Tensorflow implementation of handwritten sequense of small letter recognition.

• The handwritten dataset used is IAM
• The wrapper for code was taken from youtube-8m contest.
• The lstm2d.py and lstm1d.py was taken from Tensorflow contrib.
• The input from training and testing are stored in tfrecods files.

Dependencies

Python 2.7 Tensorflow 1.1

In order to run the training with multidimentional lstm:

python train.py --slices 26 --width 12 --stride 3 --Bwidth 90 --train_data_pattern ../tf-data/handwritten-test-{}.tfrecords --train_dir separable_lstm --test_data_pattern ../tf-data/handwritten-test-{}.tfrecords  --max_steps 6000 --batch_size 20 --beam_size 3 --input_chanels 1 --model MDLSTMCTCModel --base_learning_rate 0.001 --num_readers 2 --export_model_steps 500 --display_step 10 --display_step_lme 100 --start_new_model

options for training

• --slices: number of slices
• --width: width of the window
• --stride: step for
• --Bwidth: image width
• --train_data_pattern ../tf-data/handwritten-test-{}.tfrecords
• --train_dir separable_lstm
• --test_data_pattern ../tf-data/handwritten-test-{}.tfrecords
• --max_steps 6000
• --batch_size 20
• --beam_size 3
• --input_chanels 1
• --model: the class from handwritten models.py
• --base_learning_rate 0.001
• --num_readers 2
• --export_model_steps 500
• --display_step 10
• --display_step_lme 100
• --start_new_model
• --hidden: lstm number of neurons
• --layers: number of layers of lstm cell

In order to make tfrecord

• download the images(whole images) and xml from IAM web site (one have to register first)
• change the path to the xml folder and images folder in the jupyter notebook

In order to see statistics in tensorboard:

tensorboard --logdir=separable_lstm --port=8080

label rate error for test images

ctc loss for test images

Inference

• In order to see some prediction example checkout the jupyter notebook