app.py

# app.py - Gradio demo for testing your image-classification model (Food-101)
# - Replace FOOD101_MODEL_ID with your model repo id (e.g., "yourname/your-food-model")
# - If your model is private, add HF_TOKEN as a secret in your Space (or set env var)

import os
from PIL import Image
import torch
from transformers import AutoImageProcessor, AutoModelForImageClassification
import gradio as gr

# ---- CONFIG: 改成你的 model id ----
Tell_Me_Recipe = os.environ.get("Tell_Me_Recipe", "YOUR_USERNAME/YOUR_FOOD101_MODEL")
# optional gate model (food / not-food). Change or set to None to skip.
GATE_MODEL_ID = os.environ.get("GATE_MODEL_ID", "prithivMLmods/Food-or-Not-SigLIP2")
HF_TOKEN = os.environ.get("HF_TOKEN")  # 用於 private 模型

# ---- device ----
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

# ---- load models (一次載入) ----
def load_model(model_id):
    if model_id is None or model_id.strip() == "":
        return None, None
    try:
        processor = AutoImageProcessor.from_pretrained(model_id, use_auth_token=HF_TOKEN)
        model = AutoModelForImageClassification.from_pretrained(model_id, use_auth_token=HF_TOKEN)
        model.to(device)
        model.eval()
        return processor, model
    except Exception as e:
        print(f"Failed to load {model_id}: {e}")
        raise

print("Loading main classification model:", Tell_Me_Recipe)
clf_processor, clf_model = load_model(Tell_Me_Recipe)

if GATE_MODEL_ID:
    try:
        print("Loading gate model (food vs not-food):", GATE_MODEL_ID)
        gate_processor, gate_model = load_model(GATE_MODEL_ID)
    except Exception as e:
        print("Gate model load failed — continuing without gate:", e)
        gate_processor, gate_model = None, None
else:
    gate_processor, gate_model = None, None

# ---- small recipe template mapping (示範用) ----
RECIPES = {
    "pizza": {
        "ingredients": ["高筋麵粉 300g", "水 180ml", "酵母 3g", "番茄醬", "Mozzarella 起司"],
        "steps": ["揉麵發酵", "桿皮抹醬加起司配料", "220°C 烤 10-15 分鐘"]
    },
    "ramen": {
        "ingredients": ["中華麵", "高湯 400ml", "醬油", "叉燒", "溏心蛋"],
        "steps": ["煮麵 / 熬高湯 / 擺盤"]
    },
    "cheesecake": {
        "ingredients": ["奶油乳酪 200g", "蛋 2 顆", "砂糖 60g", "消化餅底"],
        "steps": ["餅乾壓底 / 乳酪餡打勻倒入 / 160°C 烤 40-50 分鐘"]
    }
}

def simple_recipe_for(label: str):
    key = label.replace("_", " ").lower()
    for k in RECIPES:
        if k in key:
            r = RECIPES[k]
            ing = "\n".join(f"- {i}" for i in r["ingredients"])
            steps = "\n".join(f"{idx+1}. {s}" for idx, s in enumerate(r["steps"]))
            return f"【材料】\n{ing}\n\n【步驟】\n{steps}"
    return f"找不到精確食譜。模型預測：{label.replace('_',' ')}。\n建議搜尋該菜名的食譜或連到 RecipeNLG 做檢索。"

# ---- inference helpers ----
@torch.inference_mode()
def is_food_image(image: Image.Image, threshold: float = 0.5):
    """如果有 gate_model，回傳 (is_food:bool, score:float, label:str)"""
    if gate_model is None or gate_processor is None:
        return True, 1.0, "no-gate"
    inputs = gate_processor(images=image, return_tensors="pt")
    inputs = {k: v.to(device) for k, v in inputs.items()}
    out = gate_model(**inputs)
    probs = out.logits.softmax(-1).squeeze(0)
    topv, topi = torch.max(probs, dim=-1)
    label = gate_model.config.id2label[int(topi)]
    # 假設 gate 的 label 包含 'not' 或 'not_food' 來表現非食物
    not_food_names = ["not-food", "not_food", "not food", "notfood", "no-food"]
    is_food = True
    if any(n in label.lower() for n in not_food_names):
        is_food = False
    return is_food, float(topv), label

@torch.inference_mode()
def predict_label(image: Image.Image, topk: int = 3):
    inputs = clf_processor(images=image, return_tensors="pt")
    inputs = {k: v.to(device) for k, v in inputs.items()}
    out = clf_model(**inputs)
    probs = out.logits.softmax(-1).squeeze(0)
    topv, topi = torch.topk(probs, k=min(topk, probs.shape[0]))
    labels = [clf_model.config.id2label[int(i)] for i in topi]
    return [(l.replace("_"," "), float(v)) for l, v in zip(labels, topv)]

# ---- gradio UI function ----
def analyze_image(image: Image.Image, topk: int=3, gate_threshold: float=0.5, use_gate: bool=True):
    if image is None:
        return "請上傳圖片", "", "", ""
    try:
        # 1) gate (optional)
        if use_gate and gate_model is not None:
            is_food, score, gate_label = is_food_image(image, gate_threshold)
            if not is_food and score >= gate_threshold:
                return f"判斷：非食物（{gate_label}，score={score:.2f}）", "", "", "這張圖被判定為「非食物」，不做菜名預測。"
        # 2) predict top-k
        preds = predict_label(image, topk=topk)
        topk_text = "\n".join([f"{lbl}：{p:.3f}" for lbl,p in preds])
        best_label = preds[0][0]
        recipe_txt = simple_recipe_for(best_label)
        return f"判斷：是食物（gate ok）", best_label, topk_text, recipe_txt
    except Exception as e:
        return f"發生錯誤：{e}", "", "", ""

# ---- build UI ----
with gr.Blocks() as demo:
    gr.Markdown("# Food → 菜名 + 簡易食譜 Demo")
    with gr.Row():
        img = gr.Image(type="pil", label="上傳一張照片（任意圖片）")
        with gr.Column():
            topk = gr.Slider(1, 10, value=3, step=1, label="Top-K")
            gate_check = gr.Checkbox(value=True, label="使用 food vs not-food gate（建議開）")
            gate_th = gr.Slider(0.0, 1.0, value=0.5, label="Gate threshold")
            run_btn = gr.Button("分析照片")
    with gr.Row():
        out0 = gr.Textbox(label="是否為食物 / Gate 訊息", lines=1)
    with gr.Row():
        out1 = gr.Textbox(label="預測菜名（Top 1）", lines=1)
    with gr.Row():
        out2 = gr.Textbox(label="Top-K 預測 & 機率", lines=6)
    with gr.Row():
        out3 = gr.Textbox(label="自動回傳的簡易食譜（示範）", lines=12)
    run_btn.click(fn=analyze_image, inputs=[img, topk, gate_th, gate_check], outputs=[out0, out1, out2, out3])

if __name__ == "__main__":
    demo.launch(server_name="0.0.0.0", server_port=int(os.environ.get("PORT", 7860)))