Implement fullflow of attention ocr using Tensorflow's seq2seq API and tf.estimator

COCO Text dataset

Here

Source structure

Location	Content
/data	Fonts file, character table and corpus
train.py	Train script
word_generator.py	Utility to generate synthetic word from corpus
params.json	Configuration
'COCO-Text-words-trainval'	COCO text dataset for train and eval

Generated images from word corpus

Train OCR model

E.g:
python train.py --model_dir ./model_dir/ --coco_path ./COCO-Text-words-trainval/

Predict

Tensorflow serving is used to perform prediction for all models in this repo. Details of tensorflow serving is here

Firstly, we need to export the trained model to serving

python export.py --model_dir ./model_dir --export_dir ./serving/

Secondly, serve the model with docker, by default i only support restful api call

docker run -t --rm -p 8501:8501 -v "/home/nghi/ocr_attention/serving/:/models/ocr" -e MODEL_NAME=ocr tensorflow/serving &

Serving model

Here

Prediction

Coupons.

999999502

80.57