Deformable DETR

Deformable DETR for 2D object detection. Uses deformable attention for efficient multi-scale feature processing. Lighter than DINO with competitive accuracy.

Uses pretrained weights. Set model.pretrainedbackbonepath for backbone-only loading or train.pretrainedmodelpath for full model initialization.

Supported parent model actions are train, evaluate, inference, export, and quantize. The PyT model container does not support a native gentrtengine subtask for this network. The gentrtengine action declared in references/skillinfo.yaml must run with the TAO Deploy container. 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 still requires schemas/train.schema.json and references/spectemplatetrain.yaml to exist and parse. Use the packaged train 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.

Training Requirements

• Dataset type: object_detection
• Formats: coco, coco_raw
• Monitoring metric: valmAP50 for AP50; valmAP 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"

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": "<object classes> + 1",
    "dataset.eval_class_ids": [1, 2, "..."],
    "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_TRAIN}/images.tar.gz", "json_file": f"{S3_TRAIN}/annotations.json"}],
}

evaluate (mandatory data sources):

{
    "dataset.num_classes": "<object classes> + 1",
    "dataset.eval_class_ids": [1, 2, "..."],
    "dataset.test_data_sources.image_dir": f"{S3_EVAL}/images.tar.gz",
    "dataset.test_data_sources.json_file": f"{S3_EVAL}/annotations.json",
}

If the train or AutoML run changed architecture-affecting fields such as model.enclayers, model.declayers, model.numqueries, or model.numselect, carry the same values into evaluate, export, inference, and deploy actions with the selected checkpoint. In addition to the fields above, carry model.numfeaturelevels, model.dim_feedforward, input image dimensions, and dataset class metadata when they were changed. Loading a checkpoint into the default architecture can fail with tensor shape mismatches, especially when smoke-test runs shrink the transformer for speed.

export (mandatory data sources):

{
    "dataset.num_classes": "<object classes> + 1",
    "dataset.eval_class_ids": [1, 2, "..."],
    "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_TRAIN}/images.tar.gz", "json_file": f"{S3_TRAIN}/annotations.json"}],
}

TensorRT engine generation:

Use the deploy spec templates after export. Do not call deformabledetr gentrtengine from the parent PyT model container; that CLI advertises convert, evaluate, export, inference, quantize, train, and defaultspecs, but not gentrtengine. The model action metadata selects the TAO Deploy container for engine generation.

Deploy engine generation needs the exported ONNX file as input and creates the engine at gentrtengine.trt_engine.

{
    "gen_trt_engine.tensorrt.data_type": "FP16",
    "dataset.num_classes": "<object classes> + 1",
    "gen_trt_engine.tensorrt.calibration.cal_image_dir": [f"{S3_TRAIN}/images.tar.gz"],
}

inference (mandatory data sources):

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

quantize (mandatory data sources):

{
    "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_TRAIN}/images.tar.gz", "json_file": f"{S3_TRAIN}/annotations.json"}],
    "dataset.quant_calibration_data_sources": {"image_dir": f"{S3_TRAIN}/images.tar.gz", "json_file": f"{S3_TRAIN}/annotations.json"},
}

Eval Dataset

Optional. If provided, validation mAP is computed at each checkpoint interval.

Checkpoint Handling

Training emits epoch-and-step checkpoints using the pattern modelepoch<epoch>step<step>.pth, plus a ddmodellatest.pth symlink. For dependent actions, use the model-specific or SDK-provided checkpoint resolver to select the intended artifact. Evaluation, inference, export, and quantize should receive the selected exact checkpoint path, not the ddmodellatest.pth symlink, unless the user explicitly asked for latest. Resume/retrain should set train.resumetrainingcheckpoint_path to the exact checkpoint being resumed from.

Important Parameters

• dataset.num_classes: Number of object classes plus the background class. Default 91 (COCO). Must match annotations.
• dataset.evalclassids: Foreground category ids to include in COCO metrics. Set this to every object category id in custom datasets; the template default evaluates class id 1 only.
• model.backbone: Default resnet50. Supported: resnet50, gcvittiny, gcvitsmall, gcvitbase, gcvitlarge, gcvitlarge384 (more limited than DINO).
• train.optim.lr: Learning rate. Default 2e-4 (AdamW). lr_backbone is 2e-5.
• train.optim.lr_steps: MultiStep LR schedule. Default [40]. For short runs, set to match ~80% of total epochs.
• model.num_queries: Number of object queries. Default 300. Valid range 100-900.
• model.dropout_ratio: Dropout in transformer layers. Default 0.3 (higher than DINO's 0.0). Reduce for large datasets, increase for small datasets.
• model.dim_feedforward: FFN hidden dim. Default 1024 (vs DINO's 2048). Increasing improves capacity but costs memory.

Multi-GPU / Multi-Node

Launch method: Lightning-managed (single python process, Lightning spawns workers).

Same DDP/FSDP behavior as DINO. Multi-node requires WORLDSIZE, NODERANK, MASTERADDR, MASTERPORT env vars set by orchestrator.

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].

Export / TRT Defaults

• Export input: 640x640, opset 17
• TRT data types: FP32, FP16, INT8
• TRT workspace: 1024 MB
• TRT maxbatchsize: 1

Hardware

Minimum 1 GPU(s), recommended 4 GPU(s). 16GB+ (V100 or A100) VRAM per GPU. Slightly lighter than DINO due to smaller FFN. batch_size=4 fits on most 16GB+ GPUs.

Error Patterns

CUDA out of memory: Reduce batch_size (4 -> 2 -> 1).

**numselect must be < numqueries * num_classes**: Same constraint as DINO.

returnintermindices length must match numfeaturelevels: Default [1,2,3,4] with numfeaturelevels=4.

Dataset size smaller than total batch size: Reduce batchsize or numgpus.

AutoML metric extraction: Deformable 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 valmAP50. Prefer resultsdir/train/status.json or AutoML result state before parsing raw logs. Do not optimize valloss for default detection model invocations.

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 deformable_detr.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-deformable-detr](references/tao-deploy-deformable-detr.md)