TextBoxes re-implementation using tensorflow.
Much more info can be found in Textbox and SSD
This project is greatly inspired by slim project
And many functions are modified based on SSD-tensorflow project
Now the pipeline is much clear and can be resued in any tf projects.
Author: Daitao Xing : [email protected] Jin Huang : [email protected]
2017/ 03/14
data_processing phase finished Test：
1. Download the dataset， put 1/ folder and gt.mat uner ddata/sythtext/ folder（will wirte script）
2. python datasets/data2record.py
3. python image_processing.py
output： batch_size * 300 * 300 * 3 image
2017/ 03/17
Finish the design of training(can start training)
TASET_DIR=./data/sythtext/
TRAIN_DIR=./logs/
python Textbox_train.py \
--train_dir=${TRAIN_DIR} \
--dataset_dir=${DATASET_DIR} \
--save_summaries_secs=60 \
--save_interval_secs=600 \
--weight_decay=0.0005 \
--optimizer=adam \
--learning_rate=0.001 \
--batch_size=2 \
--gpu_data=/cpu:0 \
--gpu_train=/cpu:0
2017/ 03/29
Overwrite all files, so make the training pipeline much clear.
1. Write the load_batch . This can be resued in any preproceesing jobs.
2. Rewrite the traning file, so make the pipeline more clear.
To generate tfrecords for both test and training sets for ICDAR2013 dataset:
cd datasetsGenerate tfrecord for test set:
python ICDAR2013ToRecord.py --dataset=test --ground_truth_path_test=../data/ICDAR2013/ICDAR-Test-GT/Generate tfrecord for training set:
python ICDAR2013ToRecord.py --dataset=train --ground_truth_path_train=../data/ICDAR2013/ICDAR-Training-GT/The generated tfrecords are in TextBoxes-TensorFlow/data/ICDAR2013 folder.
1. The loss decreases slowly after 2000 iterations, find why?
2. Prepare the other two datasets(transform into tf.record)
3. Evaluation part scripts.