# Origin

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

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.

 
# Handwriting synthesis

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

## Downloading online handwriting dataset

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.

## Preparing the training data

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.

## Gradient check

To gain some confidence that the build is fine run the gradient check:

gradient_check --autosave=false check_synth2.config

## Training

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.

### Step 1

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@<time step>.best_loss.save file

### Step 2

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

### Synthesis

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.

### Plotting the results

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.

### Rnnlib configuration file

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

# Contact

Please create github issues to discuss the problems