Generating Sentences Using LSTM Networks
Generating English sentences using Long Short-Term Memory (LSTM) Networks
Try it yourself!Generating English Sentences Using LSTMs.
The goal of this work is to model the English language as closely as humanly possible through the use of Long Short-Term Memory (LSTM) Networks. To achieve this, every word is transformed into a feature vector, which is used as input; and the output is the next most probable word to follow. As a result, the network will produce a sequence of words, which should resemble a coherent English sentence after enough training. The network will learn these probabilities through the use of a large list of words; more specifically, the literature novel “Twelve Short Mystery Stories” by Sapper (Ronald Standish) 1.
Running the Code
Dependencies
Python 3.6.1Tensorflow 1.3.0Keras 2.1.2matplotlib 2.0.2numpy 1.12.1
Command-line Execution
Directory Path: /src/scripts/
-
Train the model:
python train.py [-h] [-v] [--data DATA] [--epochs EPOCHS] [--temp TEMP] [--slen SLEN] [-win WIN]Optional arguments:
-h,--helpshow help message and exit -v,--versionshow program’s version number and exit --data DATApath to training data (default: sample_data_short.txt)--epochs EPOCHSnumber of epochs to train for (default: 50) --temp TEMPselect temperature value for prediction diversity (default: 1.0) --slen SLENmaximum length for a training sequence (default: 15) -win WINselect sliding window for text iteration and data collection for training (default: 3) -
Make predictions:
python predict.py [-h] [-v] [--data DATA] [--seed SEED] [--nwords NWORDS] [--temp TEMP] [--slen SLEN] [-win WIN]Optional arguments:
-h,--helpshow help message and exit -v,--versionshow program’s version number and exit --data DATApath to training data (default: sample_data_short.txt)--seed SEEDprovide a word or sentence to seed the program with --nwords NWORDSnumber of words to generate (default: 400) --temp TEMPselect temperature value for prediction diversity (default: 1.0) --slen SLENmaximum length for a training sequence (default: 15) -win WINselect sliding window for text iteration and data collection for training (default: 3)
1: This novel is available on the Project Gutenberg website.