Backend: The package uses a Deep LSTM to generate presidential speeches. Developed all in Python. We also provide a HMM option, which is considerably faster to train. The comparison of generated text is shown below in the #comparison section.
Frontend: The speech is fed through AWS to Alexa for text-to-speech. Developed in Javascript.
The scripts included in the backends can be run individually without having a front-facing interface. We will demonstrate how to install all the dependencies.
The backend is developed in Python, and we will be using pip for installing that.
The system is assumed to have installed CUDA (v8.0), cuDNN, Keras (version 1.2.2) with mxnet (pip install mxnet-cu80==0.11.0) backend.
pip install tqdm
pip install autocorrect
pip install nltk
pip install numpy
git clone https://github.com/hi-im-ryanli/deep-president.git
Run the code with our pretrained model inside of models.
python pipeline.py
Based on your GPU specs, training might be potentially a long wait. We trained our model on AWS p2.16xlarge and the scaling is awesome. The network defined here has around 10,000,000 parameters to train, and we are feeding it with around 1,500,000 training samples.
You can modify our model architecutre based on your hardware. Fewer layers and smaller unit sizes will speed up the training process, with the trade of losing expressive power on the LSTM networks.
The network training is done inside of deep-text-char.py, and modify the network as you see fit in LSTM().
Compile and run your final model like this:
ryan = DeepText(folderName='democrat')
raw = ryan.LoadText()
ryan.GenrateData(raw)
model = ryan.LSTM()
ryan.Compile(model)
Add additional notes about how to deploy this on a live system
-
LTSM model: Keras, Python
-
Data scraping: BeautifulSoup, Python
-
Server: Amazon EC2 (AWS), Flask
-
User Interfaces: React, HTML, CSS, Amazon Alexa Echo
- Aidan Curtis
- Titus Deng
- Ryan Li
- Linus Shih
- Gary Wang
This project is licensed under the MIT License - see the LICENSE.md file for details