flow_inference.infer_textlines.InferenceHandler

class flow_inference.infer_textlines.InferenceHandler(model, processor, device)[source]

Run TrOCR inference for line-level OCR/HTR records.

The handler wraps a loaded model, processor, and device. It prepares batches with line metadata, executes generation in inference mode, and returns decoded text predictions mapped back to project, filename, region, and line identifiers.

Parameters:

  • model (PreTrainedModel)

  • processor (TrOCRProcessor)

  • device (device)

__init__(model, processor, device)[source]

Initialize the inference handler.

Parameters:

  • model (PreTrainedModel) – Loaded TrOCR-compatible model used for generation.

  • processor (TrOCRProcessor) – Processor used to decode generated token IDs.

  • device (device) – Torch device used for inference.

Methods

__init__(model, processor, device)

Initialize the inference handler.

custom_collate_fn(batch)

Stack image tensors and preserve line-level metadata.

infer(records, image_handler, **kwargs)

Run inference for a collection of line-level records.

run_batch_inference(inference_dataloader, ...)

Run model generation for all batches in a DataLoader.