Recognize handwritten text in scanned documents using MultiDimensional Recurrent Neural Networks
Creates a network based on MultiDimensional RNNs architecture using python and cython with Connectionist Temporal Classification (CTC) cost function.
- Creates Multidimensional LSTM network.
- No need to extract features before feeding it to RNN or LSTM framework.
- The current configuration takes 2D input but can be extended to N-dimensional input.
- Uses forward backward algorithm with CTC loss function. This is taken from Andrew Mass stanford-ctc.
- Runs on Multi-Cores.
- Uses cython for fast execution.
- Installing Cython
- Installing dill. dill extends python’s pickle module for serializing and de-serializing python objects.
Downloading IAM dataset for handwriting recognition from IAM. To create data splits for training, validation and testing
python create_data.py path_to_xml_files path_to_words train_samples val_samples
path_to_xml_files: folder where xml files are stored, path_to_words: folder where images of handwritten words are stored, train_samples: no of training samples, val_samples: no of validation samples.
The IAM dataset contains 115149 images of words, so the following command will create 80k training_data 15k validation_data and 20k testing_data.
python create_data.py /home/xml_files/ /home/data/words/ 80000 15000
First create .so file, which will be used for calling cython functions.
python setup_cython_3.py build_ext --inplace
For training, run the following command
python train.py learning_rate momentum regularization update batch_size epochs
For example,
python train.py --learning_rate=0.001 --momentum=0.9 --reg=0.0 --update=rmsprop --batch_size=200 --epochs=50
Intial weights are initialized using xavier initialization. After every epoch parameters are saved using cPickle as model_parameters.
Run the following command
python test.py