Full PAIDF AnomalyGen pipeline — fine-tune on a new anomaly dataset, generate synthetic anomaly images (SDG), evaluate quality (nn_score), and search per-sample (guidance, crop_ratio) parameters. Three modes: full (Phase 0→7: finetune then generate), finetune_only (Phase 0→1: train only), inference_only (Phase 0, 2→7: generate from an existing checkpoint). Use when the user asks to "fine-tune AnomalyGen", "generate anomaly images", "run PAIDF SDG", "evaluate SDG output quality", "run per-sample search", or run any part of the AnomalyGen pipeline, even if they only mention one phase.
Instrucciones de origen · Vista previa de solo lectura
name
paidf-anomalygen
license
Apache-2.0
compatibility
Requires docker + nvidia-container-toolkit and a CUDA GPU. Pulls the `metropolis_sdg.paidf_anomalygen` image declared in `versions.yaml` at the skill bank root.
metadata
{"author":"NVIDIA Corporation","version":"1.0.1"}
allowed-tools
Read Bash
description
Full PAIDF AnomalyGen pipeline — fine-tune on a new anomaly dataset, generate synthetic anomaly images (SDG), evaluate quality (nn_score), and search per-sample (guidance, crop_ratio) parameters. Three modes: full (Phase 0→7: finetune then generate), finetune_only (Phase 0→1: train only), inference_only (Phase 0, 2→7: generate from an existing checkpoint). Use when the user asks to "fine-tune AnomalyGen", "generate anomaly images", "run PAIDF SDG", "evaluate SDG output quality", "run per-sample search", or run any part of the AnomalyGen pipeline, even if they only mention one phase.
tags
["tao","data"]
PAIDF AnomalyGen
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).
Multi-phase pipeline (0–7); the mode flag selects which phases run.
Assemble best-of-rounds into searched/ (stitch only), plus rounds/search_summary.csv
full, inference_only
7
Filter searched/ by nn_threshold (default 0.4), regen dropped samples, then canonical bucket eval → searched/{per_sample.csv, eval.log}
full, inference_only
Run every phase through to completion without mid-run pauses. Collect all
required parameters up front, and run every command from the repo root.
Shell setup. All ${ANOMALYGEN_SCRIPTS} references resolve to the packaged
helper-script directory. Inside the container this is preset (); on the host, export it once per
shell:
python3 -m scripts.utilities.<name> invocations work from any CWD inside the
container (PYTHONPATH is preset) and from the repo root on the host. When inside
a product container (ANOMALYGEN_PRODUCT_MODE=1), invoke anomalygen-guard
before any GPU work; if it reports BLOCKED, fix the listed issues before
continuing.
Quick Start
The pipeline runs inside the metropolis_sdg.paidf_anomalygen container
(declared in versions.yaml) or any host with the cosmos-predict2 conda env
active. All phase commands assume that environment, at the repo root, with
ANOMALYGEN_SCRIPTS exported.
Minimal end-to-end run (mode=full):
# 1. Set the shared variables (see "Shared variables" for the full set).export ANOMALYGEN_SCRIPTS="$(git rev-parse --show-toplevel)/scripts/utilities"
MODE=full
NAME=my_exp
DATASET_DIR=/data/uc1
DEFECT_DESC=assets/defect_spec_template.jsonl
NUM_SDG=20
MODEL_SIZE=2b
# 2. Phase 0 — verify / download checkpoints (~40 GB for the 2B default; needs HF_TOKEN).${ANOMALYGEN_SCRIPTS}/check.sh --model-sizes ${MODEL_SIZE^^} \
|| ${ANOMALYGEN_SCRIPTS}/download_checkpoints.sh --model-sizes ${MODEL_SIZE^^}# 3. Walk Phases 1→7 in order (see each Phase section).
For mode=inference_only (reuse a checkpoint) also set CKPT/STEP and skip
Phase 1. For mode=finetune_only run only Phases 0–1.
Running in Docker — container launch, mounts & permissions
The paidf-anomalygen image runs as a non-root baked-in user (USER anomalygen, uid=10000), independent of your host uid. Docker does not remap
uids on bind mounts, so a host directory owned by your uid is not writable by
uid 10000 and the container fails the instant it tries to create a file there.
Run as your host uid with --user "$(id -u):$(id -g)" plus the mandatory
/etc/passwd+/etc/group and HOME/cache-redirect companions, and run the
fail-fast write preflight before Phase 0. See references/docker.md for the
full docker run command, the load-bearing-flag table, the preflight snippet,
and the uid-10000 chown/chmod fallback.
Reference files — read before executing phases
Read references/finetune.md before Phase 0/1 and references/inference.md
before any of Phases 2–7; for mode=full read both before starting. The
remaining references below are on-demand — read when troubleshooting or needing
full detail for a specific phase.
File
Read when
references/finetune.md
Before Phase 0/1: env check, checkpoint download, dataset validation, config generation, training commands, best-checkpoint selection
references/finetune-commands.md
Exact Phase 1 Step 1–4 commands and CKPT/STEP derivation
references/inference-commands.md
Exact Phase 5 run_round.sh and Phase 7 filter_with_regen commands
num_SDG allocation depends on prep_testcase.sh --mode: inference (default,
Phase 2) is uniform across defect types, override per-defect via
--per-defect-counts; validation (Phase 1's validation JSONL) is proportional
to training mask counts (largest-remainder rounding) and enforces ≥1 per defect.
See references/prep-testcase.md for the full mode table.
Parameter
Description
mode
full (Phase 0→7), inference_only (skip Phase 1), or finetune_only (Phase 0→1 only).
name
Experiment label.
dataset_dir
Training/reference dataset root. Drives mask-count allocation, AMP submask templates, and holds semantic_segmentation_labels.json for cad defects.
defect_spec
JSONL tagging each defect spatial_dependency as free/text/cad. text entries need roi_prompt_defect_location. Template: assets/defect_spec_template.jsonl.
num_SDG
Total output samples per bucket. (Ignored when mode=finetune_only.)
Conditionally required
Parameter
Required when
Description
checkpoint_dir / step
mode=inference_only
Pre-existing fine-tuned model. In mode=full these are auto-derived after Phase 1; passing them is an error. In mode=finetune_only silently ignored — Phase 1 always trains from scratch (no resume-from-checkpoint support). Both must be present together — supplying only one is an error.
Optional parameters
Parameter
Default
Description
clean_dir
dataset_dir
Clean images. Set only when they live outside the training dataset. Forwarded as --clean-dir to prep-testcase and --clean-image-path to finetune.
validation_jsonl
auto-generated
Pre-built validation JSONL for Phase 1. When supplied, preflight verifies every defect_spec type appears and paths exist.
num_search_run
3
Per-sample search budget for Phase 5. 0 skips search (only original/). (Ignored when mode=finetune_only.)
nn_threshold
0.4
nn_score cutoff for Phase 7 (DINOv2 correspondence to real defects — key KPI). Samples below are regenerated; final searched/ always has num_SDG. 0 disables filtering.
mode=inference_only missing either checkpoint_dir or step → halt: "inference_only requires both checkpoint_dir and step."
mode=full with checkpoint_dir or step supplied → halt: "full mode runs finetune; use mode=inference_only to reuse an existing checkpoint."
Shared variables
Set once before Phase 0:
MODE=<full|inference_only|finetune_only>
NAME=<exp>
DATASET_DIR=<dataset_dir>
CLEAN_DIR=${clean_dir:-${DATASET_DIR}}
CKPT=<checkpoint_dir> # required iff MODE=inference_only; auto-derived after Phase 1 when MODE=full
STEP=<iter> # required iff MODE=inference_only; auto-derived after Phase 1 when MODE=full
NUM_SDG=<N>
DEFECT_DESC=<defect_spec.jsonl>
DEFECTS=(T+A T+B) # TEXTURE+TYPE names. For mode=inference_only, derive from ${CKPT}/ag_config.yaml → dataloader_train.dataset.anomaly_types (also printed by validate_checkpoint.py in Phase 0). For mode=full, take from DEFECT_DESC entries. See references/inference.md §Phase 0.
NUM_SEARCH_RUN=${num_search_run:-3}
NN_THRESHOLD=${nn_threshold:-0.4}
MODEL_SIZE=<2b|14b>
NUM_GPUS=${num_gpus:-1}
MAX_ITER=${max_iter:-75000}
SAVE_ITER=${save_iter:-5000}
VALIDATION_ITER=${validation_iter:-5000}
LR=${lr:-0.02}
BATCH_SIZE=${batch_size:-2}
IMAGE_SIZE=${image_size:-512}
VALIDATION_JSONL=${validation_jsonl:-}# optional; set by Phase 1 Step 2 if not user-supplied
BASE=results/${NAME}
JSONL=ag_inference/${NAME}/testcase.jsonl
ORIGINAL=${BASE}/original
SEARCHED=${BASE}/searched
ROUNDS=${BASE}/rounds
REGENS=${BASE}/regens
Guard preflight (product mode only)
When ANOMALYGEN_PRODUCT_MODE=1, run
.agents/skills/anomalygen-guard/scripts/preflight.py before any GPU work and
fix any BLOCKED issues. --validation-jsonl is forwarded only when the user
supplied one; for MODE=finetune_only omit --num-sdg if not supplied. See
references/guard-and-custom-counts.md for the full preflight command with all
forwarded flags and the validation-JSONL / allocate_samples.py 0-entry
checks.
Phase 0 — checkpoints
Read references/finetune.md §Phase 0 for HF_TOKEN requirements and what gets
downloaded. Both scripts default to the 2B base + t5-large (~40 GB); pass
--model-sizes ${MODEL_SIZE^^} so the chain checks and fetches the base size
this run actually uses (2b→2B, 14b→14B) — otherwise a 14b run
silently passes the 2B-only check and never downloads its checkpoint. Verify
first; download only what is missing.
Phase 1 — fine-tune (skip when MODE=inference_only)
Read references/finetune.md §Phase 1 for dataset structure, config template
details, and best-checkpoint selection. Four steps: (1) validate dataset /
derive anomaly types, (2) generate the validation JSONL (skip if user supplied
VALIDATION_JSONL), (3) generate the training config — show it to the user
and confirm before writing — (4) launch training in the background. Then
derive CKPT (path encodes upper-case MODEL_SIZE) and STEP (highest
nn_score step from validation logs). If MODE=finetune_only, stop after
training. See references/finetune-commands.md for the exact Step 1–4 commands
and the CKPT/STEP derivation snippet.
Phase 2 — prep-testcase (skip when MODE=finetune_only)
Read references/inference.md §Phase 2 for AMP routing detail and n_seeds
sizing. Do NOT pass --seeds — it is auto-computed and is not a recognized
flag. prep_testcase.sh defaults to --mode inference (uniform allocation
across defect types, no KPI floor), which Phase 2 always uses.
Custom per-defect counts: when the user specifies counts per defect type,
translate to --num-sdg plus a --per-defect-counts JSON dict (types absent
from the dict get 0; sum should equal --num-sdg, else the script warns on
stderr and uses the override sum). Confirm the allocation when intent is
ambiguous. See references/guard-and-custom-counts.md for the full
--per-defect-counts command example and the ambiguity-handling detail.
Phase 3 — SDG → original/
Read references/inference.md §Phase 3 for JSONL validation against the
checkpoint, multi-GPU caveats, and output verification.
Read references/inference.md §Eval for score interpretation and feature-count
explanation. run_eval.sh writes per_sample.csv and eval.log inside
original/ and merges nn_score into SDG_result.csv.
Read references/inference.md §Phase 5 for draw strategy, ranges, and re-AMP
guidance. For r in 1..NUM_SEARCH_RUN:
Read prior round's per_sample.csv (or ${ORIGINAL}/per_sample.csv for r=1).
Write ${ROUNDS}/round_${r}/draws.json with selected (guidance, crop_ratio) per sample.
Run round via ${ANOMALYGEN_SCRIPTS}/run_round.sh (SDG + eval; the round dir
gets its own sdg/{SDG_result.csv, per_sample.csv, eval.log}). See
references/inference-commands.md §Phase 5 for the full command and flags.
NUM_SEARCH_RUN=0 is valid — skip this phase entirely and let Phase 6
clone original/ into searched/.
Phase 6 — assemble searched/ (stitch only)
Always run assemble (works with 0 rounds — searched/ clones original/, so
downstream always reads searched/ regardless of num_search_run). Stitch-only:
copies winning images per sample-index into searched/ and carries over
per-sample nn_score / mnn_score from each pick's source-round per_sample.csv.
No eval — Phase 7 emits the canonical searched/eval.log.
Phase 7 runs by default (nn_threshold=0.4) on every mode=full and
mode=inference_only invocation; pass nn_threshold=0 to skip it. It filters
searched/ by nn_threshold, regenerates dropped samples via re-AMP (fresh
(clean, submask) pairing in the same defect type) for up to 5 attempts, then
falls back to best-scoring non-passing regens and finally to dropped originals,
so the final bucket always equals num_SDG.
Run python3 -m scripts.utilities.filter_with_regen. Pass --allocation ag_inference/${NAME}/allocation.json so regen targets the intended
per-defect counts — without it a bucket left short (e.g. by an interrupted SDG)
cannot be topped back up to num_SDG. It runs the final run_eval.sh
internally — the only eval against searched/. Read
references/inference.md §Phase 7 for regen mechanics, source-column tracing,
and the regens/regen_summary.csv schema; see
references/inference-commands.md §Phase 7 for the full command and flags.
Output layout
Every bucket that gets eval'd carries the same triad of files:
SDG_result.csv (generation params + nn_score + guardrail_pass),
per_sample.csv (per-sample nn + mnn), and eval.log (aggregate FID /
per-defect avg). Buckets live under results/<name>/ as original/ (Phase
3+4), searched/ (Phase 6 stitch + Phase 7 filter+regen+eval),
rounds/round_NN/ (Phase 5, plus search_summary.csv), and regens/regen_NN/
(Phase 7, plus regen_summary.csv).
Image content guardrail. A SigLIP content-safety check runs on every
generated image and records its verdict in SDG_result.csv.guardrail_pass
(1 safe / 0 blocked). A blocked image is replaced with an all-black
image that still occupies its slot on disk, so counting files cannot detect
it — never hand a guardrail_pass=0 sample downstream. Blacked-out samples
score near-zero nn_score, so Phase 7 regenerates them like any other
sub-threshold sample. Disable with ANOMALYGEN_IMAGE_GUARDRAIL=0.
See references/output-layout.md for the full directory tree with per-file
annotations, the guardrail semantics, and the post-run Verification
checklist (image counts per bucket, search_summary.csv /
regen_summary.csv row checks, the per-type nn_score / mnn_score / fid
fields in each eval.log, and the guardrail_pass sweep).
Error handling
Common pipeline failure modes (missing mask dirs, short/empty AMP output and
the 0 entries written halt, mid-round SDG failure resume, off-boundary
step, validate_dataset.py's non-zero exit on any pairing issue, the
2B-only check.sh default, and guardrail-blocked black images) are covered in
references/error-handling.md; see also references/finetune.md and
references/inference.md for phase-specific error handling.
