tao-train-visual-changenet

Visual ChangeNet

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).

Visual ChangeNet is a TAO Toolkit model for visual inspection and defect detection. It supports two tasks:

• Classify — Binary image classification using a siamese-style architecture with a shared backbone (C-RADIO ViT) and a learnable difference module. Compares image pairs to classify defects as PASS/NO_PASS.
• Segment — Pixel-level change segmentation using a ViT-Large NVDINOv2 backbone. Compares before/after image pairs to produce a binary change mask.

The backbone weight (c_radio_v2_vit_base_patch16_224) is the public nvidia/C-RADIOv2-B model from HuggingFace, distributed as model.safetensors (~393 MB). The TAO container does not auto-fetch from HF URLs — ptm_utils.load_pretrained_weights() hands the pretrained_backbone_path value to torch.load(path) / safetensors.torch.load_file(path) directly. Passing an https://huggingface.co/... URL or a repo id produces FileNotFoundError and the run fails with Execution status: FAIL within a few seconds. Stage the file locally before launch with the bundled helper (idempotent — reuses an already-staged file):

python3 skills/models/tao-train-visual-changenet/scripts/stage_backbone.py --workspace <workspace>
# -> <workspace>/pretrained_models/C-RADIOv2_B.safetensors

This is a public download — no NGC CLI, no NGC org, and no credentials are required, and there is no ngc:// transfer-learning checkpoint dependency (VCN trains from this backbone). Run it in the CPU shell, where host network and HF_TOKEN live; HF_TOKEN is read only if set (gated mirror / rate limit) and is needed at staging time only, never inside the training container. Mount the staged file into the container (-v <workspace>/pretrained_models/C-RADIOv2_B.safetensors:/data/pretrained_models/C-RADIOv2_B.safetensors) and set the spec model.backbone.pretrained_backbone_path to the container path.

Segment specs use model.backbone.type: vit_large_nvdinov2 and the NVDINOv2 checkpoint family. Keep the checkpoint architecture aligned with the backbone type: NV_DINOV2_518_16_256.ckpt is compatible with the packaged segment templates, but it must not be used with fan_small_12_p4_hybrid. If you switch to a different segment backbone, use a matching checkpoint or leave model.backbone.pretrained_backbone_path empty for default initialization.

Dataclass Schemas

Generated TAO Core schemas are packaged in schemas/<action>.schema.json, with schemas/manifest.json listing available actions. Each generated schema also emits references/spec_template_<action>.yaml from the schema top-level default field. AutoML enablement is declared at the model layer in references/skill_info.yaml via automl_enabled. Runnable AutoML still requires schemas/train.schema.json and references/spec_template_train.yaml to exist and parse. Use the packaged train schema for automl_default_parameters, automl_disabled_parameters, defaults, min/max bounds, enums, option weights, math conditions, dependencies, and popular parameters. Do not expect ~/tao-core at runtime; maintainers regenerate schemas/templates before packaging the skill bank.

Train Action Policy

This model is AutoML-enabled at the model layer. Before handling any train-stage request, read references/skill_info.yaml and resolve the run override from either an explicit automl_policy value or the user's workflow request. Use automl_policy: on by default and only expose on / off in new launch prompts. Treat phrases like "turn off AutoML", "disable AutoML", "no HPO", or "plain training" as automl_policy: off for this run only. When automl_policy: on, automl_enabled: true, and both schemas/train.schema.json and references/spec_template_train.yaml are packaged, route the train action through tao-skill-bank:tao-run-automl by default with this model's skill_dir. Preserve workflow/application overrides for datasets, specs, output directories, GPU/platform settings, parent checkpoints, and automl_policy. Use direct model training only when automl_policy: off or the packaged train schema/template is missing; in the missing-schema case, report that AutoML is enabled but not runnable for this model until schemas are generated.

Checkpoint retention is an orchestration policy, not an HPO parameter. Both packaged train templates default train.checkpointer.enable_topk and train.checkpointer.replace_periodic to false, preserving periodic saves controlled by train.checkpoint_interval. When AutoML checkpoint retention is enabled, the AutoML runner sets both flags to true, monitors val_loss in min mode, and uses save_top_k: 1; this replaces the periodic series with the single best checkpoint. When AutoML checkpoint retention is disabled, leave the bounded-retention overrides unset so periodic checkpoint behavior remains.

Non-train actions declared by this model skill (evaluate, inference, export, quantize, segment_evaluate, and segment_inference) stay in this model skill. Do not present segment_export or segment_quantize as runnable parent-skill actions until matching entries are packaged in schemas/manifest.json. Prune and retrain are not declared in the current parent references/skill_info.yaml; do not present them as runnable parent-skill actions unless the metadata is extended with matching action wiring and schemas. The per-run automl_policy override does not change model metadata.

For TAO Deploy TensorRT actions (gen_trt_engine, TensorRT evaluate, and TensorRT inference for classify and segment variants), read references/tao-deploy-visual-changenet.md first. Deploy spec templates live in this skill's references/ folder with the spec_template_deploy_*.yaml prefix. Deploy requires an exported ONNX artifact as parent_model. If no ONNX artifact exists and the main skill does not expose an export action, report deploy as blocked instead of inventing an artifact.

Training Requirements

Visual ChangeNet has two separate task modes with different dataset types and data source structures.

Classify

• Dataset type: visual_changenet_classify
• Formats: default
• Accepted dataset intents: training, evaluation, testing, calibration
• Monitoring metric: val_loss

Per-Action Dataset Requirements (Classify)

The quantize and gen_trt_engine rows below describe TAO spec data requirements only. They are not parent-skill actions unless the corresponding action is declared in references/skill_info.yaml or deploy/skill_info.yaml.

Segment

• Dataset type: visual_changenet_segment
• Formats: default
• Accepted dataset intents: training, calibration
• Monitoring metric: val_loss

Segment uses a paired directory structure (A/, B/, list/, label/) instead of CSV + images. The root_dir spec key points to the top-level directory containing all four subdirectories.

Required files per dataset: A.tar.gz, B.tar.gz, list.tar.gz, label.tar.gz

Per-Action Dataset Requirements (Segment)

The quantize and gen_trt_engine rows below describe TAO spec data requirements only. They are not parent-skill actions unless the corresponding action is declared in references/skill_info.yaml or deploy/skill_info.yaml.

Typical Spec Overrides

Data source overrides are mandatory for every action — the agent MUST construct data source paths from the Per-Action Dataset Requirements table above and include them in spec_overrides.

S3_TRAIN = "s3://bucket/data/train"
S3_EVAL

train (classify, mandatory data sources):

train (segment, mandatory data sources):

export (classify):

export (segment):

quantize (classify, mandatory data sources):

evaluate (classify, mandatory data sources):

inference (classify, mandatory data sources):

gen_trt_engine (classify, mandatory data sources):

quantize (segment, mandatory data sources):

evaluate (segment, mandatory data sources):

inference (segment, mandatory data sources):

gen_trt_engine (segment, mandatory data sources):

Optional: running via the TAO SDK

When running without the TAO SDK (local docker), use the pinned TAO pyt image and invoke visual_changenet <train|evaluate|inference|export|quantize> directly. --shm-size=8g is required, the C-RADIO .safetensors must be mounted to /data/pretrained_models/C-RADIOv2_B.safetensors, and checkpoint/results_dir can be overridden on the command line. See references/local-docker.md for the full docker run command, mounts, and overrides.

Tasks

Classify (default)

Uses actions: train, evaluate, inference. Defaults template: references/spec_template_train.yaml. evaluate / inference need a checkpoint from a prior 7.1 train under results_dir — there is no pretrained 7.1 classify checkpoint on NGC (the 7.0-era visual_changenet_nvpcb_trainable_v1.0 fails to load on 7.1 with a radio.* KeyError). On a fresh workspace, train from the public backbone first; do not try to download an NGC full_model classify checkpoint, and do not hardcode an NGC org.

Segment

Uses skill action names segment_train, segment_evaluate, and segment_inference. When invoking local Docker directly, run TAO CLI subcommands train, evaluate, and inference with task: segment in the spec. The schema-driven action templates are references/spec_template_segment_train.yaml, references/spec_template_segment_evaluate.yaml, and references/spec_template_segment_inference.yaml; the compact direct-Docker example template is references/spec_template_segment.yaml.

Segmentation requires compiling custom CUDA ops (MultiScaleDeformableAttention) on first run, which takes ~5 minutes. The ViT adapter backbone uses these for multi-scale feature extraction.

Dataset structure for segmentation differs from classify — uses paired directories (A/, B/, list/, label/) instead of CSV files. See dataset.segment.root_dir in the defaults.

Data Format

Classify needs a 4-column CSV (input_path,golden_path,label,object_name) plus an images directory; segment uses a paired directory structure (A/, B/, list/, label/) under dataset.segment.root_dir instead of CSV. The image_ext field (default .jpg) must match the actual file extensions; if images are .png, set dataset.classify.image_ext: .png. Multi-lighting input is configured via dataset.classify.input_map (each lighting name maps to a channel index) with dataset.classify.num_input set to match. See references/data-formats.md for the per-field input tables (classify train/eval/inference, segment), CSV column semantics, lighting/path-concatenation conventions, the segment directory layout, and input_map/grid_map examples.

Pre-Flight: validate the classify dataset (mandatory before every classify run)

Before launching a classify train, evaluate, or inference job, validate the CSV so a malformed dataset fails in <1s on the host instead of minutes into the GPU container (or, for a single-class train set, only after a checkpoint is written). Run:

Exit 0 → launch. Exit 2 → fix the dataset, do not launch. The script rejects absolute CSV paths, flat filenames where a per-sample directory is required, single-class training sets, and a batch larger than the dataset. See references/data-formats.md for the per-check contract and the --light / --image-ext options.

Important Parameters

Key knobs include train.validation_interval (default 50, must be ≤ num_epochs), train.checkpoint_interval (default 200, must be ≤ num_epochs when periodic checkpointing is active), train.num_epochs (default 100), model.classify.eval_margin (default 0.3, the precision/recall threshold), model.classify.train_margin_euclid (default 2.0), model.classify.embedding_vectors (default 5), dataset.classify.batch_size (default 16, must be > 1), dataset.classify.fpratio_sampling (default 0.25), and train.classify.cls_weight (default [1.0, 10.0]). The train.checkpointer fields are fixed lifecycle controls, not HPO search parameters. Hardware: minimum 1 GPU with 16GB+ VRAM, recommended 8 GPUs (DDP); do not set gpu_spec_key (GPU count is managed internally by TAO), num_nodes (default 1) controls multi-node. See references/tuning-parameters.md for the full per-parameter guidance and hardware detail.

Error Patterns

For checkpoint-not-found, CSV format mismatch, image extension mismatch, OOM, low evaluation accuracy, the contrastive-loss AssertionError, checkpoint load key mismatch at evaluate/inference, non-convergence, segment-only backbone dimension mismatch, the MultiScaleDeformableAttention OSError, the Lightning MisconfigurationException, ModuleNotFoundError: nvidia_tao_pytorch, and epoch defaults, see references/troubleshooting.md for the full symptom-and-fix list.

Spec Param / Parent Model Inference

Model-specific parent-model mappings are declared in references/skill_info.yaml under spec_params, so generated runners and agents resolve checkpoints before create_job() instead of guessing file names. For parent_model or parent_model_folder, pass the upstream train/export/AutoML child job id as parent_job_id; the SDK lists the parent result folder, filters checkpoint artifacts, and returns the selected model file or folder. See references/parent-model-inference.md for the full per-action spec-field-to-inference-function mapping table.

Deployment

• tao-deploy-visual-changenet