Awesome

Hand Writing RNN

This RNN is a reconstruction of Alex Graves' work presented in his paper: http://arxiv.org/abs/1308.0850

alt tag

Dataset

Data was provided by the IAM On-Line Handwriting Database http://www.fki.inf.unibe.ch/databases/iam-on-line-handwriting-database

alt tag

Model

The model is precisely the same as the one presented in the paper above. Essentially, this is an RNN with three LSTM layers, a windowing layer that learns attention for each character in the input string, and a mixture of Gaussians criterion layer.

To train the model:

th driver.lua -inputSize INPUT_DIMENSION -hiddenSize LSTM_UNIT_SIZE 
-lr LEARNING_RATE -maxlen MAXIMUM_NUMBER_OF_TIMESTEPS 
-batchSize BATCH_SIZE_OF_DATA -numPasses NUMBER_OF_PASSES
-valData PATH_TO_VALIDATION_DATA -trainData PATH_TO_TRAINING_DATA

Explore driver.lua for command options. Including specifying a validation and training set. Currently toy.t7 is provided, but a larger training set can be created by downloading the IAM handwriting database and following the directions in Graves' paper: http://www.fki.inf.unibe.ch/databases/iam-on-line-handwriting-database

Results

Each multicolored line represents a new pen stroke.

th testNet.lua -string 'How are you today'

alt tag

th testNet.lua -string 'Testing testing'

alt tag

th testNet.lua -string 'this is awesome'

alt tag

Below is the attention plot which represents by how many timesteps the model will focus on a certain character in the string 'this is awesome' (moving in a left to right direction along each character in the string)

alt tag