Qsevent77/jina-clip-implementation

Jina CLIP

Core implementation of Jina CLIP. The model uses:

• —the EVA 02 architecture for the vision tower
• —the Jina XLM RoBERTa with Flash Attention model as a text tower

Models that use this implementation

• —jinaai/jina-clip-v2
• —jinaai/jina-clip-v1

Requirements

To use the Jina CLIP source code, the following packages are required:

• —torch
• —timm
• —transformers
• —einops
• —xformers to use x-attention
• —flash-attn to use flash attention
• —apex to use fused layer normalization