The original RNNLIB is hosted at http://sourceforge.net/projects/rnnl while this "fork" is created to repeat results for the online handwriting prediction and synthesis reported in http://arxiv.org/abs/1308.0850. The later by now is Alex Graves's classic paper on LSTM networks showing of what RNN can learn about the structure present in the sequential input.
The copy of this software runs on http://www.inkposter.com
Building rnnlib requires the following:
- C++11 compiler
- fortran for building OpenBLAS
- cmake
- libcurl
- automake
- libtool
- texinfo
In addition, the following python packages are needed for the auxiliary scripts in the 'utils' directory:
- SciPy
- PyLab
- PIL
And this package is needed to create and manipulate netcdf data files with python, and to run the experiments in the 'examples' directory:
- ScientificPython (NOT Scipy)
To build RNNLIB do
$ cmake -DCMAKE_BUILD_TYPE=Release . $ cmake --build .
Cmake run creates the binary files 'rnnlib', 'rnnsynth' and 'gradient_check' in the current directory.
It is recommended that you add the directory containing the 'rnnlib' binary to your path, as otherwise the tools in the 'utilities' directory will not work.
Project files for the integrated development environments can be generated by cmake. Run cmake --help to get list of supported IDEs.
Step in to examples/online_prediction and go through few steps below to prepare the training data, train the model and eventually plot the results of the synthesis
Start by registering and downloading pen strokes data from http://www.iam.unibe.ch/~fkiwww/iamondb/data/lineStrokes-all.tar.gz Text lables for strokes can be found here http://www.iam.unibe.ch/~fkiwww/iamondb/data/ascii-all.tar.gz Then unzip ./lineStrokes and ./ascii under examples/online_prediction. Data format in the downloaded files can not be used as is and requires further preprocessing to convert pen coordinates to offsets from previous point and merge them into the single file of netcdf format.
Run ./build_netcdf.sh to split dataset to training and validation sets. The same script does all necessary preprocessing including normalisation of the input and makes corresponding online.nc and online_validation.nc files for use with rnnlib .
Each point in the input sequences from online.nc consists of three numbers: the x and y offset from the previous point, and the binary end-of-stroke feature.
To gain some confidence that the build is fine run the gradient check:
gradient_check --autosave=false check_synth2.config
The training goes in two steps. First it is done without weights regularization and then repeated again with adaptive weight noise (MDL in rnnlib terms) from the best network recorded by step one. Training with MDL from the beginning will have too slow convergence rate.
rnnlib --verbose=false synth1d.config
Where synth1d.config is 1st step configuration file that defines network topology: 3 LSTM hidden layers of 400 cells, 20 gaussian mixtures as output layer, 10 mixtures for character warping window layer Somewhere between training epoch 10-15 it will find optimal solution and will do "early stopping" w/o improvement for 20 epoch. "Early" here takes 3 days on Intel Sandybridge CPU. Normally training can be stopped as long as loss starts rising up for 2-3 consequent epochs. The best solution found is stored in synth1d@.best_loss.save file
Best loss error from step 1 is expected to be around -1080 nats and it can be further improved (ca. 10%) by using weights regularisation. Loss error goes up and down during the training unlike in Step 1. Therefore one must be more patient to declare early stopping and wait for 20 epochs with loss worse then the best result so far. Rnnlib has implementation of MDL regulariser which is used in this step. The command line is as following:
rnnlib --mdl=true --mdlOptimiser=rmsprop from_step1.best_loss.save
Handwriting synthesis is done by rnnsynth binary using network parameters obtained by step 2:
rnnsynth from_step2.best_loss.save
The character sequence is given to stdin and output is written to stdout. The output sequence is the same as input where each data point has x,y offsets and end-of-stroke flag.
Rnnsynth output is the sequence of x,y offsets and end-of-stroke flags. To visualise it one can use show_pen.m Octave script:
octave:>show_pen('/tmp/trace1')
Where /tmp/trace1 contains stdout from rnnsynth.
Configuration options are exlained in http://sourceforge.net/p/rnnl/wiki/Home/. Since then there are few things added:
- lstm1d as hiddenType layer type - optimised LSTM layer when input dimension is 1d
- rmsprop optimizer type
- mixtures=N where N is number of gaussians in the output layer
- charWindowSize=N where N is the number of gaussians in the character window layer
- skipConnections=true|false - whether to add skip connections; default is true
Please create github issues to discuss the problems