AltCLIP

名称 Name	任务 Task	语言 Language(s)	模型 Model	Github
AltCLIP	text-image representation	中英文 Chinese&English	CLIP	FlagAI

简介 Brief Introduction

我们提出了一个简单高效的方法去训练更加优秀的双语CLIP模型。命名为AltCLIP。AltCLIP基于 Stable Diffusiosn 训练，训练数据来自 WuDao数据集 和 LIAON

AltCLIP模型可以为本项目中的AltDiffusion模型提供支持，关于AltDiffusion模型的具体信息可查看此教程 。

模型代码已经在 FlagAI 上开源，权重位于我们搭建的 modelhub 上。我们还提供了微调，推理，验证的脚本，欢迎试用。

We propose a simple and efficient method to train a better bilingual CLIP model. Named AltCLIP. AltCLIP is trained based on Stable Diffusiosn with training data from WuDao dataset and Liaon.

The AltCLIP model can provide support for the AltDiffusion 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}
}

训练 Training

训练共有两个阶段。 在平行知识蒸馏阶段，我们只是使用平行语料文本来进行蒸馏（平行语料相对于图文对更容易获取且数量更大）。在双语对比学习阶段，我们使用少量的中-英 图像-文本对（一共约2百万）来训练我们的文本编码器以更好地适应图像编码器。

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 bilingual comparison learning phase, we use a small number of Chinese-English image-text pairs (about 2 million in total) to train our text encoder to better fit the image encoder.

下游效果 Performance

Language	Method	Text-to-Image Retrival			Image-to-Text Retrival			MR
		R@1	R@5	R@10	R@1	R@5	R@10
English	CLIP	65.0	87.1	92.2	85.1	97.3	99.2	87.6
	Taiyi	25.3	48.2	59.2	39.3	68.1	79.6	53.3
	Wukong	-	-	-	-	-	-	-
	R2D2	-	-	-	-	-	-	-
	CN-CLIP	49.5	76.9	83.8	66.5	91.2	96.0	77.3
	AltCLIP	66.3	87.8	92.7	85.9	97.7	99.1	88.3
	AltCLIP∗	72.5	91.6	95.4	86.0	98.0	99.1	90.4
Chinese	CLIP	0.0	2.4	4.0	2.3	8.1	12.6	5.0
	Taiyi	53.7	79.8	86.6	63.8	90.5	95.9	78.4
	Wukong	51.7	78.9	86.3	76.1	94.8	97.5	80.9
	R2D2	60.9	86.8	92.7	77.6	96.7	98.9	85.6
	CN-CLIP	68.0	89.7	94.4	80.2	96.6	98.2	87.9
	AltCLIP	63.7	86.3	92.1	84.7	97.4	98.7	87.2
	AltCLIP∗	69.8	89.9	94.7	84.8	97.4	98.8	89.2

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.

模型推理 Inference

Please download the code from FlagAI AltCLIP

from PIL import Image
import requests

# transformers version >= 4.21.0
from modeling_altclip import AltCLIP
from processing_altclip import AltCLIPProcessor

# now our repo's in private, so we need `use_auth_token=True`
model = AltCLIP.from_pretrained("BAAI/AltCLIP")
processor = AltCLIPProcessor.from_pretrained("BAAI/AltCLIP")

url = "http://images.cocodataset.org/val2017/000000039769.jpg"
image = Image.open(requests.get(url, stream=True).raw)

inputs = processor(text=["a photo of a cat", "a photo of a dog"], images=image, return_tensors="pt", padding=True)

outputs = model(**inputs)
logits_per_image = outputs.logits_per_image # this is the image-text similarity score
probs = logits_per_image.softmax(dim=1) # we can take the softmax to get the label probabilities