🚀 WD SwinV2 Tagger v3 with 🤗 transformers

本项目将 SmilingWolf/wd-swinv2-tagger-v3 模型转换为了transformers库的格式，可用于图像分类任务。

🚀 快速开始

点击下面的链接，在Colab中体验该模型：

📦 安装指南

使用以下命令安装transformers库：

pip install transformers

💻 使用示例

基础用法

使用pipeline进行图像分类：

from transformers import pipeline

pipe = pipeline(
    "image-classification",
    model="p1atdev/wd-swinv2-tagger-v3-hf",
    trust_remote_code=True,
)

print(pipe("sample.webp", top_k=15))
#[{'label': '1girl', 'score': 0.9973934888839722},
# {'label': 'solo', 'score': 0.9719744324684143},
# {'label': 'dress', 'score': 0.9539461135864258},
# {'label': 'hat', 'score': 0.9511678218841553},
# {'label': 'outdoors', 'score': 0.9438753128051758},
# ...

高级用法

使用AutoModel进行图像分类：

from PIL import Image

import numpy as np
import torch

from transformers import (
    AutoImageProcessor,
    AutoModelForImageClassification,
)

MODEL_NAME = "p1atdev/wd-swinv2-tagger-v3-hf"

model = AutoModelForImageClassification.from_pretrained(
    MODEL_NAME,
)
processor = AutoImageProcessor.from_pretrained(MODEL_NAME, trust_remote_code=True)

image = Image.open("sample.webp")
inputs = processor.preprocess(image, return_tensors="pt")

with torch.no_grad():
  outputs = model(**inputs.to(model.device, model.dtype))
logits = torch.sigmoid(outputs.logits[0]) # take the first logits

# get probabilities
results = {model.config.id2label[i]: logit.float() for i, logit in enumerate(logits)}
results = {
    k: v for k, v in sorted(results.items(), key=lambda item: item[1], reverse=True) if v > 0.35 # 35% threshold
}
print(results)  # rating tags and character tags are also included
#{'1girl': tensor(0.9974),
# 'solo': tensor(0.9720),
# 'dress': tensor(0.9539),
# 'hat': tensor(0.9512),
# 'outdoors': tensor(0.9439),
# ...

使用🤗 Optimum加速

使用🤗 Optimum可以使模型推理速度提升约30%，模型大小减少约50%，但准确率会略有下降。

pip install optimum[onnxruntime] 

-from transformers import pipeline
+from optimum.pipelines import pipeline

pipe = pipeline(
    "image-classification",
    model="p1atdev/wd-swinv2-tagger-v3-hf",
    trust_remote_code=True,
)

print(pipe("sample.webp", top_k=15))
#[{'label': '1girl', 'score': 0.9966088533401489},
# {'label': 'solo', 'score': 0.9740601778030396},
# {'label': 'dress', 'score': 0.9618403911590576},
# {'label': 'hat', 'score': 0.9563733339309692},
# {'label': 'outdoors', 'score': 0.945336639881134},
# ...

📚 详细文档

所有评级标签都有前缀 rating:，角色标签都有前缀 character:。

• 评级标签：rating:general, rating:sensitive, ...
• 角色标签：character:frieren, character:hatsune miku, ...

📄 许可证

本项目使用 apache-2.0 许可证。