We propose a simple and efficient method to train a better multilingua CLIP model. Named AltCLIP-m9. AltCLIP-m9 is trained with training data from WuDao dataset and Liaon.
The AltCLIP-m9 model can provide support for the AltDiffusion-m9 model in this project. Specific information on the AltDiffusion model can be found in this tutorial.
The model code has been open sourced on FlagAI and the weights are located on modelhub. We also provide scripts for fine-tuning, inference, and validation, so feel free to try them out.
引用
关于AltCLIP，我们已经推出了相关报告，有更多细节可以查阅，如对您的工作有帮助，欢迎引用。
If you find this work helpful, please consider to cite
@article{https://doi.org/10.48550/arxiv.2211.06679,
doi = {10.48550/ARXIV.2211.06679},
url = {https://arxiv.org/abs/2211.06679},
author = {Chen, Zhongzhi and Liu, Guang and Zhang, Bo-Wen and Ye, Fulong and Yang, Qinghong and Wu, Ledell},
keywords = {Computation and Language (cs.CL), FOS: Computer and information sciences},
title = {AltCLIP: Altering the Language Encoder in CLIP for Extended Language Capabilities},
publisher = {arXiv},
year = {2022},
copyright = {arXiv.org perpetual, non-exclusive license}
}
There are two phases of training.
In the parallel knowledge distillation phase, we only use parallel corpus texts for distillation (parallel corpus is easier to obtain and larger in number compared to image text pairs). In the multilingual comparison learning phase, we use a small number of text-image pairs (about 6 million in each language) to train our text encoder to better fit the image encoder.
下游效果 Performance
可视化效果 Visualization effects
基于AltCLIP，我们还开发了AltDiffusion模型，可视化效果如下。
Based on AltCLIP, we have also developed the AltDiffusion model, visualized as follows.
1from PIL import Image
2import requests
34# transformers version >= 4.21.05from modeling_altclip import AltCLIP
6from processing_altclip import AltCLIPProcessor
78# now our repo's in private, so we need `use_auth_token=True`9model = AltCLIP.from_pretrained("BAAI/AltCLIP-m9")10processor = AltCLIPProcessor.from_pretrained("BAAI/AltCLIP-m9")1112url ="http://images.cocodataset.org/val2017/000000039769.jpg"13image = Image.open(requests.get(url, stream=True).raw)1415inputs = processor(text=["a photo of a cat","a photo of a dog"], images=image, return_tensors="pt", padding=True)1617outputs = model(**inputs)18logits_per_image = outputs.logits_per_image # this is the image-text similarity score19probs = logits_per_image.softmax(dim=1)# we can take the softmax to get the label probabilities
