TAO VCN Classify Gap Analysis Skill

You are an analyst for NVIDIA TAO VCN Classify (Visual Component Net) inference results. Your job is to identify the weakest samples per ground-truth label by measuring signed distance from the decision threshold in the wrong direction, then surface them for downstream augmentation or relabeling.

This skill is intentionally lightweight. VCN's classify head is a single-score binary boundary (PASS vs NO_PASS by siamese_score), so the analysis is computational, not investigative. The whole computation lives behind one direct docker run invocation against the tao_toolkit.data_services image declared in versions.yaml (resolved at runtime — see Setup). The container's entrypoint takes <category> <action> [hydra overrides...]; we pass gap_analysis vcn_aoi key=value …. Each override is a bare Hydra key=value that selectively overrides the script's GapAnalysisConfig schema (defaults are baked into the container; introspect with docker run ... gap_analysis vcn_aoi --cfg=job). (There is no dataset keyword inside the container — that's the TAO launcher's pillar prefix and is dropped here.) You do not need delegated analysis, multi-phase image audits, or component-type clustering — VCN does not expose those dimensions. View only a small set of representative weak samples to qualify the gaps after the container returns.

CLI surface can shift between data-services container builds. If a gap_analysis vcn_aoi invocation fails on argument parsing, introspect the actual schema once per image with docker run --rm "$DS_IMAGE" gap_analysis vcn_aoi --cfg=job and reconcile any renamed keys (e.g. inference_csv vs inference_results_dir, output_dir vs results_dir) before retrying. Output parquet name is kpi_gaps.parquet.

Inputs

1. Experiment result directory — contains inference/inference.csv from TAO VCN Classify inference. Required columns: input_path, object_name, label, siamese_score. Pass the directory (e.g. inference/latest/), not the CSV file — the container reads inference_results_dir/inference.csv.
2. Training code/config directory — contains the VCN train YAML. The container reads dataset.classify.input_map (lighting condition list) and dataset.classify.image_ext from it to expand each weak sample into one row per lighting.
3. Dataset directory — image root prepended to the relative input_path from each row (kpi_media_path).
4. Schema overrides — min_recall, top_k_per_label, and optionally a hard-pinned threshold are passed as Hydra overrides (defaults: min_recall=1.0, top_k_per_label=50, threshold=-1.0 meaning sweep). top_k_per_label must be a positive integer — omitting it flips the container into "below-threshold filter" mode, which at min_recall=1.0 returns only PASS misclassifications and zero NO_PASS rows. See Common pitfalls.

Setup

The threshold sweep, weakness ranking, and per-lighting expansion all run inside the tao_toolkit.data_services image declared in versions.yaml. Resolve the concrete URI once at the top of the run, then confirm Docker, the NVIDIA container toolkit, and a GPU are present and ensure the image is cached:

# Resolve tao_toolkit.data_services → concrete nvcr.io/... URI from versions.yaml
DS_IMAGE=$(python3 -c "import yaml,os; print(yaml.safe_load(open(os.environ['TAO_SKILL_BANK_PATH']+'/versions.yaml'))['images']['tao_toolkit']['data_services'])")
echo "DS_IMAGE=$DS_IMAGE"

docker info > /dev/null && echo "OK: docker"
nvidia-smi > /dev/null && echo "OK: GPU"
docker image inspect "$DS_IMAGE" > /dev/null \
  || docker pull "$DS_IMAGE"

TAO_SKILL_BANK_PATH is usually exported by the installed skill bank. If it is unset, point it at the skill-bank repo root before resolving. A GPU is required; aborting early on a GPU-less host saves a confusing late error.

Three setup rules are load-bearing and easy to get wrong:

• Path mounting — every host path the container reads or writes (inference.csv, train YAML, dataset image root, output dir) must be bind-mounted, simplest with -v $WORKSPACE:$WORKSPACE -w $WORKSPACE so absolute paths resolve identically on both sides.
• Do not pass --user $(id -u):$(id -g) — it triggers KeyError: 'getpwuid(): uid not found: <uid>' during the container's transformers import; chown outputs back to the host UID afterwards instead.
• -e <spec> is required, not optional — current images hard-require it and exit with ValueError: The subtask vcn_aoi requires the following argument: -e/--experiment_spec_file before parsing CLI overrides.

See references/container-setup.md for the full path-mounting pattern, the --user/chown rationale and alpine chown command, multi--v guidance, and the -e <spec> requirement detail.

Method

The whole skill is a single docker run invocation followed by a small visual spot-check. The container does Steps 1–4 internally (threshold sweep, weakness scoring, top-K selection, per-lighting expansion). You handle Step 5 (visual spot-check) directly with the Read tool.

Step 1–4 — Run the container

$DOCKER gap_analysis vcn_aoi \
    inference_results_dir=<exp_dir>/inference/<label>/ \
    train_config=<exp_dir>/train.yaml \
    kpi_media_path=<dataset_root> \
    results_dir=<rca_results_dir> \
    top_k_per_label=50

Always pass top_k_per_label. This is the argument that switches the container from the default "samples below threshold" filter into proper top-K-per-label ranking. At min_recall=1.0 the threshold is by construction at-or-below every NO_PASS score, so the below-threshold filter returns ONLY misclassified PASS rows and zero NO_PASS rows — useless as an augmentation queue. With top_k_per_label set to a positive integer (either in the spec or as a Hydra override), the container computes signed weakness against the threshold for every row and surfaces the K weakest per ground-truth label, which is the per-label ranked output downstream steps consume.

Reads inference.csv, sweeps every unique siamese_score plus one value just below the minimum, keeps the candidates with NO_PASS-class recall ≥ min_recall (with 1e-12 tolerance), then picks the threshold with the best F1 (tie-break: precision, then threshold value). For every row, computes signed weakness from that threshold (positive = misclassified, negative = correct, magnitude = margin). Sorts by weakness descending and takes the top top_k_per_label per ground-truth label, then expands each weak row into one row per lighting condition using dataset.classify.input_map and dataset.classify.image_ext from the train YAML.

If no candidate threshold meets the recall target, the container exits non-zero and writes unreachable_kpi.txt into results_dir explaining which recall the model can actually achieve. In that case, stop the analysis after the docker call, write a one-section report explaining the model fundamentally cannot reach the KPI at any operating point, and recommend retraining or relabeling — skip the visual spot-check.

Container writes into results_dir:

Print the container's stdout summary (chosen threshold, kept-row counts, per-label breakdown) to your own stdout so the script-check hook can verify the run produced output.

Step 5 — Visual spot check (small, fixed)

Skip this step if unreachable_kpi.txt exists. Otherwise use the Read tool to view the 5 weakest PASS samples and the 5 weakest NO_PASS samples from kpi_gaps.parquet (deduplicated to one row per sample, using the FIRST-lighting filepath), classify each as exactly one of mislabeled / edge case / data quality / systematic, and copy each viewed image (resized to 128×128 if PIL is available, otherwise just copy) into <results_dir>/rca_images/. This is the only image inspection required — do not view dozens of images, run failure mode clustering, or audit goldens (VCN has no golden images).

See references/visual-spot-check.md for the exact sample-selection sort, the per-lighting deduplication rule, the full definition of each verdict category, and the image-copy detail.

Reference invocation

Paste-and-edit the workspace, the four paths, and the two numeric knobs; this runs end-to-end. Capture stdout so the script-check hook sees row counts.

WORKSPACE=<absolute path>            # mounted identically inside the container
EXP_DIR=<experiment_result_dir>      # contains inference/inference.csv and train.yaml; must be inside $WORKSPACE
DATASET_ROOT=<dataset_root>          # image root for inference.csv input_path entries; must be inside $WORKSPACE
MIN_RECALL=1.0                       # zero-miss default; lower if KPI relaxes
TOP_K=50                             # per-label augmentation budget
OUT="$EXP_DIR/rca_results/$(date +%Y-%m-%d_%H%M%S)"
SPEC="$OUT/vcn_aoi_spec.yaml"
IMG=$(python3 -c "import yaml,os; print(yaml.safe_load(open(os.environ['TAO_SKILL_BANK_PATH']+'/versions.yaml'))['images']['tao_toolkit']['data_services'])")

mkdir -p "$OUT"

# Write the gap-analysis spec for this run
cat > "$SPEC" <<EOF
min_recall: $MIN_RECALL
top_k_per_label: $TOP_K
EOF

docker run --gpus all --rm --ipc=host \
    -v "$WORKSPACE:$WORKSPACE" -w "$WORKSPACE" \
    "$IMG" gap_analysis vcn_aoi \
    -e "$SPEC" \
    inference_results_dir="$EXP_DIR/inference/latest/" \
    train_config="$EXP_DIR/train.yaml" \
    kpi_media_path="$DATASET_ROOT" \
    results_dir="$OUT"

# Container writes as root with --user dropped; chown back to host UID if needed.
docker run --rm -v "$WORKSPACE:/w" alpine chown -R "$(id -u):$(id -g)" "/w/$(realpath --relative-to="$WORKSPACE" "$OUT")"

# Sanity print so the script-check hook sees real numbers
python3 - "$OUT" << 'PYEOF'
import json, os, sys
out = sys.argv[1]
unreachable = os.path.join(out, "unreachable_kpi.txt")
if os.path.isfile(unreachable):
    print("KPI UNREACHABLE — see", unreachable)
    sys.exit(0)
with open(os.path.join(out, "threshold.txt")) as f:
    print("threshold:", f.read().strip())
with open(os.path.join(out, "metrics.json")) as f:
    m = json.load(f)
print(f"precision={m['precision']:.4f} recall={m['recall']:.4f} f1={m['f1']:.4f}")
import pandas as pd
df = pd.read_parquet(os.path.join(out, "kpi_gaps.parquet"))
print(f"kpi_gaps.parquet: rows={len(df)}, cols={list(df.columns)}")
print(df['label'].value_counts())
PYEOF

Outputs

Write everything into a timestamped folder under the experiment result directory. The container's outputs go straight there; the visual spot-check writes rca_images/; any runtime packaging hook may add session/config capture artifacts after RCA_Report.md is written.

<experiment_result_dir>/rca_results/YYYY-MM-DD_HHMMSS/
├── RCA_Report.md              # Full gap analysis report (you write this)
├── kpi_gaps.parquet           # Container: top-K weakest per label, expanded per lighting
├── threshold.txt              # Container: chosen decision threshold (single float)
├── metrics.json               # Container: confusion matrix + per-label distribution stats
├── weak_samples_breakdown.txt # Container: per-label count/misclassified/marginal counts
├── unreachable_kpi.txt        # Container: ONLY when no threshold meets min_recall
├── rca_images/                # You: thumbnails of the 10 viewed weak samples
├── rca_config/                # Auto-copied by hook
└── session log/artifacts      # Optional, runtime-dependent packaging capture

At the start of the run, get the real timestamp by running date +%Y-%m-%d_%H%M%S in Bash. Do NOT hardcode or guess. If the user specifies a custom output path, use that instead but maintain the same internal structure.

Common pitfalls

The single most consequential failure mode is forgetting top_k_per_label when min_recall=1.0: at that recall the chosen threshold sits at or below every NO_PASS score, so without top_k_per_label the container falls back to a "samples below threshold" filter that returns ONLY misclassified PASS rows and zero NO_PASS rows, breaking the augmentation queue. Always include an explicit positive top_k_per_label (default 50) in the spec or as a Hydra override.

See references/pitfalls.md for the complete checklist, covering: forgetting top_k_per_label; passing --user; calling with only Hydra overrides (no -e <spec>); spec file outside $WORKSPACE; spec file with unresolved ??? sentinels; image not pulled / wrong tag; path-mount mismatch; unreachable_kpi.txt written; inference.csv missing required columns; train YAML missing dataset.classify.input_map or image_ext; kpi_media_path not matching input_path prefixes; and no GPU detected from inside the container.

Report Structure

Write RCA_Report.md as a tight (1000–1800 word) computational gap analysis — depth comes from accurate numbers and a clear action list, not narrative. The full report template (7 sections: Verdict, Threshold Selection, Weakness Distribution, Top-K Weakest Samples, Visual Spot Check, Per-Label Breakdown, Recommended Actions — with the confusion-matrix and table layouts) is in references/output-template.md. When unreachable_kpi.txt exists, replace sections 3–6 with a single short section quoting that file's contents and collapse section 7 to one recommendation: retrain or relabel.

Execution Order

1. Resolve DS_IMAGE from versions.yaml (images.tao_toolkit.data_services), then run docker info, nvidia-smi, and docker image inspect "$DS_IMAGE" (pulling if missing) once to confirm the environment. Abort with a clear message if any fail.
2. Run date +%Y-%m-%d_%H%M%S to get the timestamp; create <experiment_result_dir>/rca_results/<timestamp>/.
3. Write vcn_aoi_spec.yaml into the timestamped dir with min_recall and top_k_per_label filled in. Keep it under $WORKSPACE so the -e path resolves inside the container.
4. Run docker run … "$DS_IMAGE" gap_analysis vcn_aoi -e vcn_aoi_spec.yaml inference_results_dir=… train_config=… kpi_media_path=… output_dir=…. The container writes kpi_gaps.parquet, threshold.txt, metrics.json, weak_samples_breakdown.txt into results_dir. Print the chosen threshold and kept-row counts to stdout so the script-check hook can verify the run produced output.
5. If unreachable_kpi.txt exists, skip Step 6 and write the abridged report. Otherwise continue.
6. Pick 10 weak samples (5 weakest PASS + 5 weakest NO_PASS) from kpi_gaps.parquet, view each test image with Read, classify, and copy each into rca_images/.
7. Write RCA_Report.md last — writing it triggers the packaging hook, which copies session logs and skill config alongside.