Haiku Generator

Description

This Repository contains an implementation of the SeqGan Architecture proposed in this Paper for generating Haikus. This Branch, master, generates word vectors using multivariate Gaussian Distributions. The charlevel branch is a somewhat working implementation of the original SeqGAN and Èmbeddings is similar to master but utilizes Deep Deterministic Policy Gradients (DDPG) in order to generate the word vectors. Because DDPG takes a long time to train and involves a second non-generator mode, the Critic, this Branch will likely be deleted in the future. This project is my submission for the BW-KI 2020 Competition.

Functionality

The SeqGAN Algorithm was first proposed by this Paper. Like in a normal GAN, the Discrimnator Model rates the samples produced by the Generator on a scale from zero to one based on how realistic they seem. However, since the samples are sequential and the Discriminator cannot judge partial Sequences, it is unable to directly provide a reward for each timestep to the Generator. This Problem is solved by using the classic RL-Algorithm REINFORCE as a Generator Model and rolling out the sequences using Monte-Carlo. One Downside of this Procedure is that a REINFORCE Agent can only act in a discrete environment since its output is a probability distribution over a finite number of actions. A somewhat working Implementation of the classic Algorithm can be found in the charlevel Branch. I am trying to extend the Generator into a continuous Action Space by having the Generator output a multivariate Gaussian Distribution. Exploration can be ensured by having the Model only output the mean but not the standard deviation for each of the num_actions distributions.

Usage

For the sake of keeping it simple, I am not uploading my dataset here. However, you can just use the dataset from this Repository and preprocess it using DatasetPreparation.ipynb. The resulting file should be placed in the data/ directory unless specified otherwise.

To make sure you have all the required dependencies installed, run pip install -r requirements.txt

The training is done in main.py. Run python3 main.py -h to see a list of all parameters you can set. Since i am not uploading my Word2Vec Model, you need to set the --pretrain_w2v parameter on your first run to train one yourself. Unless the path_model Parameter is set, the trained models will be stored in models/.

Future Goals

• Have the model output a full Covariance Matrix instead of only the diagonal values
• Use Pandas for handling the dataset (.csv instead of .txt)
• Maybe use something more advanced than REINFORCE
• Scrape Haikus from Twitter
• Improve charlevel branch
• Create contributing guidelines

Training Images

Credits

Though i plan to use my own data eventually, right now i am using a processed version of the dataset from this HaikuRNN Repository for training my model.

License

This program is licensed under the GNU General Public License v3.0. For more Information, see LICENSE.txt