AutoGineer
Generated banner for Tao Run Deft Object Detection
Generated logo for Tao Run Deft Object Detection

Tao Run Deft Object Detection

NVIDIAOfficialCreated by NVIDIA Corporation

  • Skill
  • Workflow
  • Software development
  • Apache-2.0
  • Medium risk
  • Has scripts
Download ZIP

From the author

Run the full DEFT smart-data-augmentation loop for NVIDIA TAO Grounding DINO object detection: zero-shot baseline inference, KPI analysis, per-class gap analysis, SigLIP embedding of weak images, unique-neighbor mining against a source pool, ODVG dataset staging, and retraining — repeated for a fixed number of iterations. Also prepares the source pool the loop mines from, as a separate run: Co-DETR pseudo-labeling, folding to the target classes, KITTI→COCO→ODVG conversion, and embedding. Use for prompts like "run the DEFT OD loop", "run smart data augmentation for grounding dino", "mine and retrain my detection model", "improve OD mAP with gap analysis and mining", "prep the source pool", or "pseudo-label my unlabeled images for mining"; do not use for standalone TAO training, one-off inference, or gap analysis alone.
— the description by NVIDIA Corporation

AutoGineer summary

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

This skill executes a multi-stage, iterative workflow designed to improve object detection model performance through smart data augmentation. It manages a complete loop including zero-shot baseline inference, KPI analysis, per-class gap analysis, SigLIP-based embedding, unique-neighbor mining from a source pool, dataset staging, and retraining. The skill also provides a separate invocation for preparing the source pool, which involves pseudo-labeling raw images with Co-DETR, remapping classes, and converting data formats. It operates as a state-managed loop, ensuring consistency across iterations and providing automated reporting.

What it does

  • Performs zero-shot baseline inference and KPI analysis on Grounding DINO models.
  • Conducts per-class gap analysis to identify model weaknesses.
  • Mines unique, high-value training samples from a prepared source pool using SigLIP embeddings.
  • Stages mined data into the ODVG format for iterative model retraining.
  • Prepares source pools by pseudo-labeling raw images and converting them to the required ODVG format.
  • Maintains a persistent state machine to track progress and ensure auditability across training iterations.

When to use it

  • When you need to improve the mAP of a Grounding DINO model through iterative data mining.
  • When you have a large pool of unlabeled images that need to be pseudo-labeled and mined for training.
  • When you want to automate the gap analysis, mining, and retraining cycle.
  • When you need to prepare a source pool of images for use in a DEFT-based detection workflow.

Advantages

  • Automates complex, multi-step data augmentation and training loops.
  • Reduces manual effort in identifying and labeling high-value training data.
  • Ensures consistent data processing and state management across iterations.
  • Provides clear visibility into model performance trends through automated KPI reporting.

Usage examples

“run the DEFT OD loop”
The agent initiates the iterative mining and retraining process, starting with baseline inference and gap analysis.
“prep the source pool”
The agent runs the Co-DETR pseudo-labeling and embedding pipeline to prepare your raw images for future mining.
“mine and retrain my detection model”
The agent performs gap analysis, mines new samples from the source pool, stages the data, and triggers a retraining iteration.
“improve OD mAP with gap analysis and mining”
The agent executes the full DEFT loop to identify performance gaps and iteratively grow the training dataset to boost mAP.

Install

  1. Download the skill as a ZIP:

    https://autogineer.com/agent-skills/nvidia/tao-run-deft-object-detection/download

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

    mkdir -p .agents/skills && unzip tao-run-deft-object-detection.zip -d .agents/skills

Documentation by NVIDIA Corporation

Skill: tao-run-deft-object-detection

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).

Execution Contract

Treat this as a disk-backed state machine, not as a prose recipe.

  1. Preserve every explicit user value. epoch 1 means train.num_epochs=1; a spec value or documented default applies only when the user did not supply that parameter. Show the source of every run parameter (user, spec, or default) in the Pre-Flight Summary.

  2. After the user approves the Summary, initialize deft_state.json once with scripts/init_deft_state.py. Never hand-author or reinitialize it on resume.

  3. Run every bundled or inline host-Python command through scripts/deft_python.sh. On startup, after context compaction, before every stage, and before any completion claim, run:

    <skill_root>/scripts/deft_python.sh \
  <skill_root>/scripts/audit_deft_run.py --results-dir "${RESULTS_DIR}"

    If it prints DEFT_RUN_STATUS=INVALID, stop and repair the listed disk inconsistency; do not launch another stage. Read the path printed as read_before_action before continuing.

  4. Invoke the mapped underlying skill after reading the DEFT overlay. Do not replace a missing or unread stage reference, or a failed skill call, with guessed shell commands, inline Python, a different output tree, or fabricated data.

  5. Commit every stage with scripts/commit_stage.py; it verifies artifacts, updates deft_state.json, appends exactly one ordered loop_log.jsonl event, and rolls back if its audit fails.

  6. Claim the loop complete only when this exits zero:

    <skill_root>/scripts/deft_python.sh \
  <skill_root>/scripts/audit_deft_run.py \
  --results-dir "${RESULTS_DIR}" --require-complete

Context Discipline

  • Load references just in time. Run the audit, read only its read_before_action file and the current stage's named section, then act. Never preload all references.

  • Redirect verbose train, inference, and Docker output to files. Inspect at most the final 40 lines or a one-line artifact check; never print a full spec, state file, or loop log into the conversation.

  • A Skill-tool call loads stage instructions; it does not start a background orchestrator. Continue the documented stage in the parent immediately after it returns.

When to Use This Skill

Use this skill when the user wants an agent to run the full smart-data-augmentation loop for a TAO Grounding DINO detection model: zero-shot baseline, gap analysis, mining, dataset growth, and retraining across N iterations.

  • "Run the DEFT OD loop"

  • "Run smart data augmentation for grounding dino"

  • "Mine more training data for my detection model and retrain"

  • "Improve detection mAP with gap analysis and unique-neighbor mining"

Also use it to prepare the source pool, which is a separate run that completes before the loop launches (see ## Two Invocations: Prep, Then Loop):

  • "Prep the source pool"

  • "Pseudo-label my unlabeled images for mining"

  • "Build the mining pool from these raw images"

Do not use this skill for a single standalone TAO training run, one-off inference, or gap analysis alone. Invoke the relevant leaf skill directly instead.

Scope: Grounding DINO + ODVG

This loop targets Grounding DINO with ODVG training annotations (tmm_odvg.jsonl + labelmap.json), matching the reference pipeline. dataset.train_data_sources is a list; each iteration appends one new ODVG source rather than rewriting a combined CSV. DINO and RT-DETR (COCO) are not supported by this workflow — use the leaf skills directly for those.

The loop does not train at baseline. It evaluates the supplied zero-shot / pretrained checkpoint as iteration 0 and only trains from iteration 1 onward, once mining has produced data to add.

Two Invocations: Prep, Then Loop

The source pool is prepared by its own run, before the loop launches. They are separate because a pool is prepped once and then serves many loop runs — coupling them would re-label and re-embed the same images on every launch.

InvocationDoesProduces
"Prep the source pool"Co-DETR pseudo-labels raw pool images, folds to the target classes, converts KITTI→COCO→ODVG, verifies, embedscoco.json, odvg/, source_embeddings.parquet, pool_report.json
"Run the DEFT loop"baseline → iterationscheckpoints, KPI, the mAP trend

Follow references/prep-source-pool.md for the first. The loop then takes those four paths as inputs; Pre-Flight validates them and init_deft_state.py pins them, so a run cannot reach mine with no corpus to search.

Launch Intake

Do not ask which platform to run on. This workflow supports one: every stage launches with docker run, and its compatibility line requires docker plus the nvidia-container-toolkit. The bank's rule in AGENTS.md is to treat a single supported platform as the selection and only ask when an application supports several — asking here would offer a choice that nothing downstream can act on.

Read tao-run-on-docker's ## Credentials section.

Never ask for or read credential values. Check only whether the required environment variable is set; if unset, tell the user which variable to export.

Agent Behavior

There is exactly one user gate: pre-flight confirmation. Print the Pre-Flight Summary (see references/preflight.md), then STOP and wait for explicit approval ("go", "yes", "looks good"). Do not launch any side-effecting step before that approval.

After the gate, the skill is fully autonomous. Run the entire loop without asking for confirmation. Only stop if a step fails with an unrecoverable error or a hard-stop gate fires. Print a one-line status update at each stage milestone.

Auto-mode required. The post-gate loop fires constant side-effecting calls; without auto-accept mode it stalls on the first prompt. Remind the user at the Pre-Flight Summary to enable auto-mode (shift+tab) before approving.

Non-zero command rule. Never repeat an unchanged failed command. Read the final error block, map it to the loaded stage reference, make one evidence-based correction, and rerun its documented verification. If the reference does not cover the failure, commit status=error and halt.

Revised plan. If any run parameter changes after the Summary was shown, re-run Pre-Flight and show an updated Summary before proceeding.

Workflow

Full detail in references/pipeline-and-state.md.

  1. Pre-Flight. Run every check in references/preflight.md. Resolve workspace, specs, annotations, the zero-shot checkpoint, the source-pool embedding parquet, and container images. Hard stop only on missing input you cannot resolve yourself.

  2. Prep (once, before baseline). If the source pool is not already labeled and embedded, pseudo-label it with Co-DETR, fold the predictions onto the user's target classes, convert KITTI→COCO→ODVG, and embed the pool. Skip any step whose output already exists — no script does this for you, so check before each step. See references/prep-source-pool.md.

  3. Baseline (iter_0) — no training. Run inference with the supplied zero-shot / pretrained checkpoint, then kpi_analyze. Seed train_grounding_dino.yaml from the user's template for later iterations to extend.

  4. Iterate. For each iteration 1..max_iterations, run the seven stages in order: gap_analysis → embed → mine → stage → train → inference → kpi_analyze. Each iteration's gap_analysis consumes the previous phase's inference labels. Between stages run the audit and follow its one-line disk-backed next action.

  5. Stop when max_iterations is reached or a hard-stop gate fires. mAP is reported, not gated — the loop does not early-exit on a metric target.

  6. The report renders itself. results/DEFT_Loop_Report.md is written by init_deft_state.py and re-rendered by every accepted commit_stage.py call. Do not spawn a subagent for it and do not compose it by hand; to refresh it out of band, run scripts/render_report.py --results-dir "${RESULTS_DIR}".

All stages run inline in the parent context. Prefer invoking the underlying tao-skill-bank:* skills via the Skill tool, layering loop conventions on top via the matching references/*.md overlay.

Using Bundled Scripts

Run bundled scripts through <skill_root>/scripts/deft_python.sh. Resolve every path argument to an absolute host path first. Use commit_stage.py for all state and log writes. See references/scripts-and-agents.md.

Stage Reference Modules

Each stage maps to one underlying skill or to bundled glue. Read only the current stage's overlay, then invoke.

The overlays are written against two loop variables that Pre-Flight sets and every stage uses: N, the current iteration number, and PHASE, which is baseline or iter${N}. They are stated here because a stage overlay is read on its own — see references/preflight.md for the full set.

If an overlay is missing, stop and ask the user to reinstall the plugin — the loop cannot run a stage whose settings it does not have. If a mapped skill is unavailable, do not stop and do not improvise: fall back to the overlay's documented docker run as described in references/scripts-and-agents.md, and commit that stage with --execution-path direct-container instead of the documented skill. The overlay carries everything the invocation needs, so the fallback produces the same artifacts.

On a runtime without a Skill tool (Codex, for example), the mapped skills are still on disk under the plugin root: read skills/<layer>/<name>/SKILL.md there for the invocation contract, then run it through the direct-container fallback above, committing the stage with --execution-path direct-container. On such a runtime that fallback is the expected route for every stage, not an exception.

StageOverlayUnderlying skill
prep (once)references/prep-source-pool.mdtao-skill-bank:tao-train-codetr + tao-generate-image-embeddings (+ bundled glue)
gap_analysisreferences/tao-analyze-gaps-od-map.mdtao-skill-bank:tao-analyze-gaps-od-map
embedreferences/tao-generate-image-embeddings.mdtao-skill-bank:tao-generate-image-embeddings
minereferences/tao-mine-od-images.mdtao-skill-bank:tao-mine-od-images
stagereferences/stage-mined-data.md(bundled glue — no leaf skill)
train, inferencereferences/grounding-dino.mdtao-skill-bank:tao-train-grounding-dino
kpi_analyzereferences/tao-analyze-detection-kpi.mdtao-skill-bank:tao-analyze-detection-kpi

Path rule (invariant). Record absolute host paths under ${RESULTS_DIR}. Mount "$WORKSPACE:$WORKSPACE" with identical host and container paths. TAO's update_results_dir appends the task name to results_dir, so passing results_dir=X to train writes X/train/ and to inference writes X/inference/. Never append the subdirectory yourself.

Data, Pre-Flight, Pipeline, and State References

TopicReference
Data contract, ODVG layout, source pool, output treereferences/data-layout.md
One-time source-pool prep (pseudo-label, remap, convert, embed)references/prep-source-pool.md
Pre-Flight checks, defaults, Summary templatereferences/preflight.md
Pipeline stages, state schema, loop-end sequencereferences/pipeline-and-state.md
Bundled scripts, glue, the report renderer, stage tablereferences/scripts-and-agents.md

max_iterations defaults to 1 — one mine, train and score pass, the smallest run that yields a comparison against the baseline. Confirm it with the user when they have not said how many iterations they want; an unattended run takes the default rather than stopping to ask.

Gating

Run the full Pre-Flight, print the Summary, then STOP at the one user gate. After approval, run the baseline and the seven-stage iteration pipeline.

Hard-stop and never auto-retry on: any stage status=error; a missing or zero-row source-pool embedding parquet; a zero-row mining result when weak images were present; a missing ODVG annotation source; an image/annotation mismatch after staging; or a train exit that emits no new iteration checkpoint. The loop stops when max_iterations is reached or an unrecoverable gate fires. Each terminal path commits loop_stop through commit_stage.py, then follows the loop-end sequence in references/pipeline-and-state.md.

Files

references/tao-mine-od-images.md

DEFT OD — Mining Stage Overlay

Layers loop conventions on top of tao-skill-bank:tao-mine-od-images. Read that skill's SKILL.md for the full field reference.

When to invoke

Iteration stage 3, after embed. Mines the source pool for images resembling this iteration's weak set.

Inputs from state

State fieldDescription
iterations.iter${N}.embeddings_parquetThis iteration's weak_images_embeddings.parquet — the mining target.
config.source_pool_embeddingsPre-embedded source pool — the mining source. Must use the same encoder as the target.
iterations.iter$((N-1)).exclude_parquetPrevious cumulative mined set. Absent on iteration 1.
config.source_detection_fileSource-pool COCO JSON — required for class_stratified.
config.target_detection_fileKPI COCO JSON — required for class_stratified. It describes the KPI set, not the pool: stratified allocation reads it to find how many boxes of each rare class the target set holds. If the KPI package ships one, use it. If the ground truth is KITTI, build one with annotations convert the same way prep converts the pool (references/prep-source-pool.md, step 3), pointing it at the KPI labels. Without it, use --allocation-policy global.
config.rare_class_listComma-separated rare classes, e.g. "person,bicycle".

exclude_path must be null on iteration 1

This is a hard failure, not a warning. TAO DS load_datasets raises FileNotFoundError when exclude_path is set but is not a file — and on iteration 1 no cumulative parquet exists yet. Pointing at the not-yet-written path kills the stage before any mining happens.

A spent pool fails differently and worse. When the exclude set already covers every pool image, split_datasets_by_class raises TypeError: Series object is not iterable from selection.py — an unhandled cudf error naming neither the pool nor the exclude set. Count the pool rows not already excluded before invoking, and stop the loop instead of calling the miner with nothing left to give (commit_stage.py --stage loop_stop --pool-exhausted --pool-remaining N).

  • Iteration 1: emit exclude_path: null.

  • Iteration N > 1: emit the absolute path to ${RESULTS_DIR}/iter$((N-1))/mined_cumulative.parquet, and confirm the file exists before writing the spec.

detection_format is never inferred

TAO DS raises detection_format is required (coco or kitti) when a detection file is provided rather than guessing. Set it explicitly to coco and supply COCO JSONs — a KITTI label directory passed as a detection file would be misparsed.

Note this stage uses COCO detection files while gap_analysis in the same loop reads KITTI inference labels. That split is correct: they are different inputs to different steps.

Rare classes

class_stratified allocation needs to know which classes are rare. The loop derives it from pool_report.json: any target class holding a below-mean share of the pool's annotations. Scarcity in the pool is the signal, not scarcity in the KPI set — stratified allocation exists so classes the pool holds few of still get their share of the budget.

The derivation is reported with the counts behind it, so the choice is visible rather than silent. It belongs in tmm unique_neighbor_matching itself, which already reads both inputs it would need; tracked separately as a TAO DS request.

Spec

Write per-iteration under ${RESULTS_DIR}/iter${N}/mining/unique_neighbor_matching.yaml — the invocation below reads it from $MINE_SPEC, so bind the two:

MINE_SPEC="${RESULTS_DIR}/iter${N}/mining/unique_neighbor_matching.yaml"
source_path:            <config.source_pool_embeddings>
target_path:            <iterations.iter${N}.embeddings_parquet>
output_dir:             <absolute path to ${RESULTS_DIR}/iter${N}/mining>
desired_unique_count:   <from prepare_budget_for_mining.py — see below>
allocation_policy:      class_stratified     # or global — see below
distance_metric:        euclidean
candidate_expansion_factor: 5
source_embedding_column: embedding
target_embedding_column: embedding
source_filepath_column: filepath
target_filepath_column: filepath
exclude_path:           null                 # iteration 1 ONLY; else the prev cumulative path
source_detection_file:  <config.source_detection_file>
target_detection_file:  <config.target_detection_file>
detection_format:       coco
rare_class_list:        "person,bicycle"
save_embeddings:        false
visualize:              false

Allocation policy. Use class_stratified when the user supplied rare classes — budget is apportioned by target class ratio using largest-remainder, with a non-rare fallback pool, and each source is owned by the first class that claims it. Use global when no rare classes are configured. The reference pipeline's class_balanced mode maps to class_stratified.

rare_class_list comes from state, not from this stage. Read config.rare_class_list and write it into the spec verbatim. It is derived once, at the prep commit, from the pool's own annotation counts — every target class holding a below-mean share. Under class_stratified the commit refuses a mine while it is still unset, since allocation apportions the budget by exactly that list.

candidate_expansion_factor is not the loop's iteration count. It seeds the miner's internal candidate-pool growth (starting at 5, widening each internal retry, capped at 8 internal passes). It is unrelated to max_iterations.

Invocation

<skill_root>/scripts/deft_python.sh <skill_bank>/skills/data/tao-mine-od-images/scripts/verify_unique_neighbor_matching_spec.py \
  --spec "$MINE_SPEC"

docker run --rm --name "deft_iter${N}_mine" --gpus all --shm-size=8g --user "$(id -u):$(id -g)" $DOCKER_IDENTITY \
  -v "$WORKSPACE:$WORKSPACE" $EXTRA_MOUNTS -w "$WORKSPACE" \
  "$TAO_DS_IMAGE" \
  tmm unique_neighbor_matching -e "$MINE_SPEC"

Outputs

ArtifactPath
final_unique_files.parquet${RESULTS_DIR}/iter${N}/mining/final_unique_files.parquet
summary.json${RESULTS_DIR}/iter${N}/mining/summary.json

final_unique_files.parquet is a single column named by source_filepath_column (filepath). It feeds the stage step.

Read summary.json:

  • retrieved_unique_count == 0 with weak images present → hard stop. The source pool has nothing left to give, or the exclude set has consumed it.

  • coverage_pct < 50 → warn in the iteration summary; the pool is running dry and later iterations will add little.

  • A target class with no images in resultant_dataset.per_class → name it in the iteration summary, with its allocation.per_class.desired_count where it has one. Its slots are backfilled from other classes, so retrieved_unique_count and coverage_pct stay healthy and neither rule above fires; this is the only place the starvation shows. The loop continues — the iteration still has data to train on — and DEFT_Loop_Report.md flags the class too.

  • Otherwise proceed.

Commit

<skill_root>/scripts/deft_python.sh <skill_root>/scripts/commit_stage.py \
  --results-dir "${RESULTS_DIR}" --iter-label "iter${N}" --stage mine \
  --execution-path skill \
  --mining-output "${RESULTS_DIR}/iter${N}/mining/final_unique_files.parquet" \
  --mining-summary "${RESULTS_DIR}/iter${N}/mining/summary.json" \
  --duration-sec "$(( SECONDS - started ))" \
  --summary "mined <retrieved>/<desired> unique images (<coverage_pct>% coverage)"

Computing desired_unique_count

The budget is produced before this stage by prepare_budget_for_mining.py, and its full invocation lives in references/stage-mined-data.md. Reading only this overlay leaves the field with no documented way to compute it, so the short form is repeated here:

<skill_root>/scripts/deft_python.sh <skill_root>/scripts/prepare_budget_for_mining.py \
  --weak-parquet "${RESULTS_DIR}/iter1/gaps/weak_images.parquet" \
  --multiplier "<multiplier>" --pool-size "<pool rows>" \
  --already-mined "<cumulative exclude rows, 0 on iteration 1>" \
  --remaining-iterations "<iterations left, including this one>"

Seed it from iteration 1's weak parquet on every iteration, so the amount of data added per iteration stays constant as the gap set shrinks.

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

The author says: Requires docker + nvidia-container-toolkit and one or more CUDA GPUs. Workflows declare additional requirements.

Security

What our scanner noticed· Medium risk

  • Large base64-encoded blobs
  • References to credential files
  • Scripts that your agent may run

This skill includes scripts that your agent may run. Read them before you use it.

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-run-deft-object-detection · commit fce3cca
Licence
Apache License 2.0 (redistribution allowed)
  • LICENSE (added by AutoGineer from LICENSE-APACHE)
  • ATTRIBUTION.md (written by AutoGineer)
Synced

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-run-deft-object-detection.zip