Context Compression using 2D Grids

This repository contains a project that compresses context using 2D grids. The main components are:

inference.py: This script demonstrates how to use the trained model for inference. It takes a source text and generates a visual representation of the compressed context using a canvas.

utils.py: Contains utility functions for training the model, computing embeddings, and encoding the context into grids/canvases.

Run the inference script:

python inference.py

This will load the pre-trained model, generate embeddings for the source text, and display the visual representation of the compressed context. You can input any text within the source variable. Additionally, you can set DEBUG to false if you wish to obtain the true grids rather than the human-optimized grids.

Saving Grids To save the generated grids for later use, you can modify the inference.py script to include:

torch.save(grids, 'path/to/saved/grids.pt')

Then, to load the saved grids:

grids = torch.load('path/to/saved/grids.pt')