Tao Run Deft Object Detection

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:

   Copy

   <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:

   Copy

   <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.

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.

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

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.