FAQ - Frequently Asked Questions¶

Q: Where does MyOCR look for models by default?¶

A: The default path is configured in myocr/config.py (MODEL_PATH) and usually resolves to ~/.MyOCR/models/ on Linux/macOS. Pipeline configuration files (myocr/pipelines/config/*.yaml) reference model filenames relative to this directory. You can change MODEL_PATH or use absolute paths in the YAML configuration if you store models elsewhere.

Q: How do I switch between CPU and GPU inference?¶

A: When initializing pipelines or models, pass a Device object from myocr.modeling.model.

For GPU (assuming CUDA is set up): Device('cuda:0') (for the first GPU).
For CPU: Device('cpu').

Ensure you have the correct onnxruntime package installed (onnxruntime for CPU, onnxruntime-gpu for GPU) and compatible CUDA drivers for GPU usage.

Q: The `StructuredOutputOCRPipeline` isn't working or gives errors.¶

A: This pipeline relies on an external Large Language Model (LLM).

Check Configuration: Ensure the myocr/pipelines/config/structured_output_pipeline.yaml file has the correct model, base_url, and api_key for your chosen LLM provider (e.g., OpenAI, Ollama, a local server).
API Key: Make sure the API key is correctly specified (either directly in the YAML or via an environment variable if the YAML points to one, like OPENAI_API_KEY).
Connectivity: Verify that your environment can reach the base_url specified for the LLM API.
Schema: Ensure the Pydantic json_schema passed during initialization is valid and the descriptions guide the LLM effectively.

Q: What's the difference between a Predictor and a Pipeline?¶

A:

Predictor: A lower-level component that wraps a single Model with its specific pre-processing and post-processing logic (defined in a CompositeProcessor). It handles one specific task (e.g., text detection).
Pipeline: A higher-level component that orchestrates multiple Predictors to perform a complete workflow (e.g., end-to-end OCR combining detection, classification, and recognition). Pipelines provide the main user-facing interface for common tasks.

Q: How can I use my own custom models?¶

A:

ONNX Models: Place your .onnx file in the model directory and update the relevant pipeline configuration YAML file (myocr/pipelines/config/*.yaml) to point to your model's filename. See the Overview section.
Custom PyTorch Models: Define your model architecture using components from myocr/modeling/ (backbones, necks, heads) and create a YAML configuration file specifying the architecture. Load it using ModelLoader().load(model_format='custom', ...) or create a custom pipeline/predictor. See the Models Documentation for details on CustomModel and YAML configuration.