File size: 12,997 Bytes
3f628f6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f1615f0
3f628f6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f1615f0
 
 
3f628f6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
import os
import sys
import types
import importlib.machinery
from typing import List, Dict

import gradio as gr
import torch
from PIL import Image

# ========== 1) 偽裝 flash_attn,避免 remote code 硬性檢查 ==========
def _make_pkg_stub(fullname: str):
    m = types.ModuleType(fullname)
    m.__file__ = f"<stub {fullname}>"
    m.__package__ = fullname.rpartition('.')[0]
    m.__path__ = []  # 標記為 package
    m.__spec__ = importlib.machinery.ModuleSpec(fullname, loader=None, is_package=True)
    sys.modules[fullname] = m
    return m

for name in [
    "flash_attn",
    "flash_attn.ops",
    "flash_attn.layers",
    "flash_attn.functional",
    "flash_attn.bert_padding",
    "flash_attn.flash_attn_interface",
]:
    if name not in sys.modules:
        _make_pkg_stub(name)

# ========== 2) Florence-2 載入(eager + dtype 對齊) ==========
from transformers import AutoProcessor, AutoModelForCausalLM

MODEL_ID = os.getenv("MODEL_ID", "microsoft/Florence-2-base")
device = "cuda" if torch.cuda.is_available() else "cpu"

TASK_TOKENS = {
    "caption": "<CAPTION>",
    "object_detection": "<OBJECT_DETECTION>",
}

_processor = None
_model = None

def get_florence2():
    global _processor, _model
    if _processor is None or _model is None:
        _processor = AutoProcessor.from_pretrained(MODEL_ID, trust_remote_code=True)
        _model = AutoModelForCausalLM.from_pretrained(
            MODEL_ID,
            trust_remote_code=True,
            attn_implementation="eager",  # 關鍵:不依賴 flash_attn
            torch_dtype=torch.float16 if device == "cuda" else torch.float32
        ).to(device).eval()
        _model.config.use_cache = False
    return _processor, _model

@torch.inference_mode()
def florence2_text(image: Image.Image, task: str = "caption"):
    proc, mdl = get_florence2()
    token = TASK_TOKENS.get(task, "<CAPTION>")
    text = token  # 這兩個任務都是「不帶輸入」的格式

    # 先在 CPU 做處理,再手動搬到正確 device / 對齊 dtype
    batch = proc(text=text, images=image, return_tensors="pt")
    inputs = {}
    for k, v in batch.items():
        if isinstance(v, torch.Tensor):
            if v.is_floating_point():
                inputs[k] = v.to(device=device, dtype=mdl.dtype)
            else:
                inputs[k] = v.to(device=device)
        else:
            inputs[k] = v

    ids = mdl.generate(
        **inputs,
        max_new_tokens=128,
        do_sample=False,
        num_beams=1,
        use_cache=False,      # ← 關掉 KV-cache(關鍵)
        early_stopping=False, # ← 與 num_beams=1 時無效,但設 False 更乾淨
        eos_token_id=getattr(getattr(proc, "tokenizer", None), "eos_token_id", None),
    )
    out = proc.batch_decode(ids, skip_special_tokens=True)[0].strip()
    if ">" in out:
        out = out.split(">", 1)[-1].strip()
    return out

# ========== 3) 後端邏輯(食物 DB / 同義詞 / 估重 / 規則) ==========
FOOD_DB = {
    "rice":  {"kcal":130, "carb_g":28, "protein_g":2.4, "fat_g":0.3, "sodium_mg":0,   "cat":"全榖雜糧類", "base_g":150, "tip":"主食可改糙米/全穀增加膳食纖維"},
    "noodles":{"kcal":138, "carb_g":25, "protein_g":4.5, "fat_g":1.9, "sodium_mg":170, "cat":"全榖雜糧類", "base_g":180, "tip":"盡量選清湯少油,避免重鹹湯底"},
    "bread": {"kcal":265, "carb_g":49, "protein_g":9.0, "fat_g":3.2, "sodium_mg":490, "cat":"全榖雜糧類", "base_g":60,  "tip":"可選全麥減少抹醬、甜餡"},
    "broccoli":{"kcal":35, "carb_g":7, "protein_g":2.4, "fat_g":0.4, "sodium_mg":33,  "cat":"蔬菜類",     "base_g":80,  "tip":"川燙/清炒保留口感與維生素"},
    "spinach":{"kcal":23, "carb_g":3.6,"protein_g":2.9,"fat_g":0.4,"sodium_mg":70,    "cat":"蔬菜類",     "base_g":80,  "tip":"川燙後快炒,少鹽少油"},
    "chicken":{"kcal":215,"carb_g":0,  "protein_g":27, "fat_g":12, "sodium_mg":90,    "cat":"豆魚蛋肉類", "base_g":120, "tip":"去皮烹調、烤/氣炸取代油炸"},
    "soy_braised_chicken_leg":{"kcal":220,"carb_g":0,"protein_g":24,"fat_g":12,"sodium_mg":550,"cat":"豆魚蛋肉類","base_g":130,"tip":"減醬油與滷汁、可先汆燙再滷"},
    "salmon":{"kcal":208,"carb_g":0,  "protein_g":20, "fat_g":13, "sodium_mg":60,     "cat":"豆魚蛋肉類", "base_g":120, "tip":"烤/蒸保留 Omega-3,少鹽少醬"},
    "pork_chop":{"kcal":242,"carb_g":0,"protein_g":27,"fat_g":14,"sodium_mg":75,      "cat":"豆魚蛋肉類", "base_g":120, "tip":"少裹粉油炸,改煎烤並瀝油"},
    "tofu":  {"kcal":76, "carb_g":1.9,"protein_g":8.1,"fat_g":4.8,"sodium_mg":7,      "cat":"豆魚蛋肉類", "base_g":120, "tip":"少勾芡、少滷汁,清蒸清爽"},
    "egg":   {"kcal":155,"carb_g":1.1,"protein_g":13, "fat_g":11, "sodium_mg":124,    "cat":"豆魚蛋肉類", "base_g":60,  "tip":"水煮/荷包少油,避免重鹹醬料"},
    "banana":{"kcal":89, "carb_g":23, "protein_g":1.1,"fat_g":0.3,"sodium_mg":1,      "cat":"水果類",     "base_g":100, "tip":"控制份量,避免一次過量"},
    "miso_soup":{"kcal":36,"carb_g":4.3,"protein_g":2.0,"fat_g":1.3,"sodium_mg":550,  "cat":"湯品/飲品",  "base_g":200, "tip":"味噌湯偏鹹,建議少量品嚐"},
}

ALIASES = {
    "white rice":"rice","steamed rice":"rice","飯":"rice","白飯":"rice",
    "麵":"noodles","拉麵":"noodles","麵條":"noodles","義大利麵":"noodles",
    "麵包":"bread","吐司":"bread",
    "雞肉":"chicken","雞胸":"chicken","烤雞":"chicken",
    "滷雞腿":"soy_braised_chicken_leg","醬油雞腿":"soy_braised_chicken_leg",
    "鮭魚":"salmon","三文魚":"salmon",
    "豬排":"pork_chop",
    "豆腐":"tofu",
    "蛋":"egg","水煮蛋":"egg","荷包蛋":"egg",
    "花椰菜":"broccoli","青花菜":"broccoli","菠菜":"spinach",
    "香蕉":"banana","味噌湯":"miso_soup",
}

RULES = {"T2DM": {"carb_g_per_meal_max": 60}, "HTN": {"sodium_mg_per_meal_max": 600}}
PORTION_MUL = {"小":0.8, "中":1.0, "大":1.2}

def detect_foods_from_text(text: str) -> List[str]:
    lower = text.lower()
    labels = set()
    for k in FOOD_DB.keys():
        if k in lower:
            labels.add(k)
    for alias, key in ALIASES.items():
        if alias in text or alias.lower() in lower:
            labels.add(key)
    return list(labels)

# 自由抽詞(允許未知)
import re
DEFAULT_BASE_G = 100
STOPWORDS = {
    "a","an","the","with","and","of","on","in","to","served","over","side","sides",
    "plate","bento","box","set","dish","meal","mixed","assorted","fresh","hot","cold",
    "grilled","roasted","fried","deep","steamed","boiled","braised","stir","stirred","sautéed",
    "sauce","soup","salad","topped","seasoned","style","japanese","taiwanese","korean","chinese",
    "便當","套餐","一盤","一碗","配菜","附餐","湯","沙拉","醬","佐","搭配","附","拌","炒","滷","炸","烤","蒸","煮"
}
def extract_food_terms_free(text: str):
    parts = re.split(r"(?:,|\.|;|\band\b|\bwith\b|\bserved with\b|\baccompanied by\b|\n)+", text, flags=re.I)
    hits = set()
    for p in parts:
        if not p: continue
        toks = re.findall(r"[A-Za-z\u4e00-\u9fff]+", p.lower())
        toks = [w for w in toks if len(w) >= 2 and w not in STOPWORDS]
        if not toks: continue
        head = toks[-1]
        hits.add(ALIASES.get(head, head))
    return list(hits)

def estimate_weight(name: str, plate_cm: int, portion: str) -> int:
    base = FOOD_DB.get(name, {}).get("base_g", DEFAULT_BASE_G)
    mul = PORTION_MUL.get(portion, 1.0)
    grams = int(base * mul * (plate_cm / 24))
    return max(10, grams)

def grams_to_nutrition(name: str, grams: int) -> Dict:
    info = FOOD_DB[name]
    ratio = grams / 100.0
    out = {"name": name, "cat": info["cat"], "weight_g": grams, "tip": info.get("tip","")}
    for k in ("kcal","carb_g","protein_g","fat_g","sodium_mg"):
        out[k] = round(info[k] * ratio, 1)
    return out

def make_placeholder_item(name: str, plate_cm: int, portion: str):
    grams = int(DEFAULT_BASE_G * (plate_cm / 24) * PORTION_MUL.get(portion, 1.0))
    return {
        "name": name, "cat": "未分類", "weight_g": grams,
        "kcal": "待新增資訊", "carb_g": "待新增資訊", "protein_g": "待新增資訊",
        "fat_g": "待新增資訊", "sodium_mg": "待新增資訊", "tip": "待新增資訊"
    }

def eval_rules(items: List[Dict], conditions: List[str]):
    totals = {}
    for it in items:
        # 只累加可數值的項
        if isinstance(it.get("kcal"), (int, float)):
            for k in ("kcal","carb_g","protein_g","fat_g","sodium_mg"):
                totals[k] = round(totals.get(k,0) + float(it[k]), 1)
    advice = []
    if "T2DM" in conditions and totals.get("carb_g",0) > RULES["T2DM"]["carb_g_per_meal_max"]:
        advice.append("【糖尿病】碳水偏高,建議主食減量或改全穀。")
    if "HTN" in conditions and totals.get("sodium_mg",0) > RULES["HTN"]["sodium_mg_per_meal_max"]:
        advice.append("【高血壓】鈉含量偏高,少鹽、避免重口味與滷味/湯品。")
    cats = {}
    for it in items:
        cats[it["cat"]] = cats.get(it["cat"], 0) + 1
    return totals, advice, cats

# ========== 4) Gradio 介面 ==========
def run_pipeline(image, plate_cm, portion, conditions, task_mode, dev_mode):
    if image is None:
        return "請先上傳一張照片。", "", [], {}

    # 1) 文字輸出(Dev 模式跳過模型)
    if dev_mode:
        txt = "A bento with white rice, broccoli and grilled chicken thigh."
    else:
        t = "caption" if task_mode == "描述 (Caption)" else "object_detection"
        txt = florence2_text(image, task=t)

    # 2) 合併偵測
    labels_known = detect_foods_from_text(txt)
    labels_free  = extract_food_terms_free(txt)
    labels_all = []
    seen = set()
    for term in labels_free + labels_known:
        key = ALIASES.get(term, term)
        if key not in seen:
            labels_all.append(key)
            seen.add(key)

    # 3) 產生 items(未知也照列)
    items = []
    for name in labels_all[:6]:
        if name in FOOD_DB:
            g = estimate_weight(name, plate_cm, portion)
            items.append(grams_to_nutrition(name, g))
        else:
            items.append(make_placeholder_item(name, plate_cm, portion))

    totals, advice, cats = eval_rules(items, conditions)

    # 4) 組合輸出
    lines = [f"模型輸出:{txt}", ""]
    if labels_all:
        lines.append("偵測到: " + ", ".join(labels_all))
    else:
        lines.append("偵測到: (無)")

    lines.append("")
    for it in items:
        kcal = it['kcal'] if isinstance(it['kcal'], (int, float)) else it['kcal']
        carb = it['carb_g'] if isinstance(it['carb_g'], (int, float)) else it['carb_g']
        prot = it['protein_g'] if isinstance(it['protein_g'], (int, float)) else it['protein_g']
        fat  = it['fat_g'] if isinstance(it['fat_g'], (int, float)) else it['fat_g']
        na   = it['sodium_mg'] if isinstance(it['sodium_mg'], (int, float)) else it['sodium_mg']
        lines.append(f"- {it['name']} ({it['cat']}) {it['weight_g']} g → "
                     f"{kcal} kcal, C{carb} g, P{prot} g, F{fat} g, Na{na} mg")

    if totals:
        lines.append("")
        lines.append(f"總計:{totals.get('kcal',0)} kcal,碳水 {totals.get('carb_g',0)} g,蛋白 {totals.get('protein_g',0)} g,脂肪 {totals.get('fat_g',0)} g,鈉 {totals.get('sodium_mg',0)} mg")
    if advice:
        lines.append("建議:" + " ".join(advice))

    return "\n".join(lines), txt, items, totals

with gr.Blocks(title="FoodAI · Florence-2 Demo") as demo:
    gr.Markdown("# 🍱 FoodAI · Florence-2 Demo\n上傳餐點 → 產生描述/偵測 → 估營養/建議\n\n> 開發模式:不跑模型,固定假字串方便測試 UI/流程。")
    with gr.Row():
        with gr.Column(scale=1):
            img = gr.Image(type="pil", label="上傳圖片")
            plate = gr.Slider(18, 28, value=24, step=1, label="盤子直徑 (cm)")
            portion = gr.Radio(["小", "中", "大"], value="中", label="份量")
            cond = gr.CheckboxGroup(["T2DM", "HTN"], label="狀況")
            task_mode = gr.Radio(["描述 (Caption)", "偵測 (Object Detection)"], value="描述 (Caption)", label="任務")
            dev_mode = gr.Checkbox(label="開發模式(不跑模型)", value=False)
            btn = gr.Button("開始分析", variant="primary")
        with gr.Column(scale=1):
            out_md = gr.Markdown(label="結果")
            raw = gr.Textbox(label="模型原始輸出", lines=4)
            js = gr.JSON(label="逐項結果")
            total = gr.JSON(label="總計")

    btn.click(run_pipeline, inputs=[img, plate, portion, cond, task_mode, dev_mode], outputs=[out_md, raw, js, total])

if __name__ == "__main__":
    # 在本地/Spaces 都可直接執行
    PORT = int(os.getenv("PORT", "7860"))
    demo.launch(server_name="0.0.0.0", server_port=PORT)