Thai-Cross-CLIP

Thai CLIP text encoder model trained via Teacher Learning

Overview

This model is a Thai text encoder which is a part of CLIP. This model is trained via Teacher Learning method using OpenAi's CLIP as Teacher

This repo contain only Text encoder which is compatible with ViT-B/32 OpenAi's Image encoder. To use this model as CLIP, you need to clone this repo (text encoder) and OpenAi's CLIP (image encoder). Look furthermore in Tutorial notebook Here!

Demo

[TBA]

Pre-trained Models

This Text encoder is trained on 2M Thai captions translated from English by AiResearch's MT model using WangchanBERTa as a pretrained model with an additional linear layer on top.

Name	Model Base	Vision Model	Vision Dimensions	#Parameters
WangchanBERTa ViT-B/32	WangchanBERTa	OpenAI ViT-B/32	512	106 M

other relate repo(s)

• VQGAN-ThCLIP: Text-to-image synthesis model in Thai language

Acknowledgements

• AI Builders for providing knowledge and support along the way
  
• Multilingual-CLIP for Teacher learning method
  
• OpenAI's CLIP
  
• AIResearch's translation model