This repository provides a complete implementation of the CoCa (Contrastive Captioner) architecture, which jointly trains an image encoder and text decoder using both contrastive and captioning losses. The model uses a vision transformer for image encoding and a PaLM-style decoder for text generation, enabling image-text understanding and captioning capabilities. It supports cross-attention layers for multimodal fusion and achieves state-of-the-art accuracy on ImageNet benchmarks.