Introduction

keras-auto-hwr is a micro-framework for building a handwritten text recognition (HTR) system built on top of Keras and TensorFlow.

The central philosophy is to enable one to quickly train and test an RNN to perform HTR with minimum upfront efforts. One only needs to implement a thin API specifying how and where raw data for training should be fetched from. The framework's pipeline will automatically take care of steps such as pre-processing, normalization, converting target sequences into one-hot representation and more.

Features

• minimal pre-processing
• automatic data normalization
• automatic encoding of text into a sequence of one-hot vectors
• built-in model (bidirectional LSTM model trained with CTC loss)
• train on GPU, predict on CPU
• saving/resuming training after any epoch
• interactive demo for testing a trained model live
• language-agnostic (work in progress)
• deploy a trained model for use with TensorFlow.js
• built-in Token Passing decoding algorithm

Limitations

At the time of this writing, there are a few major limitations such as:

• No support for building an offline recognition system.
• Currently, the only model that is supported is the architecture based around connectionist temporal classification
• Fixed number of layers (2) and units (100)
• No built-in encoder-decoder network architecture

Installation

Clone the repository

git clone <clone_url>

, create a virtualenv environment using Python 3

virtualenv --python='/path/to/python3/executable' venv

, activate the environment

. venv/bin/activate

, go inside the directory containing a hidden git folder and install all python modules required for the app

pip install -r requirements.txt

Quick Start

Training from scratch on Iam Online DB data set

First, download the data set, unzip it and place it under ./datasets/iam_online_db/ folder. The layout of datasets folder should be as follows:

iam_online_db/

--ascii-all/

--lineImages-all/

--lineStrokes-all/

--original-xml-all/

--original-xml-part/

Compile the data set:

python compile.py 'IamSource' 'default' --num_examples=7000

This command will compile a data set using IAM-OnDB data set, default pre-processor. It will create training/validation/test data split containing a total of 7000 examples.

Train a bidirectional LSTM model with CTC loss:

python train_ctc.py --cuda=true

Set flag "cuda" to use CuDNNLSTM implementation instead of LSTM.

You can suspend the script after any epoch and resume training later.

When training is complete, run an OCR demo script:

python demo/ocr_demo.py

Finally, deploy the model to be used in javascript:

python deploy.py

Your deployed model will be in ./weights/deployed/blstm folder.

Training on arbitrary data set

Create a subclass of BaseSource class in the data/providers.py module by implementing constructor, get_sequences and len methods. You are free to do anything in those methods. For instance, you can read the data in from the file or fetch them through the network.

The get_sequences method should return a generator that returns raw handwriting data and corresponding transcription text. Each handwriting should be a list of strokes. Each stroke is itself a list of the following format: (x1, y1, t1), (x2, y2, t2), ..., where x and y are pen position coordinates and t is the time respectively. Add necessary pre-processing if you need to.

Here is an example.

class MyDataProvider(BaseSource):
    def __init__(self, num_lines):
        pass

    def get_sequences(self):
        hwr = [
            [(23, 8, 323), (25, 9, 325)], # first stroke
            [(55, 2, 340), (58, 2, 380)]  # second stroke
        ]

        transcription = 'foobar'

        yield hwr, transcription

    def __len__(self):
        return 1

Next, compile a data set using a newly implemented data provider:

python compile.py 'MyDataProvider' 'default' --num_examples=1

Testing pre-trained model

Run demo script using Best Path Decoding algorithm

python demo/ocr_demo.py

Run demo script using Token Passing Decoding algorithm

Change your current working directory to ./algorithms/cpp:

cd ./algorithms/cpp

Compile and link C++ code into binary:

./make_binary.sh

Return to the root path of the project

cd ../../

Build a dictionary and language model (e. g. containing 4000 words)

python build_lang_model.py dictionary --max_words=4000

Now you should be able to use Token Passing algorithm for decoding RNN output in a demo script. For that, simply set a "token_passing" flag to true:

python demo/ocr_demo.py --token_passing="true"

About pre-trained model

The model has 1 bidirectional LSTM layer with 100 hidden units followed by a softmax layer containing 100 units (one unit per character plus special codes and blank code). The model was trained on the IAM On-Line Handwriting Database using 9450 examples of the form: handwriting patterns -> text lines. It was trained with Adam optimizer with a learning rate of 0.001 and CTC loss. The training was terminated after 31 epochs.

Running tests

python -m unittest discover -s tests

License

This software is licensed under MIT license (see LICENSE).

References

[1] Alex Graves et. al. Unconstrained Online Handwriting Recognition with Recurrent Neural Networks

[2] S. Young, N. Russell, and J. Thornton. Token passing: A simple conceptual model for connected speech recognition system

[3] Liwicki, M. and Bunke, H.: IAM-OnDB - an On-Line English Sentence Database Acquired from Handwritten Text on a Whiteboard