nieshaoshuai/Seq2seqAdapt_paper

Sequence-to-sequence Domain adaptation

for robust text image recognition

Prerequsites

tensorflow-gpu == 1.10.0 python 3.6.3 python Besides, we use python package distance to calculate edit distance for evaluation.

• Tensorflow: Installation Instructions
• Distance (Optional):

wget http://www.cs.cmu.edu/~yuntiand/Distance-0.1.3.tar.gz

tar zxf Distance-0.1.3.tar.gz

cd distance; sudo python setup.py install

Dataset

For a toy sample, we can download the following datasets.

Source synthetic text images

• A subset of Synth 90k:

wget http://www.cs.cmu.edu/~yuntiand/sample.tgz

tar zxf sample.tgz

Target text images

• Real scene text images from ICDAR03, ICDAR13, IIIT5k and SVT:

wget http://www.cs.cmu.edu/~yuntiand/evaluation_data.tgz

tar zxf evaluation_data.tgz

Usage:

(0) Preparing dataset

-Suppose DATA_HOME=/home/data/OCR

  python gen_tfrecord.py

(1) Pretraining a source model

   python main_baseline.py --phase='train'

(2) Training a domain adaptation model

   python main.py --phase='train'

Parameters:

Parameters for dataset

vi defaults_dataset.py

• DATA_HOME: The base directory of the dataset, default is '/home/data/OCR/'
• DATASET : the name of dataset, eg. svt, sample

Parameters for training

vi defaults.py

• DATA_HOME: The base directory of the dataset, default is '/home/data/OCR/'

• SOURCE_DATASET : the name of source dataset, eg. sample

• TARGET_DATASET : the name of target dataset, eg. svt

• MODEL_DIR : the directory of model

• OUTPUT_DIR : output directory

• LOG_PATH : the logging file