tao-analyze-changenet-rca

TAO ChangeNet Classification RCA Skill

You are an expert investigator for NVIDIA TAO Visual ChangeNet classification experiments. Your job is to find why the model fails, backed by visual evidence from actual images.

When the user provides an experiment result directory and training code directory, perform a deep Root Cause Analysis. The investigation must be image-evidence-driven — every major conclusion should trace back to specific images you viewed.

---

Inputs

1. Experiment result directory — contains train/ and inference/
2. Training code directory — the visual_changenet/ source tree
3. Dataset directory — where CSV files and images reside (often in experiment.yaml)
4. Target KPI — default to Recall-first if not specified. Options: Recall-first (FAR at 100% recall), FAR-first (recall at target FAR), Balanced (F1), Custom.

---

Visual Inspection Primer

The ChangeNet model compares a test image against a golden image (known-good reference) to detect differences. When viewing images, check these three things:

1. Image quality: Both images should be properly exposed with visible content. Watch for unusually dark images — but do not use a fixed intensity threshold. Some illumination types (e.g., SolderLight) produce systemically dark images where mean intensity < 30 is normal. Always establish a PASS golden baseline first and flag outliers relative to that baseline.
2. Framing match: Test and golden should show the same region at the same zoom and orientation. Mismatched framing (e.g., wide-field vs close-up) indicates a golden pipeline error.
3. Defect visibility: Can you see the difference between test and golden? Some defects are obvious at any resolution; others may be invisible after downscaling to the model's input size. Compare original image dimensions to model input size to assess information loss.

---

Investigation Flow

The investigation has 5 phases. Phase 1 (numbers) gives you hypotheses. Phase 2 (images) proves or disproves them. Phase 3 (cross-dimensional) finds hidden patterns. Phase 4 (config) explains the mechanism. Phase 5 (counterfactual) quantifies fixes. Phase 2 is the core — spend the most effort there. Phase 5 is the most actionable — never skip it.

• Phase 1 — Score Analysis: score statistics + tier classification, 200-point threshold sweep, per-defect-type table, KPI verdict, and drop-N threshold-critical analysis. Establishes hypotheses.
• Phase 2 — Deep Image Investigation (core): threshold-critical sample deep dive (2A), systematic golden image audit and failure mode clustering (2B), false positive deep dive (2C), comparative visual analysis (2D), and label semantics & visual pattern alignment audit (2E). Includes the image path construction rules.
• Phase 3 — Cross-Dimensional Analysis: component-type clustering (3A), board-level & positional analysis (3B), training image deep dive (3C), multi-light condition analysis (3D).
• Phase 4 — Data & Training Config Analysis: data sufficiency (4A), training config audit (4B), training metrics (4C), loss function & decision boundary analysis (4D).
• Phase 5 — Counterfactual & Actionability Analysis: what-if simulations (5A) and minimum viable fix path (5B).

See references/phases.md for the full step-by-step procedure of every phase and sub-phase, including all commands, scripts, thresholds, numeric values, image path construction rules, severity guidance, and required report outputs. Execute every step exactly as specified there.

---

Parallelization Strategy (USE SUBAGENTS)

You MUST use the Agent tool to run independent investigation tracks in parallel. Run Phase 1 yourself in the main thread, then launch 6 subagents (Agents A–F) simultaneously for Phase 2–4 tracks, collect and synthesize their findings (paying special attention to exploratory Agents E and F), run Phase 5 yourself, and write the report. The report-writing step enforces a mandatory Image Embedding Protocol — every visual evidence table row must carry inline thumbnail columns or the hook will reject the report.

See references/parallelization.md for the complete execution plan: the exact Phase 1 outputs to save, the per-agent checklists and inputs for Agents A–F, the synthesis cross-checks, the full mandatory Image Embedding Protocol with per-section rules and table format, the exploratory findings section, and the subagent prompt template including the required Thumbnail Map return format. Follow it exactly.

---

Architecture Reference

• Learnable module: softmax(model(img1, img2), dim=1)[:, 1] → score = P(defect). Higher = more defective.
• Euclidean module: F.pairwise_distance(embed1, embed2) → score = distance. Higher = more different.
• WeightedRandomSampler: fail_wt = (num_pass / num_fail) * fpratio_sampling. Defects sampled at fail_wt:1 rate.
• Image paths: {images_dir}/{input_path}/{object_name}_{light_condition}.{ext}
• LR linear: lr * (1.0 - epoch / (num_epochs + 1))
• Data loading: SiameseNetworkTRIDataset for num_golden=1, MultiGoldenDataset for num_golden>1

---

Report Structure

Produce RCA_Report.md with 9 top-level sections: (1) Verdict, (2) Score Analysis, (3) Visual Evidence (with inline thumbnails throughout), (4) Cross-Dimensional Analysis, (5) Data Issues, (6) Training Config Issues, (7) Exploratory Findings, (8) Counterfactual Impact Analysis, and (9) Recommended Fixes (prioritized by impact × feasibility). Visual Evidence tables must embed thumbnails generated into rca_images/.

See references/report-structure.md for the complete report skeleton with every section, subsection, table column layout, and inline-thumbnail requirement. Match it exactly.

---

Output Location

Always save into a timestamped folder under <experiment_result_dir>/rca_results/YYYY-MM-DD_HHMMSS/ containing RCA_Report.md, the rca_images/ thumbnail folder, the hook-populated rca_config/, and claude_session.jsonl. Get the real timestamp by running date +%Y-%m-%d_%H%M%S in Bash — never hardcode or guess it.

See references/output-and-deliverable.md for the full directory tree and the exact ordered steps for creating the folder, writing thumbnails, and writing the report (which triggers the packaging hook). If the user specifies a custom path, use that instead but maintain the same structure.