Training

Training

Standard Training

flashocr train --config configs/flashocr_m_coco.yaml --device cuda

Config-driven Training

flashocr train --config configs/flashocr_m_coco.yaml
flashocr train --config configs/flashocr_m_lora.yaml
flashocr train --config configs/flashocr_m_kd.yaml

Training Options

Flag	Description	Default
--config	YAML config file	—
--model-size	Model variant (m, m-0.5x, m-1.5x)	m
--epochs	Training epochs	100
--batch-size	Batch size	64
--lr	Learning rate	0.001
--device	Device (cuda/cpu)	cuda
--amp	Mixed precision training	false
--multi-gpu	DataParallel	false
--pretrained	Load pretrained weights	false

Data Format

FlashOCR expects a TSV label file with tab-separated image paths and labels:

data/
├── train/
│   ├── images/
│   │   ├── img_001.jpg
│   │   └── img_002.jpg
│   └── labels.tsv
└── val/
    ├── images/
    └── labels.tsv

labels.tsv format:

img_001.jpg	hello
img_002.jpg	world
img_003.jpg	12345

Mixed Precision & Multi-GPU

flashocr train --config configs/flashocr_m_coco.yaml --amp --multi-gpu