RobertBiehl/CLIP-tf2
📦 Open Source ProjectRobertBiehl
A complete implementation of OpenAI's CLIP model ported to TensorFlow 2 and Keras.
CLIP-tf2 bridges the gap between PyTorch-centric research and TensorFlow-based production environments. By converting the original OpenAI CLIP architecture into Keras-compatible layers, this project allows practitioners to utilize pre-trained weights for tasks like image classification, semantic search, and zero-shot object recognition within the TensorFlow ecosystem. The implementation maintains the integrity of the original CLIP model, featuring the dual-encoder architecture consisting of a vision transformer (ViT) or ResNet image encoder and a Transformer-based text encoder. This repository is particularly useful for developers who require high-performance vision-language models but are constrained by infrastructure or project requirements that mandate the use of TensorFlow 2. It supports standard Keras model loading and inference workflows, making it a drop-in solution for integrating state-of-the-art multimodal capabilities into legacy or TF-native AI applications.
💡Highlights
- ├─Native Keras/TF2 implementation
- ├─Zero-shot image classification
- └─Supports ViT and ResNet backbones
🎯For
- ├─TensorFlow Developers
- └─Computer Vision Engineers