Tyrion - Word Embedding Autoencoder
An autoencoder to calculate word embeddings as mentioned in Lebret/Collobert paper 2015
Runs the model mentined in http://arxiv.org/abs/1412.4930 on set of text files stored in directory on GPU
With GPU support code is 5-8x faster than it's CPU version.
Requirements:
-
Numpy http://www.numpy.org/
Installation
To install the package follow the instructions mentioned below
Install from source
To install the package from source run setup.py file, it will install and include necessary files in system python directory
$ python setup.py install
Installtion from pip
Tyrion can also be install from pip. Run following command from console
$ pip install tyrion
Training
To train the model on bunch of text files run following commnands in python console:
There is a sample data folder in project's folder, that can be used for sanity checking
>> from tyrion import train
>> train.train('/path/to/text/corpus')
Setting hyperparameters
Different hyperparameters are set by default in training module, to set hyperparamters manually use following command while training >> from tyrion import train
>> train.train('/path/to/text/corpus', contextSize=size, min_count=count, newdims=dims, ntimes=ntimes, lr=learningrate)
Arguments explained:
-
ContextSize is the contex window size used for constructing coocurence matrix
-
min_count is minimum threshold frequency for words, removes garbage words below this frequency
-
newdims is the dimension of embedding desired.
-
ntimes is number of epochs for training model
-
lr is the learning rate to drive optimization routine
Utility functions
To generate word embeddings and to find closest words to a word, use utils module in tyrion. Ex.
>> from tyrion import utils
>> embedding = utils.gen_embedding('word')
>> close_words = utils.closest_words('word',n=10)
Future work
- Implement a closely related paper on phrase embeddings (http://arxiv.org/abs/1506.05703) (ICML 2015)
- Try implementing AdaGrad ( for optimization )