OCRNet

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).

OCRNet for scene text recognition. Recognizes text content from cropped text region images. Supports CTC and attention-based decoders.

Set train.pretrainedmodelpath for pretrained OCR weights.

For TAO Deploy TensorRT actions (gentrtengine, TensorRT evaluate, and TensorRT inference), read references/tao-deploy-ocrnet.md first. Deploy spec templates live in this skill's references/ folder with the spectemplatedeploy_*.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/spectemplate<action>.yaml from the schema top-level default field. AutoML enablement is declared at the model layer in references/skillinfo.yaml via automlenabled. Runnable AutoML for an action requires schemas/<action>.schema.json and references/spectemplate<action>.yaml to exist and parse. Use the packaged selected-action schema for automldefaultparameters, automldisabledparameters, 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/skillinfo.yaml and resolve the run override from either an explicit automlpolicy value or the user's workflow request. Use automlpolicy: 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 automlpolicy: off for this run only. When automlpolicy: on, automlenabled: true, and both schemas/train.schema.json and references/spectemplatetrain.yaml are packaged, route the train action through tao-skill-bank:tao-run-automl by default with this model's skilldir. Preserve workflow/application overrides for datasets, specs, output directories, GPU/platform settings, parent checkpoints, and automlpolicy. 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: ocrnet
• Formats: default
• Monitoring metric: valacc1

Per-Action Dataset Requirements

Checkpoint Selection

OCRNet training writes both bestaccuracy.pth and epoch-step checkpoints such as modelepoch000step00003.pth. Use the SDK/model checkpoint resolver through the specparams mappings in references/skill_info.yaml; do not guess by sorting for the newest .pth.

• Use best_accuracy.pth for best-checkpoint evaluate, inference, export, and prune requests.
• Use the exact requested modelepochstep.pth for epoch/step-specific actions.
• Use train.resumetrainingcheckpointpath only for resume training, and use model.prunedgraph_path for retrain from a prune output. OCRNet does not expose a separate ocrnet retrain CLI subtask in the PyT image; the model-skill retrain action routes through ocrnet train -e with the pruned graph path set.
• OCRNet quantize loads the model through PyTorch. For trusted checkpoints created by the same local run, set TORCHFORCENOWEIGHTSONLY_LOAD=1 if PyTorch 2.6+ rejects the checkpoint as a weights-only load.

Typical Spec Overrides

Data source overrides are mandatory for every action. Run dataset_convert separately for train and validation splits, then pass the LMDB folders that directly contain data.mdb and lock.mdb into train, quantize, and retrain. Tarballs from remote storage must be extracted before they are used as image directories.

TRAIN_IMAGES = "<extracted train image folder>"
TRAIN_GT = "<train gt_new.txt>"
EVAL_IMAGES = "<extracted eval image folder>"
EVAL_GT = "<eval gt_new.txt>"
TRAIN_LMDB = "<train dataset_convert results_dir>"
EVAL_LMDB = "<eval dataset_convert results_dir>"
CHAR_LIST = "<character_list>"

dataset_convert (run once per split):

{
    "dataset_convert.input_img_dir": TRAIN_IMAGES,
    "dataset_convert.gt_file": TRAIN_GT,
}

train (mandatory data sources):

{
    "train.num_epochs": 30,
    "train.checkpoint_interval": 10,
    "train.validation_interval": 10,
    "train.num_gpus": 1,
    "dataset.batch_size": 16,
    "dataset.train_dataset_dir": [TRAIN_LMDB],
    "dataset.val_dataset_dir": EVAL_LMDB,
    "dataset.train_gt_file": "",
    "dataset.val_gt_file": "",
    "dataset.character_list_file": CHAR_LIST,
}

deploy/gentrtengine (mandatory data sources):

{
    "gen_trt_engine.onnx_file": "<selected export ONNX>",
    "gen_trt_engine.trt_engine": "<output engine path>",
    "gen_trt_engine.tensorrt.calibration.cal_cache_file": "<output calibration cache path>",
    "gen_trt_engine.tensorrt.data_type": "fp16",
    "gen_trt_engine.tensorrt.calibration.cal_image_dir": [TRAIN_IMAGES],
    "dataset.character_list_file": CHAR_LIST,
}

evaluate (mandatory data sources):

{
    "evaluate.checkpoint": "<selected train/AutoML checkpoint>",
    "dataset.character_list_file": CHAR_LIST,
    "evaluate.test_dataset_dir": EVAL_IMAGES,
    "evaluate.test_dataset_gt_file": EVAL_GT,
}

export (mandatory data sources):

{
    "export.checkpoint": "<selected train/AutoML checkpoint>",
    "export.onnx_file": "<output ONNX path>",
    "dataset.character_list_file": CHAR_LIST,
}

inference (mandatory data sources):

{
    "inference.checkpoint": "<selected train/AutoML checkpoint>",
    "dataset.character_list_file": CHAR_LIST,
    "inference.inference_dataset_dir": EVAL_IMAGES,
}

prune (mandatory data sources):

{
    "prune.checkpoint": "<selected train/AutoML checkpoint>",
    "prune.pruned_file": "<output pruned PTH path>",
    "dataset.character_list_file": CHAR_LIST,
}

quantize (mandatory data sources):

{
    "dataset.train_dataset_dir": [TRAIN_LMDB],
    "dataset.val_dataset_dir": EVAL_LMDB,
    "dataset.character_list_file": CHAR_LIST,
    "dataset.quant_calibration_dataset.images_dir": TRAIN_IMAGES,
    "quantize.model_path": "<selected train/AutoML checkpoint>",
}

retrain (mandatory data sources):

{
    "dataset.train_dataset_dir": [TRAIN_LMDB],
    "dataset.val_dataset_dir": EVAL_LMDB,
    "dataset.character_list_file": CHAR_LIST,
    "model.pruned_graph_path": "<selected prune output>",
}

Eval Dataset

Optional. Test data provided as separate tarball.

Important Parameters

• dataset.characterlistfile: Path to character list defining the supported character set. This determines the output vocabulary size.
• model.backbone: Default ResNet.
• model.prediction: Decoder type. CTC or Attn (attention-based).
• train.optim.lr: Learning rate. Default 1.0 (Adadelta optimizer). High default is specific to Adadelta.
• dataset.batch_size: Per-GPU batch size. Default 16.

Multi-GPU / Multi-Node

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

• Strategy: auto for single-GPU, reads train.distributed_strategy from config when multi-GPU
• No explicit num_nodes in train script — single-node oriented
• Lightweight model, single GPU typically sufficient

Hardware

Minimum 1 GPU(s), recommended 1 GPU(s). 8GB+ VRAM per GPU. OCR text recognition is lightweight. Single GPU is typically sufficient.

Error Patterns

datasetconvert required: If using raw images + gt files, run datasetconvert first to produce LMDB format.

datasetconvert output folder: Direct ocrnet datasetconvert writes data.mdb and lock.mdb directly under datasetconvert.resultsdir. Use that folder itself for dataset.traindatasetdir, dataset.valdatasetdir, quantize, and retrain inputs. SDK-backed runs may wrap the same LMDB folder inside job artifact directories; resolve the actual folder containing data.mdb and lock.mdb.

GT file BOM: Some text-recognition GT files can start with a UTF-8 BOM on the first filename. If dataset conversion logs a missing path with an invisible prefix before the first image name, strip the BOM from a local copy of the GT file before conversion or evaluation.

Character list mismatch: All characters in training data must be present in the character_list file.

Export/prune output fields required: export.onnxfile and prune.prunedfile must be writable output paths. These are declared in references/skill_info.yaml so SDK-backed model runs can create the paths automatically.

TensorRT lives in deploy: The PyT OCRNet CLI exposes datasetconvert, evaluate, export, inference, prune, quantize, and train, but not gentrtengine. Use references/tao-deploy-ocrnet.md and deploy/skillinfo.yaml for TensorRT engine generation and TensorRT-backed evaluate/inference.

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 createjob(). This mirrors the old microservices inferparams.py flow.

Inference mappings from TAO Core ocrnet.config.json:

For parentmodel or parentmodelfolder, pass the upstream train/export/AutoML child job id as parentjob_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-ocrnet](references/tao-deploy-ocrnet.md)