计算出图像和文本的向量，归一化后，计算余弦相似度。然后计算概率。
import torch
import torch.nn.functional as F
from PIL import Image
from transformers import AutoProcessor, AutoModel
model = AutoModel.from_pretrained("./my_local_siglip").eval()
processor = AutoProcessor.from_pretrained("./my_local_siglip")
# ---------- 图片特征 ----------
image = Image.open("你的图片.jpg").convert("RGB")
image_inputs = processor(images=image, return_tensors="pt")
with torch.no_grad():
image_features = model.get_image_features(**image_inputs)
if not isinstance(image_features, torch.Tensor):
image_features = image_features.pooler_output
# ---------- 文本特征 ----------
texts = ["一只乌龟", "一只猫", "一张风景照"]
text_inputs = processor(text=texts, padding="max_length", return_tensors="pt")
with torch.no_grad():
text_features = model.get_text_features(**text_inputs)
if not isinstance(text_features, torch.Tensor):
text_features = text_features.pooler_output
# ---------- 打印特征 ----------
print("image_features shape:", image_features.shape)
print("image_features[:5]:", image_features[0, :5])
print("text_features shape:", text_features.shape)
print("text_features[:3, :5]:", text_features[:3, :5])
# ---------- 归一化 + 余弦相似度 ----------
image_features_n = F.normalize(image_features, dim=-1)
text_features_n = F.normalize(text_features, dim=-1)
similarity = image_features_n @ text_features_n.T
print("similarity shape:", similarity.shape)
print("similarity (cosine):", similarity)
# ---------- 套用 SigLIP 打分公式 ----------
# 从模型里取出 logit_scale 和 logit_bias
logit_scale = model.logit_scale.exp() # 注意：模型里存的是 log(scale)，要 exp 还原
logit_bias = model.logit_bias
print("logit_scale (exp后):", logit_scale.item())
print("logit_bias:", logit_bias.item())
# SigLIP 公式: logit = scale * cosine + bias
logits = logit_scale * similarity + logit_bias
# sigmoid 得到概率
probs = torch.sigmoid(logits)
print("logits:", logits)
print("probs (与pipeline对齐):", probs)
# ---------- 按概率排序打印 ----------
for i, text in enumerate(texts):
print(f"标签：{text}，概率：{probs[0][i]:.2%}")输出结果
D:\>python siglip2_sim.py [transformers] Model config: bos_token_id must be `None` or an integer within the vocabulary (between 0 and 31999), got 49406. This may result in unexpected behavior. [transformers] Model config: eos_token_id must be `None` or an integer within the vocabulary (between 0 and 31999), got 49407. This may result in unexpected behavior. Loading weights: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 792/792 [00:01<00:00, 610.77it/s] image_features shape: torch.Size([1, 1024]) image_features: tensor([[ 0.4702, 0.3278, 0.0822, ..., 0.0059, -0.2070, 0.1430]]) text_features shape: torch.Size([3, 1024]) text_features: tensor([[-0.0905, 0.1456, 0.1879, ..., -0.1348, -0.4561, 0.3933], [ 0.5858, -0.0117, -0.2989, ..., 0.4996, -0.2518, 1.1217], [ 0.1977, 0.1644, -0.1696, ..., 0.0288, -0.0876, 0.6664]]) similarity shape: torch.Size([1, 3]) similarity: tensor([[0.1704, 0.0758, 0.0875]]) similarity shape: torch.Size([1, 3]) similarity (cosine): tensor([[0.1704, 0.0758, 0.0875]]) logit_scale (exp后): 107.9853286743164 logit_bias: -16.347614288330078 logits: tensor([[ 2.0506, -8.1654, -6.9033]], grad_fn=<AddBackward0>) probs (与pipeline对齐): tensor([[8.8601e-01, 2.8424e-04, 1.0035e-03]], grad_fn=<SigmoidBackward0>) 标签：一只乌龟，概率：88.60% 标签：一只猫，概率：0.03% 标签：一张风景照，概率：0.10%
可以看出乌龟的相似度只有0.1704，和猫的相似度0.0758差距不大。
但是算出来的概率确差距很大，一个88%接近1，一个0.03%接近0。
prob = sigmoid(logit_scale.exp() * cosine + logit_bias)
model.get_image_features() 是 SiglipModel（或 Siglip2Model）类的一个方法，它在内部调用了模型的视觉塔（vision_model），并额外接上一个投影层，最终输出图像的特征向量。