Tao Train Image Classification

Classification PyT

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).

PyTorch image classification. Supports a wide range of backbones (FAN, EfficientNet, ResNet, etc.) with distillation and quantization for deployment.

Set model.backbone.pretrained_backbone_path for backbone weights or train.pretrained_model_path for full model.

For TAO Deploy TensorRT actions (gen_trt_engine, TensorRT evaluate, and TensorRT inference), read references/tao-deploy-image-classification.md first. Deploy spec templates live in this skill's references/ folder with the spec_template_deploy_*.yaml prefix.

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 for an action requires schemas/<action>.schema.json and references/spec_template_<action>.yaml to exist and parse. Use the packaged selected-action 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.

Non-train actions such as evaluate, inference, export, and deploy flows stay in this model skill. The per-run automl_policy override does not change model metadata.

Training Requirements

• Dataset type: image_classification
• Formats: classification_pyt
• Monitoring metric: val_acc_1
• AutoML train scope: exclude distill.teacher.* parameters from ordinary train searches. Those fields belong to the separate distill action and do not change classification_pyt train behavior even though they appear in the combined generated schema.

Per-Action Dataset Requirements

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.

TRAIN_IMAGES_DIR = "/workspace/data/extracted/train/images_train"
VAL_IMAGES_DIR = "/workspace/data/extracted/val/images_val"
TEST_IMAGES_DIR = "/workspace/data/extracted/test/images_test"
CLASSES_FILE = "/workspace/data/s3/classes.txt"
WRITABLE_RESULTS_DIR = "/results"   # must be a writable bind, not the image CWD

For local Docker, download the S3 archives, extract them first, and point dataset.*.images_dir at the extracted class-root folder. Do not pass images_train.tar.gz, images_val.tar.gz, or images_test.tar.gz directly to local Docker specs; the skill metadata declares these inputs as folders.

train (mandatory data sources):

{
    "train.num_epochs": 2,
    "train.validation_interval": 2,
    "train.checkpoint_interval": 2,
    "train.num_gpus": 1,
    "dataset.train_dataset.images_dir": TRAIN_IMAGES_DIR,
    "dataset.classes_file": CLASSES_FILE,
    "dataset.val_dataset.images_dir": VAL_IMAGES_DIR,
    "dataset.root_dir": WRITABLE_RESULTS_DIR,
}

dataset.root_dir is mandatory for train, not just for export. CLDataset writes a classes.txt into root_dir during setup, so it must point at a writable bind mount (the results mount is the natural choice). The spec template defaults it to '', which makes the write land in the container working directory — that succeeds under a root container but fails under any runner that drops to a non-root user:

PermissionError: [Errno 13] Permission denied: 'classes.txt'
  .../classification_pyt/dataloader/dataset.py, in CLDataset.__init__

The local Docker launcher uses the host UID:GID whenever a writable results bind is present, so non-root training hits this unless dataset.root_dir is set.

export (mandatory data sources):

{
    "export.input_height": 224,
    "export.input_width": 224,
    "dataset.root_dir": "/workspace/data/extracted",
}

gen_trt_engine:

{
    "gen_trt_engine.tensorrt.data_type": "fp16",
}

inference (mandatory data sources):

{
    "dataset.batch_size": 1,
    "dataset.val_dataset.images_dir": VAL_IMAGES_DIR,
    "dataset.classes_file": CLASSES_FILE,
    "dataset.test_dataset.images_dir": TEST_IMAGES_DIR,
}

distill (mandatory data sources):

{
    "dataset.train_dataset.images_dir": TRAIN_IMAGES_DIR,
    "dataset.classes_file": CLASSES_FILE,
    "dataset.val_dataset.images_dir": VAL_IMAGES_DIR,
    "train.optim.policy": "step",
}

evaluate (mandatory data sources):

{
    "dataset.val_dataset.images_dir": VAL_IMAGES_DIR,
    "dataset.classes_file": CLASSES_FILE,
    "dataset.test_dataset.images_dir": TEST_IMAGES_DIR,
}

quantize (mandatory data sources):

{
    "dataset.train_dataset.images_dir": TRAIN_IMAGES_DIR,
    "dataset.classes_file": CLASSES_FILE,
    "dataset.val_dataset.images_dir": VAL_IMAGES_DIR,
    "dataset.quant_calibration_dataset.images_dir": TRAIN_IMAGES_DIR,
}

Eval Dataset

Optional. Validation images are provided as a separate tar alongside training images. For small smoke datasets that do not provide a separate images_test.tar.gz, set dataset.test_dataset.images_dir to the validation archive so evaluate and inference still exercise the checkpoint handoff.

Important Parameters

• dataset.num_classes: Number of classes. Default 20. Must match the number of subdirectories in your image tarballs.
• model.backbone.type: Default fan_small_12_p4_hybrid. Supported backbones and their head in_channels (from model_params_mapping.py): FAN: fan_tiny, fan_small_12_p4_hybrid, fan_base_16_p4_hybrid, fan_large_16_p4_hybrid. GCViT: gcvit_tiny through gcvit_large. FasterViT: fastervit_0 through fastervit_6. ViT/EVA/DINO: vit_large_patch14_dinov2, eva02_large_patch14, etc. SigLIP-CLIPA: ViT-H-14-SigLIP-CLIPA-224, etc. Some backbones require non-default input resolution (384, 512, 768).
• dataset.classes_file: Path to classes.txt listing class names.
• train.optim.lr: Learning rate. Default 6e-5.
• dataset.img_size: Input image size. Default 224.
• dataset.batch_size: Per-GPU batch size. Default 8.

Multi-GPU / Multi-Node

Launch method: Lightning-managed (single python process, Lightning spawns workers).

• Multi-GPU strategy: ddp_find_unused_parameters_true
• No fsdp support

Multi-node env vars (set by orchestrator): WORLD_SIZE, NODE_RANK, MASTER_ADDR, MASTER_PORT, NUM_GPU_PER_NODE.

Hardware

Minimum 1 GPU(s), recommended 2 GPU(s). 16GB+ (V100 or A100) VRAM per GPU. Classification is generally lightweight. Most backbones at 224x224 fit well on 16GB GPUs with batch_size=8.

Error Patterns

CUDA out of memory: Reduce batch_size or use a smaller backbone.

num_classes mismatch: Ensure dataset.num_classes matches the actual class directories in your image tarballs and classes.txt.

Empty class directory: Every class in classes.txt must have at least one image in the corresponding subdirectory.

Distill scheduler default: The bundled distill template and schema use train.optim.policy: step. Keep that setting for distill specs unless the container implementation is updated; the 7.0 PyT distiller does not assign a scheduler interval for train.optim.policy: linear.

Checkpoint handoff: Training produces model_epoch_*.pth checkpoints and a classifier_model_latest.pth symlink. For evaluate, inference, export, quantize, distill, and resume, select the exact intended epoch checkpoint through the SDK resolver; use the latest symlink only when the user explicitly requests latest.

Spec Param / Parent Model Inference

Model-specific inference mappings belong in this MD file, not in config.json. Generated runners should read this section and apply the mappings with SDK helpers before create_job(). This mirrors the old microservices infer_params.py flow.

Inference mappings from TAO Core classification_pyt.config.json:

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. Do not add these mappings back to config.json and do not patch generated runner scripts to guess checkpoint paths.

Deployment

• tao-deploy-image-classification