/text-to-weights

Produce neural network weights from a provided text

Primary LanguagePython

Text to Neural Network Weights

I generated this program with the help of AI to create model weights from Ulises Carrión's 1975 text "The New Art of Making Books." Read more here. ~MS

A Python tool that generates deterministic neural network weights from input text. The weights are initialized using Xavier/Glorot initialization and can be saved in multiple formats.

Installation

Requirements

  • Python 3.x
  • NumPy
  • Pandas (for CSV output)

Install the required packages:

pip3 install numpy pandas

Usage

Basic Usage

python3 text_to_weights.py <input_text_file> [--format FORMAT] [--output-dir DIR]

Arguments

  • input_text_file: Path to the text file to generate weights from
  • --format: Output format (choices: 'npy', 'txt', 'csv', default: 'npy')
  • --output-dir: Output directory (default: 'weights')

Output Formats

The program supports three output formats, organized in format-specific subdirectories:

  1. NPY Format (weights/npy/)

    • Binary NumPy format
    • Most efficient for loading in Python
    • Example: weights/npy/layer_0.npy

  2. TXT Format (weights/txt/)

    • Plain text, comma-separated values
    • Human-readable
    • Example: weights/txt/layer_0.txt

  3. CSV Format (weights/csv/)

    • Comma-separated values with headers
    • Compatible with spreadsheet software
    • Example: weights/csv/layer_0.csv

Examples

Generate weights in NPY format:

python3 text_to_weights.py input.txt

Generate weights in CSV format:

python3 text_to_weights.py input.txt --format csv

Use custom output directory:

python3 text_to_weights.py input.txt --format txt --output-dir my_weights

Output Structure

The program generates three weight matrices with the following dimensions:

  • Layer 1: 128 x 64
  • Layer 2: 64 x 32
  • Layer 3: 32 x 16

Each layer's weights are initialized using Xavier/Glorot initialization for better neural network training characteristics.

Loading the Weights

# Load NPY format
import numpy as np
weights = np.load('weights/npy/layer_0.npy')

# Load TXT format
weights = np.loadtxt('weights/txt/layer_0.txt', delimiter=',')

# Load CSV format
import pandas as pd
weights = pd.read_csv('weights/csv/layer_0.csv').values

Features

  • Deterministic weight generation (same input text produces same weights)
  • Multiple output formats for different use cases
  • Xavier/Glorot initialization for better neural network training
  • Organized directory structure by format
  • Configurable layer sizes

Testing

Run the test suite:

python3 -m unittest test_text_to_weights.py -v