory
train	`train.pretrained_model_path`	`ptm_if_no_resume_model`	PTM when no resume checkpoint exists
train	`train.resume_training_checkpoint_path`	`resume_model`	model file inferred from the current job results folder

For parent_model or parent_model_folder, pass the upstream train/export/AutoML child job id as parent_job_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

Files

ATTRIBUTION.mdAdded by AutoGineer918 B
BENCHMARK.md6.9 KiB
LICENSELicence file added by AutoGineer11.2 KiB
SKILL.md17.4 KiB
config/skillspector-baseline.yaml834 B
evals/evals.json794 B
references/skill_info.yaml7.1 KiB
references/spec_template_deploy_evaluate.yaml508 B
references/spec_template_deploy_gen_trt_engine.yaml665 B
references/spec_template_deploy_inference.yaml562 B
references/spec_template_distill.yaml4.1 KiB
references/spec_template_evaluate.yaml3.1 KiB
references/spec_template_export.yaml3.2 KiB
references/spec_template_inference.yaml3.1 KiB
references/spec_template_quantize.yaml2.9 KiB
references/spec_template_train.yaml2.9 KiB
references/tao-deploy-rtdetr.md6.1 KiB
references/tao-deploy-rtdetr.skill_info.yaml3.2 KiB
schemas/distill.schema.json50.6 KiB
schemas/evaluate.schema.json54.1 KiB
schemas/export.schema.json54.7 KiB
schemas/inference.schema.json55.0 KiB
schemas/manifest.json17.9 KiB
schemas/quantize.schema.json50.6 KiB
schemas/train.schema.json50.5 KiB
skill-card.md4.1 KiB
skill.oms.sig9.4 KiB

schemas/distill.schema.json

{
  "automl_default_parameters": [
    "train.optim.lr_step_size",
    "model.enc_layers",
    "dataset.augmentation.distortion_prob",
    "model.dec_layers",
    "dataset.batch_size",
    "train.optim.weight_decay",
    "train.optim.lr_decay",
    "dataset.workers",
    "train.optim.momentum",
    "dataset.augmentation.iou_crop_prob",
    "train.ema.decay",
    "model.num_queries",
    "train.optim.lr",
    "train.optim.lr_backbone",
    "model.num_select"
  ],
  "automl_disabled_parameters": [
    "quantize",
    "train.cudnn",
    "gen_trt_engine.tensorrt.calibration.cal_image_dir",
    "quantize.backend_kwargs",
    "model.use_encoder_idx",
    "train.gpu_ids",
    "model.return_interm_indices",
    "dataset.train_data_sources",
    "wandb.tags",
    "dataset.eval_class_ids",
    "quantize.skip_names",
    "model.feat_strides",
    "inference.color_map",
    "evaluate",
    "inference",
    "train",
    "model.frozen_fm",
    "model.loss_types",
    "dataset.val_data_sources",
    "distill",
    "dataset.augmentation",
    "gen_trt_engine",
    "dataset.augmentation.train_spatial_size",
    "gen_trt_engine.tensorrt.layers_precision",
    "dataset",
    "gen_trt_engine.tensorrt",
    "quantize.layers",
    "dataset.augmentation.eval_spatial_size",
    "model.backbone_names",
    "model",
    "train.optim.lr_steps",
    "train.freeze",
    "evaluate.gpu_ids",
    "train.optim",
    "train.ema",
    "gen_trt_engine.tensorrt.calibration",
    "dataset.augmentation.multi_scales",
    "model.hidden_dim",
    "model.linear_proj_names",
    "export",
    "model.feat_channels",
    "wandb",
    "dataset.infer_data_sources",
    "inference.gpu_ids",
    "dataset.test_data_sources",
    "dataset.quant_calibration_data_sources"
  ],
  "default": {
    "dataset": {
      "augmentation": {
        "distortion_prob": 0.8,
        "eval_spatial_size": [
          640,
          640
        ],
        "iou_crop_prob": 0.8,
        "multi_scales": [
          480,
          512,
          544,
          576,
          608,
          640,
          672,
          704,
          736,
          768,
          800
        ],
        "preserve_aspect_ratio": false,
        "train_spatial_size": [
          640,
          640
        ]
      },
      "batch_size": 4,
      "dataset_type": "serialized",
      "eval_class_ids": [
        1
      ],
      "infer_data_sources": {
        "classmap": "",
        "image_dir": [
          ""
        ]
      },
      "num_classes": 80,
      "pin_memory": true,
      "quant_calibration_data_sources": {
        "image_dir": "",
        "json_file": ""
      },
      "remap_mscoco_category": false,
      "test_data_sources": {
        "image_dir": "",
        "json_file": ""
      },
      "train_data_sources": [
        {
          "image_dir": "",
          "json_file": ""
        }
      ],
      "val_data_sources": {
        "image_dir": "",
        "json_file": ""
      },
      "workers": 8
    },
    "encryption_key": "",
    "model": {
      "act": "silu",
      "alpha": 0.75,
      "aux_loss": true,
      "backbone": "resnet_50",
      "backbone_names": [
        "backbone.0"
      ],
      "bbox_cost": 5.0,
      "bbox_loss_coef": 5.0,
      "class_cost": 2.0,
      "clip_max_norm": 0.1,
      "dec_layers": 6,
      "depth_mult": 1,
      "dim_feedforward": 1024,
      "distillation_loss_coef": 1.0,
      "dn_number": 100,
      "dropout_ratio": 0.0,
      "enc_act": "gelu",
      "enc_layers": 1,
      "eval_idx": -1,
      "expansion": 1,
      "feat_channels": [
        256,
        256,
        256
      ],
      "feat_strides": [
        8,
        16,
        32
      ],
      "frozen_fm": {
        "backbone": "radio_v2-l",
        "checkpoint": "",
        "enabled": false
      },
      "gamma": 2.0,
      "giou_cost": 2.0,
      "giou_loss_coef": 2.0,
      "hidden_dim": 256,
      "linear_proj_names": [
        "reference_points",
        "sampling_offsets"
      ],
      "load_teacher_enc_dec": false,
      "loss_types": [
        "vfl",
        "boxes"
      ],
      "nheads": 8,
      "num_feature_levels": 3,
      "num_queries": 300,
      "num_select": 300,
      "pe_temperature": 10000,
      "pretrained_backbone_path": "",
      "return_interm_indices": [
        1,
        2,
        3
      ],
      "train_backbone": true,
      "use_encoder_idx": [
        2
      ],
      "vfl_loss_coef": 1.0
    },
    "model_name": "",
    "quantize": {
      "algorithm": "minmax",
      "backend": "torchao",
      "backend_kwargs": {},
      "device": "cuda",
      "layers": [],
      "mode": "weight_only_ptq",
      "model_path": "",
      "results_dir": "",
      "skip_names": []
    },
    "results_dir": "",
    "train": {
      "activation_checkpoint": true,
      "checkpoint_interval": 1,
      "checkpoint_interval_unit": "epoch",
      "clip_grad_norm": 0.1,
      "cudnn": {
        "benchmark": false,
        "deterministic": true
      },
      "distributed_strategy": "ddp",
      "ema": {
        "cpu_offload": false,
        "decay": 0.999,
        "every_n_steps": 1,
        "validate_original_weights": false
      },
      "enable_ema": false,
      "freeze": [],
      "gpu_ids": [
        0
      ],
      "is_dry_run": false,
      "num_epochs": 10,
      "num_gpus": 1,
      "num_nodes": 1,
      "optim": {
        "lr": 0.0001,
        "lr_backbone": 1e-05,
        "lr_decay": 0.1,
        "lr_scheduler": "MultiStep",
        "lr_step_size": 1000,
        "lr_steps": [
          1000
        ],
        "momentum": 0.9,
        "monitor_name": "val_loss",
        "optimizer": "AdamW",
        "warmup_steps": 0,
        "weight_decay": 0.0001
      },
      "precision": "fp32",
      "pretrained_model_path": "",
      "results_dir": "",
      "resume_training_checkpoint_path": "",
      "seed": 1234,
      "validation_interval": 1
    },
    "wandb": {
      "enable": true,
      "entity": "",
      "group": "",
      "name": "TAO Toolkit Training",
      "project": "TAO Toolkit",
      "reinit": false,
      "run_id": "",
      "save_code": false,
      "sync_tensorboard": false,
      "tags": [
        "tao-toolkit"
      ]
    }
  },
  "popular": {
    "evaluate": {
      "gpu_ids": [
        0
      ],
      "num_gpus": 1,
      "num_nodes": 1
    },
    "gen_trt_engine": {
      "batch_size": -1,
      "gpu_id": 0,
      "tensorrt": {
        "calibration": {
          "cal_batch_size": 1,
          "cal_batches": 1
        },
        "min_batch_size": 1,
        "opt_batch_size": 1
      }
    },
    "inference": {
      "gpu_ids": [
        0
      ],
      "num_gpus": 1,
      "num_nodes": 1
    },
    "train": {
      "checkpoint_interval": 1,
      "gpu_ids": [
        0
      ],
      "num_epochs": 10,
      "num_gpus": 1,
      "num_nodes": 1,
      "validation_interval": 1
    }
  },
  "properties": {
    "automl_disabled_parameters": [
      "wandb",
      "model",
      "dataset",
      "train",
      "evaluate",
      "inference",
      "export",
      "gen_trt_engine",
      "distill",
      "quantize"
    ],
    "dataset": {
      "automl_default_parameters": [
        "dataset.batch_size",
        "dataset.workers"
      ],
      "automl_disabled_parameters": [
        "dataset.train_data_sources",
        "dataset.val_data_sources",
        "dataset.test_data_sources",
        "dataset.infer_data_sources",
        "dataset.quant_calibration_data_sources",
        "dataset.eval_class_ids",
        "dataset.augmentation"
      ],
      "automl_enabled": false,
      "default": {
        "augmentation": {
          "distortion_prob": 0.8,
          "eval_spatial_size": [
            640,
            640
          ],
          "iou_crop_prob": 0.8,
          "multi_scales": [
            480,
            512,
            544,
            576,
            608,
            640,
            672,
            704,
            736,
            768,
            800
          ],
          "preserve_aspect_ratio": false,
          "train_spatial_size": [
            640,
            640
          ]
        },
        "batch_size": 4,
        "dataset_type": "serialized",
        "eval_class_ids": [
          1
        ],
        "infer_data_sources": {
          "classmap": "",
          "image_dir": [
            ""
          ]
        },
        "num_classes": 80,
        "pin_memory": true,
        "quant_calibration_data_sources": {
          "image_dir": "",
          "json_file": ""
        },
        "remap_mscoco_category": false,
        "test_data_sources": {
          "image_dir": "",
          "json_file": ""
        },
        "train_data_sources": [
          {
            "image_dir": "",
            "json_file": ""
          }
        ],
        "val_data_sources": {
          "image_dir": "",
          "json_file": ""
        },
        "workers": 8
      },
      "description": "Configurable parameters to construct the dataset for a RT-DETR experiment.",
      "properties": {
        "augmentation": {
          "automl_default_parameters": [
            "dataset.augmentation.distortion_prob",
            "dataset.augmentation.iou_crop_prob"
          ],
          "automl_disabled_parameters": [
            "dataset.augmentation.multi_scales",
            "dataset.augmentation.train_spatial_size",
            "dataset.augmentation.eval_spatial_size"
          ],
          "automl_enabled": false,
          "default": {
            "distortion_prob": 0.8,
            "eval_spatial_size": [
              640,
              640
            ],
            "iou_crop_prob": 0.8,
            "multi_scales": [
              480,
              512,
              544,
              576,
              608,
              640,
              672,
              704,
              736,
              768,
              800
            ],
            "preserve_aspect_ratio": false,
            "train_spatial_size": [
              640,
              640
            ]
          },
          "description": "Configuration parameters for data augmentation",
          "properties": {
            "distortion_prob": {
              "automl_enabled": true,
              "default": 0.8,
              "description": "The probability for RandomPhotometricDistort",
              "maximum": 1.0,
              "minimum": 0.0,
              "title": "distortion probability",
              "type": "float"
            },
            "eval_spatial_size": {
              "automl_enabled": false,
              "default": [
                640,
                640
              ],
              "description": "Input resolution to run evaluation during validation and testing. This is in the [h, w] order.",
              "title": "evaluation spatial size",
              "type": "list"
            },
            "iou_crop_prob": {
              "automl_enabled": true,
              "default": 0.8,
              "description": "The probability for RandomIoUCrop",
              "maximum": 1.0,
              "minimum": 0.0,
              "title": "iou crop probability",
              "type": "float"
            },
            "multi_scales": {
              "automl_enabled": false,
              "default": [
                480,
                512,
                544,
                576,
                608,
                640,
                672,
                704,
                736,
                768,
                800
              ],
              "description": "A list of sizes to perform random resize.",
              "title": "multi-scales",
              "type": "list"
            },
            "preserve_aspect_ratio": {
              "default": false,
              "description": "Flag to enable resize with preserving the aspect ratio.",
              "title": "preserve aspect ratio",
              "type": "bool"
            },
            "train_spatial_size": {
              "automl_enabled": false,
              "default": [
                640,
                640
              ],
              "description": "Input resolution to run evaluation during training. This is in the [h, w] order.",
              "title": "train spatial size",
              "type": "list"
            }
          },
          "title": "augmentation",
          "type": "collection"
        },
        "batch_size": {
          "automl_enabled": true,
          "default": 4,
          "description": "The batch size for training and validation",
          "maximum": Infinity,
          "minimum": 1,
          "title": "batch size",
          "type": "int"
        },
        "dataset_type": {
          "default": "serialized",
          "description": "If set to default, we follow the standard CocoDetection` dataset structure\n                    from the torchvision which loads COCO annotation in every subprocess. This leads to redudant\n                    copy of data and can cause RAM to explod if workers` is high. If set to serialized,\n                    the data is serialized through pickle and torch.Tensor` that allows the data to be shared\n                    across subprocess. As a result, RAM usage can be greatly improved.",
          "enum": [
            "serialized",
            "default"
          ],
          "title": "dataset type",
          "type": "categorical"
        },
        "eval_class_ids": {
          "automl_enabled": false,
          "default": [
            1
          ],
          "description": "IDs of the classes for evaluation.",
          "title": "eval class ids",
          "type": "list"
        },
        "infer_data_sources": {
          "automl_enabled": false,
          "default": {
            "classmap": "",
            "image_dir": [
              ""
            ]
          },
          "description": "The data source for inference:\n                    * image_dir : The list of directories that contains the inference images\n                    * classmap : The path of the .txt file that contains class names",
          "title": "infer data sources",
          "type": "collection"
        },
        "num_classes": {
          "default": 80,
          "description": "The number of classes in the training data",
          "math_cond": ">0",
          "maximum": Infinity,
          "minimum": 1,
          "title": "num classes",
          "type": "int"
        },
        "pin_memory": {
          "default": true,
          "description": "Flag to enable the dataloader to allocated pagelocked memory for faster\n                    of data between the CPU and GPU.",
          "title": "pin_memory",
          "type": "bool"
        },
        "quant_calibration_data_sources": {
          "automl_enabled": false,
          "default": {
            "image_dir": "",
            "json_file": ""
          },
          "description": "The data source for quantization calibration:\n                    * image_dir : The directory that contains the quantization calibration images\n                    * json_file(optional) : The path of the JSON file, which uses quantization calibration-                        annotation COCO format",
          "title": "quantization calibration data sources",
          "type": "collection"
        },
        "remap_mscoco_category": {
          "default": false,
          "description": "Flag to enable mapping of MSCOCO 91 classes to 80. Only required if we're directly\n                    training using the original COCO annotation files.\n                    For custom dataset, this value needs to be set False",
          "title": "remap mscoco category",
          "type": "bool"
        },
        "test_data_sources": {
          "automl_enabled": false,
          "default": {
            "image_dir": "",
            "json_file": ""
          },
          "description": "The data source for testing:\n                    * image_dir : The directory that contains the test images\n                    * json_file : The path of the JSON file, which uses test-annotation COCO format",
          "title": "test data sources",
          "type": "collection"
        },
        "train_data_sources": {
          "automl_enabled": false,
          "default": [
            {
              "image_dir": "",
              "json_file": ""
            }
          ],
          "description": "The list of data sources for training:\n                    * image_dir : The directory that contains the training images\n                    * json_file : The path of the JSON file, which uses training-annotation COCO format",
          "title": "train data sources",
          "type": "list"
        },
        "val_data_sources": {
          "automl_enabled": false,
          "default": {
            "image_dir": "",
            "json_file": ""
          },
          "description": "The list of data sources for validation:\n                    * image_dir : The directory that contains the validation images\n                    * json_file : The path of the JSON file, which uses validation-annotation COCO format",
          "title": "validation data sources",
          "type": "collection"
        },
        "workers": {
          "automl_enabled": true,
          "default": 8,
          "description": "The number of parallel workers processing data",
          "maximum": Infinity,
          "minimum": 1,
          "title": "batch size",
          "type": "int"
        }
      },
      "type": "collection"
    },
    "distill": {
      "automl_enabled": false,
      "description": "Configurable parameters to construct the distiller for a RT-DETR experiment.",
      "type": "collection"
    },
    "encryption_key": {
      "default": "",
      "description": "Key for encrypting model checkpoints",
      "title": "Encryption key",
      "type": "string"
    },
    "model": {
      "automl_default_parameters": [
        "model.num_queries",
        "model.num_select",
        "model.enc_layers",
        "model.dec_layers"
      ],
      "automl_disabled_parameters": [
        "model.return_interm_indices",
        "model.feat_strides",
        "model.feat_channels",
        "model.use_encoder_idx",
        "model.hidden_dim",
        "model.loss_types",
        "model.backbone_names",
        "model.linear_proj_names",
        "model.frozen_fm"
      ],
      "automl_enabled": false,
      "default": {
        "act": "silu",
        "alpha": 0.75,
        "aux_loss": true,
        "backbone": "resnet_50",
        "backbone_names": [
          "backbone.0"
        ],
        "bbox_cost": 5.0,
        "bbox_loss_coef": 5.0,
        "class_cost": 2.0,
        "clip_max_norm": 0.1,
        "dec_layers": 6,
        "depth_mult": 1,
        "dim_feedforward": 1024,
        "distillation_loss_coef": 1.0,
        "dn_number": 100,
        "dropout_ratio": 0.0,
        "enc_act": "gelu",
        "enc_layers": 1,
        "eval_idx": -1,
        "expansion": 1,
        "feat_channels": [
          256,
          256,
          256
        ],
        "feat_strides": [
          8,
          16,
          32
        ],
        "frozen_fm": {
          "backbone": "radio_v2-l",
          "checkpoint": "",
          "enabled": false
        },
        "gamma": 2.0,
        "giou_cost": 2.0,
        "giou_loss_coef": 2.0,
        "hidden_dim": 256,
        "linear_proj_names": [
          "reference_points",
          "sampling_offsets"
        ],
        "load_teacher_enc_dec": false,
        "loss_types": [
          "vfl",
          "boxes"
        ],
        "nheads": 8,
        "num_feature_levels": 3,
        "num_queries": 300,
        "num_select": 300,
        "pe_temperature": 10000,
        "pretrained_backbone_path": "",
        "return_interm_indices": [
          1,
          2,
          3
        ],
        "train_backbone": true,
        "use_encoder_idx": [
          2
        ],
        "vfl_loss_coef": 1.0
      },
      "description": "Configurable parameters to construct the model for a RT-DETR experiment.",
      "properties": {
        "act": {
          "default": "silu",
          "description": "The activation used for top-down FPN and bottom-up PAN.",
          "title": "activation",
          "type": "string"
        },
        "alpha": {
          "default": 0.75,
          "description": "The alpha value in the varifocal loss.",
          "math_cond": "> 0.0",
          "title": "alpha",
          "type": "float"
        },
        "aux_loss": {
          "default": true,
          "description": "A flag specifying whether to use auxiliary\n                    decoding losses (loss at each decoder layer)",
          "title": "Auxiliary Loss",
          "type": "bool"
        },
        "backbone": {
          "default": "resnet_50",
          "description": "The backbone name of the model.\n                    TAO implementation of RT-DETR support ResNet, EfficientViT, FAN, and ConvNext.",
          "enum": [
            "resnet_18",
            "resnet_34",
            "resnet_50",
            "resnet_101",
            "convnext_tiny",
            "convnext_small",
            "convnext_base",
            "convnext_large",
            "convnext_xlarge",
            "fan_tiny",
            "fan_small",
            "fan_base",
            "fan_large",
            "efficientvit_b0",
            "efficientvit_b1",
            "efficientvit_b2",
            "efficientvit_b3",
            "efficientvit_l0",
            "efficientvit_l1",
            "efficientvit_l2",
            "efficientvit_l3"
          ],
          "title": "backbone",
          "type": "categorical"
        },
        "backbone_names": {
          "automl_enabled": false,
          "default": [
            "backbone.0"
          ],
          "description": "Prefix of the tensor names corresponding to the backbone.",
          "title": "Backbone tensor name prefix",
          "type": "list"
        },
        "bbox_cost": {
          "default": 5.0,
          "description": "The relative weight of the L1 error of the bounding box coordinates in the matching cost.",
          "maximum": Infinity,
          "minimum": 0.0,
          "title": "BBox cost coefficient",
          "type": "float"
        },
        "bbox_loss_coef": {
          "default": 5.0,
          "description": "The relative weight of the L1 error of the bounding box coordinates in the loss function.",
          "maximum": Infinity,
          "minimum": 0.0,
          "title": "BBox loss coefficient",
          "type": "float"
        },
        "class_cost": {
          "default": 2.0,
          "description": "The relative weight of the classification error in the matching cost.",
          "maximum": Infinity,
          "minimum": 0.0,
          "title": "Class cost coefficient",
          "type": "float"
        },
        "clip_max_norm": {
          "default": 0.1,
          "title": "clip max norm",
          "type": "float"
        },
        "dec_layers": {
          "automl_enabled": true,
          "default": 6,
          "description": "Numer of decoder layers in the transformer",
          "minimum": 1,
          "title": "decoder layers",
          "type": "int"
        },
        "depth_mult": {
          "default": 1,
          "description": "The number of RegVGGBlock used in CSPRepLayer.",
          "maximum": Infinity,
          "minimum": 1,
          "title": "expansion",
          "type": "int"
        },
        "dim_feedforward": {
          "default": 1024,
          "description": "Dimension of the feedforward network",
          "minimum": 1,
          "title": "dim feedforward",
          "type": "int"
        },
        "distillation_loss_coef": {
          "default": 1.0,
          "description": "The coefficient for the distillation loss during distill.",
          "minimum": 0.0,
          "title": "distillation loss coefficient",
          "type": "float"
        },
        "dn_number": {
          "default": 100,
          "description": "The number of denoising queries.",
          "maximum": Infinity,
          "minimum": 0,
          "title": "denoising number",
          "type": "int"
        },
        "dropout_ratio": {
          "default": 0.0,
          "description": "The probability to drop hidden units.",
          "maximum": 1.0,
          "minimum": 0.0,
          "title": "drop out ratio",
          "type": "float"
        },
        "enc_act": {
          "default": "gelu",
          "description": "The activation used for the encoder.",
          "title": "encoder activation",
          "type": "string"
        },
        "enc_layers": {
          "automl_enabled": true,
          "default": 1,
          "description": "Numer of encoder layers in the transformer",
          "minimum": 1,
          "title": "encoder layers",
          "type": "int"
        },
        "eval_idx": {
          "default": -1,
          "description": "The index of decoder layer to use for evaluation. By default, use the last decoder layer.",
          "maximum": Infinity,
          "minimum": -1,
          "title": "evaluation index",
          "type": "int"
        },
        "expansion": {
          "default": 1,
          "description": "The expansion raito for hidden dimesnion used in CSPRepLayer.",
          "maximum": Infinity,
          "minimum": 0,
          "title": "expansion",
          "type": "int"
        },
        "feat_channels": {
          "automl_enabled": false,
          "default": [
            256,
            256,
            256
          ],
          "description": "The feature channel sizes in decoder.",
          "title": "feature channels",
          "type": "list"
        },
        "feat_strides": {
          "automl_enabled": false,
          "default": [
            8,
            16,
            32
          ],
          "description": "The stride used as grid size of positional embedding at each encoder layer.",
          "title": "feature strides",
          "type": "list"
        },
        "frozen_fm": {
          "automl_enabled": false,
          "default": {
            "backbone": "radio_v2-l",
            "checkpoint": "",
            "enabled": false
          },
          "description": "Configurable parameters to construct the frozen foundation model.",
          "properties": {
            "backbone": {
              "default": "radio_v2-l",
              "description": "Name of the frozen foundation model.",
              "enum": [
                "radio_v2-b",
                "radio_v2-l",
                "radio_v2-h"
              ],
              "title": "Name of the frozen foundation model",
              "type": "categorical"
            },
            "checkpoint": {
              "default": "",
              "description": "Path to a pretrained foundation model.",
              "title": "Pretrained foundation model path or name",
              "type": "string"
            },
            "enabled": {
              "default": false,
              "description": "Flag to enable frozen foundation model to be added to RT-DETR.",
              "title": "Enable frozen FM",
              "type": "bool"
            }
          },
          "type": "collection"
        },
        "gamma": {
          "default": 2.0,
          "description": "The gamma value in the varifocal loss.",
          "math_cond": "> 0.0",
          "title": "gamma",
          "type": "float"
        },
        "giou_cost": {
          "default": 2.0,
          "description": "The relative weight of the GIoU loss of the bounding box in the matching cost.",
          "maximum": Infinity,
          "minimum": 0.0,
          "title": "GIoU cost coefficient",
          "type": "float"
        },
        "giou_loss_coef": {
          "default": 2.0,
          "description": "The relative weight of the GIoU loss of the bounding box in the loss function.",
          "maximum": Infinity,
          "minimum": 0.0,
          "title": "GIoU loss coefficient",
          "type": "float"
        },
        "hidden_dim": {
          "automl_enabled": false,
          "default": 256,
          "description": "Dimension of the hidden units.",
          "type": "int"
        },
        "linear_proj_names": {
          "automl_enabled": false,
          "default": [
            "reference_points",
            "sampling_offsets"
          ],
          "description": "Linear projection layer names.",
          "title": "linear projection names",
          "type": "list"
        },
        "load_teacher_enc_dec": {
          "default": false,
          "description": "Flag to load teacher's encoder and decoder weights.",
          "title": "Load teacher's encoder and decoder weights",
          "type": "bool"
        },
        "loss_types": {
          "automl_enabled": false,
          "default": [
            "vfl",
            "boxes"
          ],
          "description": "Losses to be used during training",
          "title": "loss_types",
          "type": "list"
        },
        "nheads": {
          "default": 8,
          "description": "Number of heads",
          "title": "nheads",
          "type": "int"
        },
        "num_feature_levels": {
          "default": 3,
          "description": "The number of feature levels to use in the model",
          "maximum": 4,
          "minimum": 1,
          "title": "number of feature levels",
          "type": "int"
        },
        "num_queries": {
          "automl_enabled": true,
          "default": 300,
          "description": "The number of queries",
          "maximum": Infinity,
          "minimum": 1,
          "parent_param": "TRUE",
          "title": "number of queries",
          "type": "int"
        },
        "num_select": {
          "automl_enabled": true,
          "default": 300,
          "depends_on": "model.num_queries",
          "description": "The number of top-K predictions selected during post-process. Must be < num_queries * num_classes",
          "maximum": 1000,
          "minimum": 1,
          "title": "num select",
          "type": "int"
        },
        "pe_temperature": {
          "default": 10000,
          "description": "The temperature applied to the positional sine embedding.",
          "maximum": Infinity,
          "minimum": 1,
          "title": "pe_temperature",
          "type": "int"
        },
        "pretrained_backbone_path": {
          "default": "",
          "description": "[Optional] Path to a pretrained backbone file.",
          "title": "pretrained backbone path",
          "type": "string"
        },
        "return_interm_indices": {
          "automl_enabled": false,
          "default": [
            1,
            2,
            3
          ],
          "description": "The index of feature levels to use in the model. The length must match `num_feature_levels`.",
          "title": "return interim indices",
          "type": "list"
        },
        "train_backbone": {
          "default": true,
          "description": "Flag to set backbone weights as trainable or frozen.\n                    When set to `False`, the backbone weights will be frozen.",
          "title": "Train backbone",
          "type": "bool"
        },
        "use_encoder_idx": {
          "automl_enabled": false,
          "default": [
            2
          ],
          "description": "The index of multi-scale backbone features to pass to encoder.",
          "title": "use encoder index",
          "type": "list"
        },
        "vfl_loss_coef": {
          "default": 1.0,
          "description": "The relative weight of the varifocal error in the loss function.",
          "maximum": Infinity,
          "minimum": 0.0,
          "title": "varifocal loss coefficient",
          "type": "float"
        }
      },
      "type": "collection"
    },
    "model_name": {
      "default": "",
      "description": "Name of model if invoking task via :code:`model_agnostic`",
      "title": "Model name",
      "type": "string"
    },
    "quantize": {
      "automl_disabled_parameters": [
        "quantize.layers",
        "quantize.skip_names",
        "quantize.backend_kwargs"
      ],
      "automl_enabled": false,
      "default": {
        "algorithm": "minmax",
        "backend": "torchao",
        "backend_kwargs": {},
        "device": "cuda",
        "layers": [],
        "mode": "weight_only_ptq",
        "model_path": "",
        "results_dir": "",
        "skip_names": []
      },
      "description": "Configurable parameters to run model quantization for a RT-DETR experiment.",
      "properties": {
        "algorithm": {
          "default": "minmax",
          "description": "Calibration/optimization algorithm. Used by ModelOpt backends (modelopt.pytorch and modelopt.onnx). Ignored by torchao backend.",
          "enum": [
            "minmax",
            "max",
            "entropy",
            "awq_clip",
            "awq_lite",
            "awq_full",
            "rtn_dq"
          ],
          "title": "Calibration algorithm",
          "type": "categorical"
        },
        "backend": {
          "default": "torchao",
          "description": "The quantization backend to use",
          "enum": [
            "modelopt.pytorch",
            "torchao",
            "modelopt.onnx"
          ],
          "title": "Quantization backend",
          "type": "categorical"
        },
        "backend_kwargs": {
          "automl_enabled": false,
          "description": "Additional keyword arguments to pass to the backend",
          "title": "Backend kwargs",
          "type": "collection"
        },
        "device": {
          "default": "cuda",
          "description": "Device to use for calibration. Accepts 'cuda' (uses default GPU), 'cpu', 'trt' (TensorRT for ONNX backend), or specific GPU device like 'cuda:0', 'cuda:1', etc. If 'cuda' is specified but no GPU is available, the framework will automatically fall back to 'cpu'. Note: 'trt' is only supported by the modelopt.onnx backend for ONNX Runtime with TensorRT execution provider.",
          "pattern": "^(cuda|cpu|trt|cuda:[0-9]+)$",
          "title": "Calibration device",
          "type": "string"
        },
        "layers": {
          "automl_enabled": false,
          "default": [],
          "description": "List of per-module quantization configurations. Each entry specifies which modules to quantize and their data types. This is the primary way to configure quantization.",
          "title": "Layer quantization configs",
          "type": "list"
        },
        "mode": {
          "default": "weight_only_ptq",
          "description": "The quantization mode to use",
          "enum": [
            "static_ptq",
            "weight_only_ptq"
          ],
          "title": "Quantization mode",
          "type": "categorical"
        },
        "model_path": {
          "default": "",
          "description": "Path to the model to be quantized. For ONNX backend, path to ONNX file.",
          "title": "Model path",
          "type": "string"
        },
        "results_dir": {
          "default": "",
          "description": "Path to where all the assets generated from a task are stored.",
          "title": "Results directory",
          "type": "string"
        },
        "skip_names": {
          "automl_enabled": false,
          "default": [],
          "description": "List of module/layer names or patterns to exclude from quantization",
          "title": "Skip names",
          "type": "list"
        }
      },
      "required": [
        "model_path",
        "results_dir"
      ],
      "type": "collection"
    },
    "results_dir": {
      "default": "",
      "description": "\n        Path to where all the assets generated from a task are stored.\n        ",
      "title": "Results directory",
      "type": "string"
    },
    "train": {
      "automl_disabled_parameters": [
        "train.gpu_ids",
        "train.cudnn",
        "train.freeze",
        "train.ema",
        "train.optim"
      ],
      "automl_enabled": false,
      "default": {
        "activation_checkpoint": true,
        "checkpoint_interval": 1,
        "checkpoint_interval_unit": "epoch",
        "clip_grad_norm": 0.1,
        "cudnn": {
          "benchmark": false,
          "deterministic": true
        },
        "distributed_strategy": "ddp",
        "ema": {
          "cpu_offload": false,
          "decay": 0.999,
          "every_n_steps": 1,
          "validate_original_weights": false
        },
        "enable_ema": false,
        "freeze": [],
        "gpu_ids": [
          0
        ],
        "is_dry_run": false,
        "num_epochs": 10,
        "num_gpus": 1,
        "num_nodes": 1,
        "optim": {
          "lr": 0.0001,
          "lr_backbone": 1e-05,
          "lr_decay": 0.1,
          "lr_scheduler": "MultiStep",
          "lr_step_size": 1000,
          "lr_steps": [
            1000
          ],
          "momentum": 0.9,
          "monitor_name": "val_loss",
          "optimizer": "AdamW",
          "warmup_steps": 0,
          "weight_decay": 0.0001
        },
        "precision": "fp32",
        "pretrained_model_path": "",
        "results_dir": "",
        "resume_training_checkpoint_path": "",
        "seed": 1234,
        "validation_interval": 1
      },
      "description": "Configurable parameters to construct the trainer for a RT-DETR experiment.",
      "popular": [
        "validation_interval",
        "checkpoint_interval",
        "num_gpus",
        "num_nodes",
        "num_epochs",
        "gpu_ids"
      ],
      "properties": {
        "activation_checkpoint": {
          "default": true,
          "description": "\n        A True value instructs train to recompute in backward pass to save GPU memory,\n        rather than storing activations.",
          "title": "enable activation checkpointing",
          "type": "bool"
        },
        "checkpoint_interval": {
          "default": 1,
          "description": "The interval (in epochs) at which a checkpoint will be saved. Helps resume training.",
          "minimum": 1,
          "popular": true,
          "title": "Checkpoint interval",
          "type": "int"
        },
        "checkpoint_interval_unit": {
          "default": "epoch",
          "description": "The unit of the checkpoint interval.",
          "enum": [
            "epoch",
            "step"
          ],
          "title": "Checkpoint interval unit",
          "type": "categorical"
        },
        "clip_grad_norm": {
          "default": 0.1,
          "description": "\n        Amount to clip the gradient by L2 Norm.\n        A value of 0.0 specifies no clipping.",
          "math_cond": "> 0.0",
          "title": "clip gradient norm",
          "type": "float"
        },
        "cudnn": {
          "automl_enabled": false,
          "default": {
            "benchmark": false,
            "deterministic": true
          },
          "properties": {
            "benchmark": {
              "default": false,
              "type": "bool"
            },
            "deterministic": {
              "default": true,
              "type": "bool"
            }
          },
          "type": "collection"
        },
        "distributed_strategy": {
          "default": "ddp",
          "description": "\n        The multi-GPU training strategy.\n        DDP (Distributed Data Parallel) and Fully Sharded DDP are supported.",
          "enum": [
            "ddp",
            "fsdp"
          ],
          "title": "distributed_strategy",
          "type": "categorical"
        },
        "ema": {
          "automl_default_parameters": [
            "train.ema.decay"
          ],
          "automl_enabled": false,
          "default": {
            "cpu_offload": false,
            "decay": 0.999,
            "every_n_steps": 1,
            "validate_original_weights": false
          },
          "description": "Hyper parameters to configure the Exponential Moving Average.",
          "properties": {
            "cpu_offload": {
              "default": false,
              "description": "Offload EMA calculation to CPU. Note that this will significantly slow down training.",
              "title": "cpu offload",
              "type": "bool"
            },
            "decay": {
              "automl_enabled": true,
              "default": 0.999,
              "description": "The decreasing factor for the exponential moving average.",
              "math_cond": "> 0.0",
              "maximum": 1.0,
              "minimum": 0.0,
              "title": "ema decay",
              "type": "float"
            },
            "every_n_steps": {
              "default": 1,
              "description": "The number of steps to perform exponential moving average.",
              "maximum": Infinity,
              "minimum": 1,
              "title": "every_n_steps",
              "type": "int"
            },
            "validate_original_weights": {
              "default": false,
              "description": "Whether to run evaluation using the non-EMA weight.",
              "title": "validate original weights",
              "type": "bool"
            }
          },
          "title": "ema",
          "type": "collection"
        },
        "enable_ema": {
          "default": false,
          "description": "Whether to enable Exponential Moving Average during training.",
          "title": "enable ema",
          "type": "bool"
        },
        "freeze": {
          "automl_enabled": false,
          "default": [],
          "description": "\n        List of layer names to freeze.\n        Example: [\"backbone\", \"encoder\", \"decoder\"].",
          "title": "freeze",
          "type": "list"
        },
        "gpu_ids": {
          "automl_enabled": false,
          "default": [
            0
          ],
          "description": "\n        List of GPU IDs to run the training on. The length of this list\n        must be equal to the number of gpus in train.num_gpus.",
          "popular": true,
          "title": "GPU IDs",
          "type": "list"
        },
        "is_dry_run": {
          "default": false,
          "description": "\n        Whether to run the trainer in Dry Run mode. This serves\n        as a good means to validate the spec file and run a sanity check on the trainer\n        without actually initializing and running the trainer.",
          "title": "Is dry run",
          "type": "bool"
        },
        "num_epochs": {
          "default": 10,
          "description": "Number of epochs to run the training.",
          "maximum": Infinity,
          "minimum": 1,
          "popular": true,
          "title": "Number of epochs",
          "type": "int"
        },
        "num_gpus": {
          "default": 1,
          "description": "The number of GPUs to run the train job.",
          "minimum": 1,
          "popular": true,
          "title": "Number of GPUs",
          "type": "int"
        },
        "num_nodes": {
          "default": 1,
          "description": "Number of nodes to run the training on. If > 1, then multi-node is enabled.",
          "minimum": 1,
          "popular": true,
          "title": "Number of nodes",
          "type": "int"
        },
        "optim": {
          "automl_default_parameters": [
            "train.optim.lr",
            "train.optim.lr_backbone",
            "train.optim.momentum",
            "train.optim.weight_decay",
            "train.optim.lr_step_size",
            "train.optim.lr_decay"
          ],
          "automl_disabled_parameters": [
            "train.optim.lr_steps"
          ],
          "automl_enabled": false,
          "default": {
            "lr": 0.0001,
            "lr_backbone": 1e-05,
            "lr_decay": 0.1,
            "lr_scheduler": "MultiStep",
            "lr_step_size": 1000,
            "lr_steps": [
              1000
            ],
            "momentum": 0.9,
            "monitor_name": "val_loss",
            "optimizer": "AdamW",
            "warmup_steps": 0,
            "weight_decay": 0.0001
          },
          "description": "Hyper parameters to configure the optimizer.",
          "properties": {
            "lr": {
              "automl_enabled": true,
              "default": 0.0001,
              "description": "The initial learning rate for training the model, excluding the backbone.",
              "math_cond": "> 0.0",
              "maximum": 1.0,
              "minimum": 0.0,
              "title": "learning rate",
              "type": "float"
            },
            "lr_backbone": {
              "automl_enabled": true,
              "default": 1e-05,
              "description": "The initial learning rate for training the backbone.",
              "math_cond": "> 0.0",
              "maximum": 1.0,
              "minimum": 0.0,
              "title": "learning rate - backbone",
              "type": "float"
            },
            "lr_decay": {
              "automl_enabled": true,
              "default": 0.1,
              "description": "The decreasing factor for the learning rate scheduler.",
              "math_cond": "> 0.0",
              "maximum": 1.0,
              "minimum": 0.0,
              "title": "learning rate decay",
              "type": "float"
            },
            "lr_scheduler": {
              "default": "MultiStep",
              "description": "The learning scheduler:\n                    * MultiStep : Decrease the lr by lr_decay from lr_steps\n                    * StepLR : Decrease the lr by lr_decay at every lr_step_size.",
              "enum": [
                "MultiStep",
                "StepLR"
              ],
              "title": "learning rate scheduler",
              "type": "categorical"
            },
            "lr_step_size": {
              "automl_enabled": true,
              "default": 1000,
              "description": "The number of steps to decrease the learning rate in the StepLR.",
              "math_cond": "> 0",
              "maximum": 10000,
              "minimum": 1,
              "title": "learning rate step size",
              "type": "int"
            },
            "lr_steps": {
              "automl_enabled": false,
              "default": [
                1000
              ],
              "description": "The steps at which the learning rate must be decreased.\n                    This is applicable only with the MultiStep LR.",
              "title": "learning rate decay steps",
              "type": "list"
            },
            "momentum": {
              "automl_enabled": true,
              "default": 0.9,
              "description": "The momentum for the AdamW optimizer.",
              "math_cond": "> 0.0",
              "maximum": 1.0,
              "minimum": 0.0,
              "title": "momentum - AdamW",
              "type": "float"
            },
            "monitor_name": {
              "default": "val_loss",
              "description": "The metric value to be monitored for the :code:`AutoReduce` Scheduler.",
              "enum": [
                "val_loss",
                "train_loss"
              ],
              "title": "monitor_name",
              "type": "categorical"
            },
            "optimizer": {
              "default": "AdamW",
              "description": "Type of optimizer used to train the network.",
              "enum": [
                "AdamW",
                "SGD"
              ],
              "type": "categorical"
            },
            "warmup_steps": {
              "default": 0,
              "description": "The number of steps to perform linear learning rate warm-up.",
              "maximum": Infinity,
              "minimum": 0,
              "title": "warm up steps",
              "type": "int"
            },
            "weight_decay": {
              "automl_enabled": true,
              "default": 0.0001,
              "description": "The weight decay coefficient.",
              "math_cond": "> 0.0",
              "maximum": 1.0,
              "minimum": 0.0,
              "title": "weight decay",
              "type": "float"
            }
          },
          "title": "optimizer",
          "type": "collection"
        },
        "precision": {
          "default": "fp32",
          "description": "Precision to run the training on.",
          "enum": [
            "bf16",
            "fp32",
            "fp16"
          ],
          "title": "precision",
          "type": "categorical"
        },
        "pretrained_model_path": {
          "default": "",
          "description": "Path to a pre-trained RT-DETR model to initialize the current training from.",
          "type": "string"
        },
        "results_dir": {
          "default": "",
          "description": "\n        Path to where all the assets generated from a task are stored.\n        ",
          "title": "Results directory",
          "type": "string"
        },
        "resume_training_checkpoint_path": {
          "default": "",
          "description": "Path to the checkpoint to resume training from.",
          "title": "Resume checkpoint path",
          "type": "string"
        },
        "seed": {
          "default": 1234,
          "description": "The seed for the initializer in PyTorch. If < 0, disable fixed seed.",
          "maximum": Infinity,
          "minimum": -1,
          "title": "Seed for randomization",
          "type": "int"
        },
        "validation_interval": {
          "default": 1,
          "description": "\n        The interval (in epochs) at which a evaluation\n        will be triggered on the validation dataset.",
          "minimum": 1,
          "popular": true,
          "title": "Validation interval",
          "type": "int"
        }
      },
      "type": "collection"
    },
    "wandb": {
      "automl_disabled_parameters": [
        "wandb.tags"
      ],
      "automl_enabled": false,
      "default": {
        "enable": true,
        "entity": "",
        "group": "",
        "name": "TAO Toolkit Training",
        "project": "TAO Toolkit",
        "reinit": false,
        "run_id": "",
        "save_code": false,
        "sync_tensorboard": false,
        "tags": [
          "tao-toolkit"
        ]
      },
      "properties": {
        "enable": {
          "default": true,
          "type": "bool"
        },
        "entity": {
          "default": "",
          "type": "string"
        },
        "group": {
          "default": "",
          "type": "string"
        },
        "name": {
          "default": "TAO Toolkit Training",
          "type": "string"
        },
        "project": {
          "default": "TAO Toolkit",
          "type": "string"
        },
        "reinit": {
          "default": false,
          "type": "bool"
        },
        "run_id": {
          "default": "",
          "type": "string"
        },
        "save_code": {
          "default": false,
          "type": "bool"
        },
        "sync_tensorboard": {
          "default": false,
          "type": "bool"
        },
        "tags": {
          "automl_enabled": false,
          "default": [
            "tao-toolkit"
          ],
          "type": "list"
        }
      },
      "type": "collection"
    }
  },
  "required": [
    "quantize"
  ],
  "type": "object",
  "x_tao_schema": {
    "action": "distill",
    "core_module": "rtdetr",
    "model": "rtdetr",
    "network_arch": "rtdetr",
    "schema_action": "distill",
    "schema_version": 1,
    "source": "tao-core dataclass config"
  }
}

Requirements and compatibility

Works with: Claude Code, Claude apps, OpenAI Codex, GitHub Copilot, Cursor, Gemini CLI, Kiro, OpenCode, Goose, Amp, Roo Code, JetBrains Junie, Factory Droid, Devin (ex-Windsurf), Mistral Vibe, OpenHands, Letta

“Works with” means the tool reads the open Agent Skills format; a skill written for one tool may use features others lack.

Needs: Declares allowed tools, Needs Python, Needs Node.js

The author says: Requires docker + nvidia-container-toolkit.

Security

What our scanner noticed· Medium risk

Large base64-encoded blobs

A static scan is not a guarantee. Skills are third-party content: review a skill before you let an agent use it.

Attribution and licence

Created by

NVIDIA Corporation

Publisher

NVIDIA (Official)

Source

NVIDIA/skills · skills/tao-train-rtdetr · commit fce3cca

Licence

Apache License 2.0 (redistribution allowed)

LICENSE (added by AutoGineer from LICENSE-APACHE)
ATTRIBUTION.md (written by AutoGineer)

Synced

2026-11-09

AutoGineer is not affiliated with or endorsed by NVIDIA Corporation or NVIDIA.

Download ZIP

The ZIP holds the skill’s files, its licence files and an ATTRIBUTION.md. Download tao-train-rtdetr.zip