import gradio as gr
from transformers import AutoImageProcessor, SiglipForImageClassification
from PIL import Image
import torch

# 加载 Trash-Net 模型
model_name = "prithivMLmods/Trash-Net"
model = SiglipForImageClassification.from_pretrained(model_name)
processor = AutoImageProcessor.from_pretrained(model_name)

# 定义垃圾分类函数
def trash_classification(image):
    """输入图片，返回垃圾分类结果"""
    if image is None:
        return {}
    
    # 转换图片为 RGB
    image = Image.fromarray(image).convert("RGB")
    
    # 转换成模型需要的 tensor
    inputs = processor(images=image, return_tensors="pt")
    
    # 模型预测
    with torch.no_grad():
        outputs = model(**inputs)
        logits = outputs.logits
        probs = torch.nn.functional.softmax(logits, dim=1).squeeze().tolist()
    
    # 分类标签
    labels = ["cardboard", "glass", "metal", "paper", "plastic", "trash"]
    
    # 返回每个类别的概率
    predictions = {labels[i]: round(probs[i], 3) for i in range(len(probs))}
    return predictions

# 创建 Gradio 接口
iface = gr.Interface(
    fn=trash_classification,
    inputs=gr.Image(type="numpy"),
    outputs=gr.Label(label="Prediction Scores"),
    title="Trash Classification",
    description="Upload an image to classify the type of waste material."
)

# 启动
if __name__ == "__main__":
    iface.launch()