app.py

import gradio as gr
from transformers import AutoImageProcessor, AutoModelForImageClassification, pipeline
import torch
import torch.nn.functional as F
from PIL import Image

# ==========================================
# 1. INISIALISASI MODEL
# ==========================================
print("Sedang memuat model... Mohon tunggu.")

# Gatekeeper: Google SigLIP
gatekeeper = pipeline("zero-shot-image-classification", model="google/siglip-base-patch16-224")

# Specialist: ConvNeXt
model_name = "louislu9911/convnextv2-base-22k-384-finetuned-cassava-leaf-disease"
processor = AutoImageProcessor.from_pretrained(model_name)
model = AutoModelForImageClassification.from_pretrained(model_name)

# Label Mapping
LABELS = {
    0: "Hawar Bakteri (CBB)",
    1: "Goresan Coklat (CBSD)",
    2: "Bercak Hijau (CGM)",
    3: "Virus Mosaik (CMD)",
    4: "Sehat (Healthy)"
}

# ==========================================
# 2. FUNGSI UTAMA
# ==========================================
def deteksi(image):
    if image is None: return "Error", "Silakan upload gambar."
    
    try:
        # ------------------------------------------------------------
        # VALIDASI INTERNAL (INPUT TETAP SPESIFIK)
        # ------------------------------------------------------------
        # Kita biarkan AI menebak secara detail di dalam hati,
        # agar dia bisa membedakan fitur visual dengan akurat.
        
        all_candidates = [
            "cassava leaf",  # TARGET
            
            # Group Tanaman Lain (Musuh)
            "papaya leaf", "castor bean leaf", "kapok leaf", "marijuana leaf",
            "corn leaf", "rice plant leaf", "banana leaf", 
            "mango leaf", "tomato leaf", "chili pepper leaf",
            "ginger turmeric leaf", # Tambahan biar jahe ketangkap
            
            # Group Digital/Palsu
            "vector art digital illustration", 
            "3D rendered image", 
            "pixel art", 
            "sketch drawing cartoon", 
            "artificial plastic plant",
            
            # Group Bukan Daun
            "cassava tuber root", "tree stem wood", "soil ground", "human person",
            
            # Group Kualitas Buruk
            "solid black or dark image", 
            "blurry out of focus image", 
            "white blank image", 
            "far away field landscape"
        ]
        
        # Jalankan Gatekeeper
        res = gatekeeper(image, candidate_labels=all_candidates)
        top_label = res[0]['label']
        top_score = res[0]['score']

        # --- LOGIKA OUTPUT (KE USER) ---
        # Di sini kita "sensor" jawaban AI biar tidak spesifik/sok tau.
        
        if top_label != "cassava leaf":
            
            pesan_error = "Objek Ditolak." # Default
            
            # 1. CEK DIGITAL / PALSU
            # Kalau labelnya mengandung kata-kata ini, kita bilang "Bukan Foto Asli"
            fake_keywords = ["vector", "illustration", "3D", "pixel", "sketch", "cartoon", "artificial", "plastic"]
            if any(k in top_label for k in fake_keywords):
                pesan_error = "Bukan foto daun asli (Terdeteksi gambar digital/palsu)."

            # 2. CEK KUALITAS FOTO
            elif any(k in top_label for k in ["black", "dark", "blurry", "white", "landscape"]):
                pesan_error = "Kualitas foto buruk (Gelap/Blur/Kejauhan)."

            # 3. CEK BUKAN DAUN (BATANG/TANAH/MANUSIA/UMBI)
            elif any(k in top_label for k in ["root", "stem", "soil", "human"]):
                pesan_error = "Objek bukan daun (Terdeteksi Batang/Tanah/Manusia)."

            # 4. SISA NYA ADALAH "TANAMAN LAIN" (JAGUNG, PEPAYA, JAHE, DLL)
            # Kita tidak perlu sebut namanya. Cukup bilang "Bukan Daun Singkong".
            else:
                pesan_error = "Bukan Daun Singkong (Terdeteksi jenis tanaman lain)."
            
            return "DITOLAK", f"{pesan_error} (Skor AI: {top_score:.1%})"
        
        # Syarat Skor Minimal
        if top_score < 0.20:
             return "MERAGUKAN", f"Mirip singkong tapi AI ragu ({top_score:.1%}). Pastikan cahaya cukup & fokus."

        # ------------------------------------------------------------
        # DIAGNOSA PENYAKIT
        # ------------------------------------------------------------
        inputs = processor(images=image, return_tensors="pt")
        with torch.no_grad():
            logits = model(**inputs).logits
        probs = F.softmax(logits, dim=-1)
        idx = probs.argmax(-1).item()
        confidence = probs[0][idx].item()
        
        if confidence < 0.50:
            return "Tidak Teridentifikasi", f"Pola penyakit tidak jelas. ({confidence:.1%})"

        return LABELS[idx], f"Keyakinan: {confidence:.1%}"

    except Exception as e:
        return "System Error", str(e)

# ==========================================
# 3. INTERFACE
# ==========================================
title = "🍂 Deteksi Singkong (Simple Validator)"
description = "Sistem hanya menerima foto Daun Singkong Asli. Foto buram, digital, atau tanaman lain akan ditolak."

iface = gr.Interface(
    fn=deteksi,
    inputs=gr.Image(type="pil", label="Upload Foto"),
    outputs=[gr.Label(num_top_classes=0, label="Status"), gr.Textbox(label="Keterangan")],
    title=title,
    description=description,
    flagging_mode="never"
)

iface.launch()