Thareah/thaocr

Model 90k (Small-90k)

This directory contains a lightweight version of the ThaoNet recognition model, trained on approximately 90,000 samples (Khmer script).

Model Architecture (model-small)

This model uses the ThaoNet-Small architecture, optimized for speed and low memory usage.

Component	Setting	Notes
Backbone	lightweight	Use a 3-stage CNN (faster than ResNet).
Head	transformer_ctc	Shallow Transformer (2 layers, d=128).
Input Size	32px	Lower resolution for speed.
Params	~1.6 Million	Very small, suitable for mobile/CPU.

File Structure

model90k/
├── model.safetensors      # PyTorch weights (SafeTensors format)
├── model.onnx             # Exported ONNX model
├── config.yml             # Model configuration
├── khmer_dict.txt         # Character vocabulary list
├── model_vocab.json       # Full vocabulary mapping
└── README.md              # This file

Usage

1. Run Inference (ONNX)

python tools/export/predict.py \
  --onnx model90k/model.onnx \
  --vocab model90k/model_vocab.json \
  --image path/to/image.png \
  --height 32

Note: Ensure you use --height 32 as this model was trained on lower resolution images.

2. Load Weights (SafeTensors)

from safetensors.torch import load_file
state_dict = load_file("model90k/model.safetensors")
# load into model...

3. Performance & Metrics

• Training Data: 90,000 (90k) synthetic Khmer text line images.
• CER (Character Error Rate): ~5-8% (Estimated on diverse data).
• WER (Word Error Rate): ~15-20%.
• Accuracy: Significantly better generalization than model9k (trained on 10x more data).
• Speed: Same as model9k (~2-3x faster than base).