joe-bender / philosophy-generator

Generate pseudo-philosophical ramblings using a neural network trained on the book The Critique of Pure Reason by Immanuel Kant.

Philosophy Text Generator

I created a text generator using a single book, The Critique of Pure Reason by Immanuel Kant, and a small neural network model trained for only a few epochs on short segments of the text.

Features

• a training loop created from scratch, implemented with a container class and many methods
• datasets, batch creation, tokenization, and numericalization from scratch using my own classes and methods
• creation of my own training samples from a raw source (a complete book) instead of a curated dataset
• a model created from the basic modules of PyTorch, implementing weight tying
• a custom text generation method, using an altered sotmax function with a customizable exponential base to tune the randomness of token selection
• a custom loss function and metric function, which ignore out-of-vocabulary tokens when evaluating the model's success

The main goal of this project was to build a model and training system from the most basic building blocks of PyTorch possible, without even using its built-in data loading functionality.

Although the text generation actually works pretty well with such a small model and short training period, the bulk of the work in this project was the surrounding infrastructure to allow for the training process to happen at all. I learned a lot about the lower level details of the full process of starting with a raw data source and ending up with a trained model that can do something useful.

Sample Philosophical Outputs

Here are a few examples of text generated by the trained model:

the conception of the understanding, in which the form of the understanding is not the cause of the soul in space, is only in the subjective conditions of time

but this is not a transcendental idea, which is not a thing in itself, but only in the sphere of experience

but this is not an object of the subject, and is not the objective validity of the possibility of a thing which is not an object

Running this project on your machine

This project has three main Jupyter notebooks, which need to be run in order. They will download the data (the book), train a model, and then generate some text. Random seeds are used throughout, so the results should be completely reproducible. The only requirements are PyTorch, NumPy, and Pandas.