RNNlib Dataset Generator for IAM Handwriting Database

This is a training data generator for RNNlib, a recurrent neural network implementation. For more information on the network, please refer to the RNNlib Wiki.

Prerequisites

In order to generate training data, please install Python, Pillow (PIL) and NetCDF. Currently, the generator is *nix-only.

Download the IAM Database

All test sequences can be downloaded from the IAM On-Line Handwriting Database. Please note, that all data belongs to the University of Bern and requires a login before downloading. This generator only works with strokes and images from the On-Line database, however it can easily be adapted to work with the Offline database aswell.

Please consider downloading the following files after registering here:

• ascii-all.tar.gz: Plain text transcriptions of all sequences.
• lineStrokes-all.tar.gz: XML files containing strokes with timestamps for online recognition.
• lineImages-all.tar.gz: TIF images for offline recognition.

After extracting all archives, the project folder should contain the following files:

ascii/
lineImages/
lineStrokes/
build_nc.sh
iam_offline.py
iam_online.py
...

After this, download the training set and validation set descriptors for the "Handwritten text recognition task IAM-OnDB-t1". These are textfiles that describe which sequences are used for which step in training the RNN. Place the text files in the same folder as the extracted archives:

ascii/
lineImages/
lineStrokes/
testset_f.txt
testset_t.txt
testset_v.txt
trainset.txt
...

Generate Training Files

The build_nc.sh utility creates NetCDF files compatible with RNNlib by processing every sequence in a file list. It takes two parameters:

1. mode: Either online (uses lineStrokes/) or offline (uses lineImages/).
2. dataset: The name of a configuration file, e.g. trainset or testset_v without the .txt file extension.

The dataset file contains a list of sample identifiers that is translated into a path depending on the mode parameter. In case one of the files is missing, the script will report an error and continue execution. For instance, to process all online samples use the following command:

$ /build_nc.sh online trainset
Loading labels
Processing sample a01/a01-020/a01-020x
...
Generating trainset.nc

The shell script uses either iam_online.py or iam_offline.py internally and creates several temporary files with a CDL extension. Those files are then merged into one CDL file and passed to NetCDF to create the training file of the same name. Please note, that the temporary CDL files might be slightly larger than the final NetCDF binary.

NOTE: In case of an error the CDL files remain in the file system and have to be cleaned up manually.

Training RNNlib

The configs folder contains sample configurations for both online and offline trainings. Copy one of these files to the project base directory and run:

/path/to/rnnlib --autosave=true <mode>.config

The sample configurations use trainset.nc for training and testset_v.nc and testset_t.nc for validation. RNNlib usually stops after 200 to 300 epochs with a sequence error rate of about 10%. It generates the following files:

online@2015.01.24-18.22.58.736519.best_ctcError.save
online@2015.01.24-18.22.58.736519.best_labelError.save
online@2015.01.24-18.22.58.736519.last.save
online@2015.01.24-18.22.58.736519.log

Each save file contains a complete network configuration including internal weights. Thus, they can be used to resume training at any point or for classification of new data. Please refer to the RNNlib Wiki for more information.