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.pretrained_backbone_path for backbone-only loading or train.pretrained_model_path for full model initialization.

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

Training Requirements

• Dataset type: object_detection
• Formats: coco, coco_raw
• Monitoring metric: val_mAP50 for AP50; 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"

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.enc_layers, model.dec_layers, model.num_queries, or model.num_select, carry the same values into evaluate, export, inference, and deploy actions with the selected checkpoint. In addition to the fields above, carry model.num_feature_levels, 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 deformable_detr gen_trt_engine from the parent PyT model container; that CLI advertises convert, evaluate, export, inference, quantize, train, and default_specs, but not gen_trt_engine. 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 gen_trt_engine.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 model_epoch_<epoch>_step_<step>.pth, plus a dd_model_latest.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 dd_model_latest.pth symlink, unless the user explicitly asked for latest. Resume/retrain should set train.resume_training_checkpoint_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.eval_class_ids: 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 resnet_50. Supported: resnet_50, gcvit_tiny, gcvit_small, gcvit_base, gcvit_large, gcvit_large_384 (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 WORLD_SIZE, NODE_RANK, MASTER_ADDR, MASTER_PORT env vars set by orchestrator.

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].

Export / TRT Defaults

• Export input: 640x640, opset 17
• TRT data types: FP32, FP16, INT8
• TRT workspace: 1024 MB
• TRT max_batch_size: 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).

num_select must be < num_queries * num_classes: Same constraint as DINO.

return_interm_indices length must match num_feature_levels: Default [1,2,3,4] with num_feature_levels=4.

Dataset size smaller than total batch size: Reduce batch_size or num_gpus.

AutoML metric extraction: Deformable 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. Prefer results_dir/train/status.json or AutoML result state before parsing raw logs. Do not optimize val_loss 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 create_job(). This mirrors the old microservices infer_params.py flow.

Inference mappings from TAO Core deformable_detr.config.json:

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-deformable-detr