app.py

import torch
import gradio as gr
from PIL import Image
import json
import clip
from transformers import AutoTokenizer, AutoModelForCausalLM
from huggingface_hub import hf_hub_download
import numpy as np
# =========================================================
# 1. 讀取台南垃圾分類資料庫 JSON
# =========================================================
with open("tainan_recycle_data.json", "r", encoding="utf-8") as f:
    tainan_db = json.load(f)

def lookup_item(name):
    """模糊比對英文/中文項目"""
    name = name.lower()
    for item in tainan_db:
        # 中文名稱比對
        if name in item["name"].lower():
            return item
        # 英文名稱比對
        if name in item["english_name"].lower():
            return item
    return None


# =========================================================
# 2. 載入模型：GroundingDINO (從你的 Dataset Repo)
# =========================================================
# ⚠️ 請把這裡的 repo_id 換成你的 HuggingFace Dataset
REPO_ID = "idkWhatToUse/groundingdino-weights" 

config_path = hf_hub_download(
    repo_id=REPO_ID,
    filename="GroundingDINO_SwinT_OGC.py",
    repo_type="dataset"
)
checkpoint_path = hf_hub_download(
    repo_id=REPO_ID,
    filename="groundingdino_swint_ogc.pth",
    repo_type="dataset"
)

from groundingdino.util.inference import load_model, load_image, predict, annotate
dino_model = load_model(config_path, checkpoint_path)


# =========================================================
# 3. 載入 CLIP（分類判斷）+ LLM（分類理由）
# =========================================================
device = "cuda" if torch.cuda.is_available() else "cpu"
clip_model, preprocess = clip.load("ViT-B/32", device=device)

labels = ["一般垃圾", "紙類", "塑膠類", "金屬類", "玻璃類", "食物垃圾", "電池", "電子產品"]

# LLM (phi-2)
tokenizer = AutoTokenizer.from_pretrained("microsoft/phi-2")
llm = AutoModelForCausalLM.from_pretrained("microsoft/phi-2").to(device)


def generate_reason(item_name, category):
    prompt = f"物品「{item_name}」被歸類為「{category}」。請用一句話簡單解釋原因："
    inputs = tokenizer(prompt, return_tensors="pt").to(device)
    outputs = llm.generate(**inputs, max_new_tokens=40)
    return tokenizer.decode(outputs[0], skip_special_tokens=True)


# =========================================================
# 4. CLIP 分類（查不到官方資料時使用）
# =========================================================
def classify_clip(image, obj_name):
    text_inputs = clip.tokenize(labels).to(device)
    img_tensor = preprocess(image).unsqueeze(0).to(device)

    with torch.no_grad():
        logits, _ = clip_model(img_tensor, text_inputs)
        probs = logits.softmax(dim=-1).cpu().numpy()[0]

    idx = probs.argmax()
    category = labels[idx]
    reason = generate_reason(obj_name, category)

    return category, reason


# =========================================================
# 5. 主流程：物件偵測 → 查資料庫 → 顯示官方資訊
# =========================================================
def pipeline(image):
    # Gradio 會回傳 numpy array，因此先轉換
    if isinstance(image, np.ndarray):
        image = Image.fromarray(image.astype("uint8"))
    image.save("temp.jpg")

    # 1. GroundingDINO 偵測物品
    img_np, img_tensor = load_image("temp.jpg")
    boxes, logits, phrases = predict(
        model=dino_model,
        image=img_tensor,
        caption="bottle, can, box, plastic, cup, phone, battery, appliance, metal",
        box_threshold=0.3,
        text_threshold=0.25
    )
    annotated = annotate(img_np, boxes, logits, phrases)

    # 2. 組合輸出
    result_text = ""

    for obj in phrases:
        obj_clean = obj.lower()

        # （A）先查台南官方資料庫
        match = lookup_item(obj_clean)
        if match:
            result_text += f"""
🧩 **物品：{match['name']}**
📘 英文名稱：{match['english_name']}
♻️ 回收指示：{match['recyclable']}
📖 官方說明：{match['notes']}
🌐 資料來源：台南市政府環保局

"""
            continue

        # （B）查不到 → 用 CLIP 推論 + LLM 理由
        category, reason = classify_clip(image, obj_clean)

        result_text += f"""
🧩 偵測到：{obj}
📦 分類推論：{category}
💡 理由：{reason}

"""

    return annotated, result_text


# =========================================================
# 6. Gradio 介面
# =========================================================
with gr.Blocks() as demo:
    gr.Markdown("# 🗑️ AI 垃圾分類助手（含台南官方資料庫）")
    gr.Markdown("上傳照片 → 物件偵測 → 查台南市環保局資料庫 → 顯示回收建議")

    img_input = gr.Image(label="上傳垃圾照片")
    btn = gr.Button("開始分析 🚀", variant="primary")

    img_output = gr.Image(label="物件偵測（GroundingDINO）")
    text_output = gr.Textbox(label="分類結果", lines=10)

    btn.click(pipeline, inputs=[img_input], outputs=[img_output, text_output])

demo.launch()