标签:

• clip
• siglip 库名称: open_clip 管道标签: 零样本图像分类 许可证: apache-2.0 数据集:
• webli

ViT-L-16-SigLIP-256 模型卡

一个在WebLI数据集上训练的SigLIP（语言图像预训练的Sigmoid损失）模型。

此模型已从Big Vision中的原始JAX检查点转换为PyTorch。这些权重可在OpenCLIP（图像+文本）和timm（仅图像）中使用。

模型详情

• 模型类型: 对比图像-文本，零样本图像分类。
• 原始来源: https://github.com/google-research/big_vision
• 数据集: WebLI
• 论文:
  • 语言图像预训练的Sigmoid损失: https://arxiv.org/abs/2303.15343

模型使用

使用OpenCLIP

import torch
import torch.nn.functional as F
from urllib.request import urlopen
from PIL import Image
from open_clip import create_model_from_pretrained, get_tokenizer # 适用于open-clip-torch>=2.23.0, timm>=0.9.8

model, preprocess = create_model_from_pretrained('hf-hub:timm/ViT-L-16-SigLIP-256')
tokenizer = get_tokenizer('hf-hub:timm/ViT-L-16-SigLIP-256')

image = Image.open(urlopen(
    'https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/beignets-task-guide.png'
))
image = preprocess(image).unsqueeze(0)

labels_list = ["一只狗", "一只猫", "一个甜甜圈", "一个贝涅饼"]
text = tokenizer(labels_list, context_length=model.context_length)

with torch.no_grad(), torch.cuda.amp.autocast():
    image_features = model.encode_image(image)
    text_features = model.encode_text(text)
    image_features = F.normalize(image_features, dim=-1)
    text_features = F.normalize(text_features, dim=-1)

    text_probs = torch.sigmoid(image_features @ text_features.T * model.logit_scale.exp() + model.logit_bias)

zipped_list = list(zip(labels_list, [round(p.item(), 3) for p in text_probs[0]]))
print("标签概率: ", zipped_list)

使用timm（用于图像嵌入）

from urllib.request import urlopen
from PIL import Image
import timm

image = Image.open(urlopen(
    'https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/beignets-task-guide.png'
))

model = timm.create_model(
    'vit_large_patch16_siglip_256',
    pretrained=True,
    num_classes=0,
)
model = model.eval()

# 获取模型特定的转换（归一化，调整大小）
data_config = timm.data.resolve_model_data_config(model)
transforms = timm.data.create_transform(**data_config, is_training=False)

output = model(transforms(image).unsqueeze(0))  # 输出是（batch_size, num_features）形状的张量

引用

@article{zhai2023sigmoid,
  title={Sigmoid loss for language image pre-training},
  author={Zhai, Xiaohua and Mustafa, Basil and Kolesnikov, Alexander and Beyer, Lucas},
  journal={arXiv preprint arXiv:2303.15343},
  year={2023}
}

@misc{big_vision,
  author = {Beyer, Lucas and Zhai, Xiaohua and Kolesnikov, Alexander},
  title = {Big Vision},
  year = {2022},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/google-research/big_vision}}
}