We built a beer review text generator in pytorch based on the dataset from https://arxiv.org/abs/1303.4402 using several different recurrent neural network topologies:
- LSTM
- RNN
- GRU
- Bi-directional LSTM
The network was trained using teacher forcing
Once trained, the network produces reviews based on a beer style, rating, and a given temperature for the softmax alphabet output. The networks were trained and using 1-2 1080-TI GPUs and tested using various temperatures and model sizes.
- main.py
Trains or generates some output from the network
- configs.py
Controls all the parameters of the training/generation
- models.py
implementations of different network types
- results
Example results at different temperatures
- weights
Weights for the trained networks
The original paper: https://arxiv.org/pdf/1511.03683.pdf which has a demo available at http://deepx.ucsd.edu/#/home/beermind