CLI Reference

CLI Reference

Kiri OCR provides a comprehensive Command Line Interface (CLI) accessed via the kiri-ocr command.

predict

Run OCR inference on a document image.

kiri-ocr predict [IMAGE_PATH] [OPTIONS]

Arguments:

Argument	Description	Default
image	Path to the input image file (Required).	-
--mode	Detection mode: lines or words.	lines
--model	Path to model file or Hugging Face repo ID.	mrrtmob/kiri-ocr
--padding	Padding (pixels) around detected text boxes.	10
--output, -o	Directory to save results.	output
--no-render	Skip generation of visual reports (images/HTML).	False
--device	Compute device: cpu or cuda.	cpu
--verbose, -v	Enable verbose logging.	False

---

train

Train the Transformer-based recognition model.

kiri-ocr train [OPTIONS]

Key Arguments:

Argument	Description	Default
--train-labels	Path to training labels file (image_path \t label).	-
--val-labels	Path to validation labels file.	-
--hf-dataset	Hugging Face dataset ID (e.g., mrrtmob/km_en_image_line).	-
--output-dir	Directory to save model checkpoints.	models
--epochs	Number of training epochs.	100
--batch-size	Batch size.	32
--height	Input image height.	48
--width	Input image width.	640
--lr	Learning rate.	0.0003
--device	cuda or cpu.	cuda
--resume	Resume from the latest checkpoint in output dir.	False

Model Architecture Arguments:

Argument	Description	Default
--encoder-dim	Encoder hidden dimension.	256
--encoder-heads	Encoder attention heads.	8
--encoder-layers	Number of encoder layers.	4
--encoder-ffn-dim	Encoder feedforward dimension.	1024
--decoder-dim	Decoder hidden dimension.	256
--decoder-heads	Decoder attention heads.	8
--decoder-layers	Number of decoder layers.	3
--decoder-ffn-dim	Decoder feedforward dimension.	1024
--dropout	Dropout rate.	0.15

See Training Guide for full details and examples.

---

generate

Generate synthetic training data (images of text lines).

kiri-ocr generate [OPTIONS]

Arguments:

Argument	Description	Default
--train-file, -t	Input text file (one line per sample).	Required
--output, -o	Output directory for images/labels.	data
--fonts-dir	Directory containing .ttf fonts.	fonts
--augment	Number of augmented versions per line.	1
--random-augment	Apply random noise, rotation, blur.	False
--height	Output image height.	32
--width	Output image width.	512

---

init-config

Create a default YAML configuration file for training.

kiri-ocr init-config -o config.yaml

---

generate-detector

Generate a dataset for training the text detector (CRAFT).

kiri-ocr generate-detector [OPTIONS]

Arguments:

Argument	Description	Default
--text-file	Source text file.	Required
--fonts-dir	Directory of fonts.	fonts
--output	Output directory.	detector_dataset
--num-train	Number of training images.	800
--num-val	Number of validation images.	200

---

train-detector

Train the text detector model.

kiri-ocr train-detector [OPTIONS]

Arguments:

Argument	Description	Default
--data-yaml	Path to dataset YAML config.	detector_dataset/data.yaml
--epochs	Number of epochs.	100
--batch-size	Batch size.	16
--image-size	Image size for training.	640