An exploration into AI-generated choose your own adventure stories. This project is heavily inspired by AI Dungeons, an AI based choose your own adventure game created by Latitude. The goal of this project is to build off of the gameplay in AI Dungeons and add other features such as inventory, health, and narrative consistency.
To run the code you will need the following packages:
Clicking any of the names above will bring you to the relavent instrucitions to install the package for your given platform.
Currently there is no offical dataset or model parameters for the game however, you can make your own! Create a .csv file in the data folder with two columns, source and target. This file will contain your dataset. In the source column place all the prompts you expect to feed your network. In the target column, write out the outputs you expect from the model. These columns should be written in standard english, no fancy formatting needed! Feel free to include punctuation as well. To train a new model open the training.py file and scroll to the very bottom. There should be a line of code inside an if statement which instantiates a new training object. Replace the current parameters used to instantiate the model with the name of the new .csv file you've made and a new name you'd like to call your model. An example is shown bellow:
if __name__ == "__main__":
training = Training("MY_DATA.csv", "MY_MODEL")Once you've filled out that line, run the training.py file and your new model will appear in the model folder!
To run the model open the adventures.py file and replace the parameters used to instantiate the model in line 22 with the new parameters you just made. An example is shown bellow:
self._model = Model("MY_MODEL", "MY_DATA.csv")Then feel free to replace the current prompt used to start the conversation and you're ready to go! The prompt can be found in the last line of the file (line 55). An example of how this should be done is shown bellow:
game.run("This is my new prompt!")To run the program, in a terminal run the adventures.py file and you'll begin having a conversation with your language model!!!
The architecture for this project is losely based off of an MVC architecture. A model class keeps track of game state. It updates the game state using a transformer based recurrent neural network to generate text given the previous game state. The neural network is contained in a separate class to make it easier to alter the text generation without affecting the rest of the game. The other two classes, view and controller, are very simple. The view class prints text to the terminal while the controller prompts the user to input text into the terminal. The reason these classes are kept seperate despite their simplicity is that they will most likely represent two different components in the final version of the game which will hopefully be hosted on the website. A simple class structure is defined bellow.
└── Adventure [Main class]
└── View [Displays game to user]
└── Controller [Gather input from user]
└── Model [Keeps track of game state]
└── Transformer [Pytorch transformer language model]
└── Generator [Standard linear layer + softmax to generate output]
└── Embeddings [Embed text as vector]
└── PositionalEncoding [Encode position data in input vector]
└── Encoder [Stack of encoder layers]
└── EncoderLayer [A single encoder layer]
└── SublayerConnection [Intermidate connection before a LayerNorm]
└── LayerNorm [A normilization layer]
└── MultiHeadedAttention [Contains multiple attention functions]
└── PositionwiseFeedForward [Feed forward network]
└── LayerNorm [A normilization layer]
└── Decoder [Stack of decoder layers]
└── DecoderLayer [A single decoder layer]
└── SublayerConnection [Intermidate connection before a LayerNorm]
└── LayerNorm [A normilization layer]
└── MultiHeadedAttention [Contains multiple attention functions]
└── PositionwiseFeedForward [Feed forward network]
└── LayerNorm [A normilization layer]
└── Training [Train the language model]
└── Batch [Holds batches of source and target data for training]
└── NoamOpt [Optim wrapper that implements changing rate]
└── LabelSmoothing [Reduce model's confidence]
└── SimpleLossCompute [Calculate incorrectness of model for training]
└── AdventureCollate [Restructure batches for use in training]
└── AdventureDataset [Load training and vocab data from csv]
└── Vocabulary [Hold model vocabulary]