ZH-CLIP: A Chinese CLIP Model

Models

You can download ZH-CLIP model from ? thu-ml/zh-clip-vit-roberta-large-patch14 . The model structure is shown below:

• Vision encoder network structure is the same as openai/clip-vit-large-patch14 , and initialize with laion/CLIP-ViT-L-14-laion2B-s32B-b82K .
• Text encoder network struceure is the same as hfl/chinese-roberta-wwm-ext-large and initialized.

Results

COCO-CN Retrieval (Official Test Set):

Model	Text-to-Image	Image-to-Text
R@1	R@5	R@10	Mean	R@1	R@5	R@10	Mean
Clip-Chinese	22.60	50.04	65.24	45.96	22.8	49.8	64.1	45.57
mclip	56.51	83.57	90.79	76.95	59.9	87.3	94.1	80.43
Taiyi-CLIP	52.52	81.10	89.93	74.52	45.80	75.80	88.10	69.90
CN-CLIP	64.10	88.79	94.40	82.43	61.00	84.40	93.10	79.5
altclip-xlmr-l	62.87	87.18	94.01	81.35	63.3	88.3	95.3	82.3
ZH-CLIP	68.00	89.46	95.44	84.30	68.50	90.10	96.50	85.03

Flickr30K-CN Retrieval (Official Test Set):

Model	Text-to-Image	Image-to-Text
R@1	R@5	R@10	Mean	R@1	R@5	R@10	Mean
Clip-Chinese	17.76	40.34	51.88	36.66	30.4	55.30	67.10	50.93
mclip	62.3	86.42	92.58	80.43	84.4	97.3	98.9	93.53
Taiyi-CLIP	53.5	80.5	87.24	73.75	65.4	90.6	95.7	83.9
CN-CLIP	67.98	89.54	94.46	83.99	81.2	96.6	98.2	92.0
altclip-xlmr-l	69.16	89.94	94.5	84.53	85.1	97.7	99.2	94.0
ZH-CLIP	69.64	90.14	94.3	84.69	86.6	97.6	98.8	94.33

Muge Text-to-Image Retrieval (Official Validation Set):

Model	Text-to-Image
R@1	R@5	R@10	Mean
Clip-Chinese	15.06	34.96	46.21	32.08
mclip	22.34	41.15	50.26	37.92
Taiyi-CLIP	42.09	67.75	77.21	62.35
cn-clip	56.25	79.87	86.50	74.21
altclip-xlmr-l	29.69	49.92	58.87	46.16
ZH-CLIP	56.75	79.75	86.66	74.38

Zero-shot Image Classification:

Model	Zero-shot Classification (ACC1)
CIFAR10	CIFAR100	DTD	EuroSAT	FER	FGVC	KITTI	MNIST	PC	VOC	ImageNet
Clip-Chinese	86.85	44.21	18.40	34.86	14.21	3.87	32.63	14.37	52.49	67.73	22.22
mclip	92.88	65.54	29.57	46.76	41.18	7.20	23.21	52.80	51.64	77.56	42.99
Taiyi-CLIP	95.62	73.30	40.69	61.62	36.22	13.98	41.21	73.91	50.02	75.28	49.82
CN-CLIP	94.75	75.04	44.73	52.34	48.57	20.55	20.11	61.99	62.59	79.12	53.40
Altclip-xlmr-l	95.49	77.29	42.07	56.96	51.52	26.85	24.89	65.68	50.02	77.99	59.21
ZH-CLIP	97.08	80.73	47.66	51.58	48.48	20.73	20.11	61.94	62.31	78.07	56.87

Getting Started

Dependency

• python >= 3.9
• pip install -r requirements.txt

Inference

You can clone code from https://github.com/thu-ml/zh-clip

from PIL import Image
import requests
from models.zhclip import ZhCLIPProcessor, ZhCLIPModel  # Code in https://github.com/thu-ml/zh-clip

version = 'thu-ml/zh-clip-vit-roberta-large-patch14'
model = ZhCLIPModel.from_pretrained(version)
processor = ZhCLIPProcessor.from_pretrained(version)

url = "http://images.cocodataset.org/val2017/000000039769.jpg"
image = Image.open(requests.get(url, stream=True).raw)
inputs = processor(text=["一只猫", "一只狗"], images=image, return_tensors="pt", padding=True)

outputs = model(**inputs)
image_features = outputs.image_features
text_features = outputs.text_features
text_probs = (image_features @ text_features.T).softmax(dim=-1)

Other Chinese CLIP Models

In addition, to compare the effectiveness of different methods, the inference methods of other Chinese CLIP models have been integrated. For the convenience of use, the inference code has also been made public, and please contact us if there is any infringement. The code only implements models at the same level as clip-vit-large-patch14, but it may be adapted for the use of more different versions of models in the future.

#	model	alias
0	ZH-CLIP	zhclip
1	AltCLIP	altclip
2	Chinese-CLIP	cnclip
3	TaiyiCLIP	taiyiclip
4	Multilingual-CLIP	mclip
5	CLIP-Chinese	clip-chinese

Usage in inference.py