Upload 5 files

Dockerfile

@@ -0,0 +1,36 @@
+# Gunakan base image Python yang ringan
+FROM python:3.9-slim as builder
+
+# Instal library sistem yang dibutuhkan untuk proses build
+RUN apt-get update && apt-get install -y \
+    build-essential \
+    && rm -rf /var/lib/apt/lists/*
+
+WORKDIR /code
+
+# Salin dan instal requirements dalam tahap builder
+COPY ./requirements.txt /code/requirements.txt
+RUN pip install --no-cache-dir --upgrade -r /code/requirements.txt
+
+# --- Tahap Final ---
+# Mulai lagi dari base image yang bersih untuk runtime
+FROM python:3.9-slim
+
+# Instal HANYA library sistem yang dibutuhkan untuk menjalankan aplikasi
+RUN apt-get update && apt-get install -y \
+    libgl1-mesa-glx \
+    libglib2.0-0 \
+    && rm -rf /var/lib/apt/lists/*
+
+WORKDIR /code
+
+# Salin paket python yang sudah terinstal dari tahap builder
+COPY --from=builder /usr/local/lib/python3.9/site-packages /usr/local/lib/python3.9/site-packages
+
+# Salin kode aplikasi Anda
+COPY . /code/
+
+EXPOSE 7860
+
+# CMD tetap sama, memastikan aplikasi dijalankan dengan Gunicorn
+CMD ["gunicorn", "--workers", "3", "--bind", "0.0.0.0:7860", "--timeout", "120", "app:app"]

app.py

@@ -0,0 +1,272 @@
+# server.py
+# Server Flask untuk prediksi multi-head captcha yang disesuaikan untuk Hugging Face Spaces.
+# Versi ini sudah dioptimalkan untuk memuat model yang terkuantisasi (INT8).
+
+from flask import Flask, request, jsonify
+from flask_cors import CORS
+import torch
+import torch.nn as nn
+import timm
+import albumentations as A
+from albumentations.pytorch import ToTensorV2
+from PIL import Image, UnidentifiedImageError
+import numpy as np
+import cv2
+import os
+import json
+import base64
+from io import BytesIO
+import sys
+import logging
+
+torch.set_num_threads(1)
+
+# ==============================================================================
+#  BAGIAN 1: PENGATURAN DASAR FLASK & LOGGING
+# ==============================================================================
+
+logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(levelname)s - %(message)s', stream=sys.stdout)
+app = Flask(__name__)
+CORS(app)
+
+MODEL = None
+MAPPINGS = None
+DEVICE = None
+TRANSFORMS = None
+
+# ==============================================================================
+#  BAGIAN 2: DEFINISI MODEL (TIDAK ADA PERUBAHAN)
+# ==============================================================================
+
+class TextHeadCTC(nn.Module):
+    def __init__(self, input_dim, hidden_dim, ctc_vocab_size):
+        super().__init__()
+        self.rnn = nn.LSTM(input_dim, hidden_dim, num_layers=1, bidirectional=True, batch_first=True)
+        self.fc = nn.Linear(hidden_dim * 2, ctc_vocab_size)
+    def forward(self, x):
+        rnn_out, _ = self.rnn(x)
+        output_logits = self.fc(rnn_out)
+        return nn.functional.log_softmax(output_logits, dim=2).permute(1, 0, 2)
+
+class MultiHeadModel(nn.Module):
+    def __init__(self, backbone_name, ctc_vocab_size, num_object_classes, num_types):
+        super().__init__()
+        self.backbone = timm.create_model(
+            backbone_name,
+            pretrained=False,
+            num_classes=0,
+            drop_path_rate=0.1
+        )
+        backbone_features_dim = self.backbone.num_features
+        rnn_hidden_dim, projected_embed_dim = 256, 256
+        self.type_head = nn.Linear(backbone_features_dim, num_types)
+        self.object_head = nn.Linear(backbone_features_dim, num_object_classes)
+        self.input_proj = nn.Conv2d(backbone_features_dim, projected_embed_dim, kernel_size=1)
+        self.text_head_ctc = TextHeadCTC(projected_embed_dim, rnn_hidden_dim, ctc_vocab_size)
+        self.pool = nn.AdaptiveAvgPool2d((1, 1))
+        
+    def forward(self, x):
+        features = self.backbone.forward_features(x)
+        pooled_features = self.pool(features).flatten(1)
+        type_logits = self.type_head(pooled_features)
+        object_logits = self.object_head(pooled_features)
+        proj_features = self.input_proj(features)
+        bs, c_proj, h_feat, w_feat = proj_features.size()
+        image_features_seq = proj_features.view(bs, c_proj, h_feat * w_feat).permute(0, 2, 1)
+        text_log_probs = self.text_head_ctc(image_features_seq)
+        return type_logits, object_logits, text_log_probs
+
+# ==============================================================================
+#  BAGIAN 3: FUNGSI HELPER (TIDAK ADA PERUBAHAN)
+# ==============================================================================
+
+def get_transforms(img_height, img_width):
+    interpolation_method = cv2.INTER_AREA
+    return A.Compose([
+        A.Resize(height=img_height, width=img_width, interpolation=interpolation_method),
+        A.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
+        ToTensorV2()
+    ])
+
+def ctc_decoder_with_confidence(log_probs, idx_to_char_map, blank_idx):
+    probs = torch.exp(log_probs)
+    max_probs, pred_indices = torch.max(probs, dim=-1)
+    max_probs = max_probs.squeeze(1).cpu().numpy()
+    pred_indices = pred_indices.squeeze(1).cpu().numpy()
+    
+    decoded_sequence = []
+    confidence_values = []
+    last_idx = -1
+
+    for i, idx in enumerate(pred_indices):
+        if idx == blank_idx or idx == last_idx:
+            last_idx = blank_idx if idx == blank_idx else last_idx
+            continue
+        
+        decoded_sequence.append(idx_to_char_map.get(str(idx), '?'))
+        confidence_values.append(max_probs[i])
+        last_idx = idx
+    
+    final_text = "".join(decoded_sequence)
+    avg_confidence = np.mean(confidence_values) if confidence_values else 0.0
+    
+    return final_text, avg_confidence
+
+# ==============================================================================
+#  BAGIAN 4: INISIALISASI SERVER (VERSI UNTUK MODEL KUANTISASI)
+# ==============================================================================
+
+def initialize_server(model_path, mappings_path):
+    global MODEL, MAPPINGS, DEVICE, TRANSFORMS
+    logging.info("Memulai inisialisasi server dengan model terkuantisasi...")
+    DEVICE = torch.device("cpu") # Kuantisasi INT8 dioptimalkan untuk CPU
+    logging.info(f"Menggunakan device: {DEVICE}")
+
+    try:
+        if not os.path.exists(mappings_path):
+            raise FileNotFoundError(f"File mappings tidak ditemukan di: {mappings_path}")
+        with open(mappings_path, 'r', encoding='utf-8') as f:
+            MAPPINGS = json.load(f)
+        logging.info("File mappings berhasil dimuat.")
+    except Exception as e:
+        logging.error(f"FATAL: Gagal memuat file mappings: {e}")
+        sys.exit(1)
+
+    TRANSFORMS = get_transforms(MAPPINGS['img_height'], MAPPINGS['img_width'])
+
+    # Langkah 1: Buat instance model dengan arsitektur asli
+    try:
+        m = MAPPINGS
+        model_to_quantize = MultiHeadModel(
+            backbone_name=m['backbone'],
+            ctc_vocab_size=len(m['ctc_char_to_idx']),
+            num_object_classes=len(m['object_to_idx']),
+            num_types=len(m['type_to_idx'])
+        )
+        logging.info(f"Instance model '{MAPPINGS['backbone']}' berhasil dibuat.")
+    except Exception as e:
+        logging.error(f"FATAL: Gagal membuat instance model. Error: {e}")
+        sys.exit(1)
+
+    # Langkah 2: Siapkan model untuk menerima weights terkuantisasi
+    MODEL = torch.quantization.quantize_dynamic(
+        model_to_quantize, {nn.Linear, nn.LSTM}, dtype=torch.qint8
+    )
+    logging.info("Arsitektur model disiapkan untuk kuantisasi dinamis.")
+        
+    # Langkah 3: Muat state_dict dari file .pth yang sudah terkuantisasi
+    try:
+        if not os.path.exists(model_path):
+             raise FileNotFoundError(f"File model tidak ditemukan di: {model_path}")
+        
+        # Langsung load state_dict karena kita sudah menyiapkan arsitekturnya
+        MODEL.load_state_dict(torch.load(model_path, map_location=DEVICE))
+        MODEL.to(DEVICE)
+        MODEL.eval()
+        logging.info("Model weights terkuantisasi berhasil dimuat dan siap digunakan.")
+
+    except Exception as e:
+        logging.error(f"FATAL: Gagal memuat model weights terkuantisasi. Error: {e}", exc_info=True)
+        sys.exit(1)
+    
+    logging.info("Inisialisasi server selesai. Siap menerima permintaan.")
+
+# ==============================================================================
+#  BAGIAN 5: ENDPOINT FLASK (TIDAK ADA PERUBAHAN)
+# ==============================================================================
+@app.route('/', methods=['GET'])
+def home():
+    """Endpoint dasar untuk memeriksa apakah server berjalan."""
+    return "<h1>Captcha Prediction Server is running (Quantized Model).</h1><p>Gunakan endpoint /predict untuk melakukan prediksi.</p>", 200
+
+@app.route('/predict', methods=['POST'])
+def predict_endpoint():
+    """Endpoint untuk menerima gambar base64 dan mengembalikan prediksi."""
+    
+    # Otentikasi
+    expected_api_key = os.environ.get('API_KEY_SECRET')
+    if not expected_api_key:
+        logging.error("FATAL: Secret 'API_KEY_SECRET' tidak diatur di server.")
+        return jsonify({"error": "Konfigurasi server error."}), 500
+
+    auth_header = request.headers.get('Authorization')
+    if not auth_header or auth_header != f"Bearer {expected_api_key}":
+        logging.warning(f"Akses ditolak untuk IP: {request.remote_addr}. Alasan: Kunci API tidak valid.")
+        return jsonify({"error": "Akses ditolak."}), 403
+
+    # Proses prediksi
+    if not request.is_json:
+        return jsonify({"error": "Request harus berupa JSON"}), 400
+    
+    data = request.get_json()
+    base64_string = data.get('image_base64')
+
+    if not base64_string:
+        return jsonify({"error": "Key 'image_base64' tidak ditemukan atau kosong"}), 400
+
+    try:
+        if ',' in base64_string:
+            _, encoded = base64_string.split(',', 1)
+        else:
+            encoded = base64_string
+        image_data = base64.b64decode(encoded)
+        img_pil = Image.open(BytesIO(image_data)).convert("RGB")
+            
+    except (base64.binascii.Error, UnidentifiedImageError) as e:
+        logging.error(f"Error memproses gambar base64: {e}")
+        return jsonify({"error": f"Data base64 tidak valid atau format gambar tidak didukung."}), 400
+    except Exception as e:
+        logging.error(f"Error tak terduga saat memproses gambar: {e}")
+        return jsonify({"error": "Gagal memproses gambar."}), 500
+
+    try:
+        image_rgb = np.array(img_pil)
+        img_tensor = TRANSFORMS(image=image_rgb)['image'].unsqueeze(0).to(DEVICE)
+
+        with torch.no_grad():
+            type_logits, object_logits, text_log_probs = MODEL(img_tensor)
+
+        type_prob = torch.softmax(type_logits, dim=1)
+        type_conf, type_pred_idx = torch.max(type_prob, dim=1)
+        pred_type = MAPPINGS['idx_to_type'].get(str(type_pred_idx.item()), 'Tipe Tidak Dikenal')
+        
+        response = {
+            "predicted_type": pred_type,
+            "type_confidence": f"{type_conf.item():.2%}",
+            "prediction": None,
+            "prediction_confidence": None,
+            "error": None
+        }
+
+        if pred_type == 'object':
+            obj_prob = torch.softmax(object_logits, dim=1)
+            obj_conf, obj_pred_idx = torch.max(obj_prob, dim=1)
+            pred_obj = MAPPINGS['idx_to_object'].get(str(obj_pred_idx.item()), 'Objek Tidak Dikenal')
+            response["prediction"] = pred_obj
+            response["prediction_confidence"] = f"{obj_conf.item():.2%}"
+        elif pred_type == 'text':
+            pred_text, confidence = ctc_decoder_with_confidence(text_log_probs, MAPPINGS['ctc_idx_to_char'], MAPPINGS['ctc_blank_idx'])
+            response["prediction"] = pred_text
+            response["prediction_confidence"] = f"{confidence:.2%}"
+        
+        logging.info(f"Prediksi berhasil: Tipe='{response['predicted_type']}', Hasil='{response['prediction']}', Conf='{response['prediction_confidence']}'")
+        return jsonify(response), 200
+
+    except Exception as e:
+        logging.error(f"Error saat inferensi model: {e}", exc_info=True)
+        return jsonify({"error": "Terjadi kesalahan pada server saat melakukan prediksi."}), 500
+
+# ==============================================================================
+#  BAGIAN 6: MENJALANKAN SERVER (UNTUK HUGGING FACE SPACES)
+# ==============================================================================
+
+# Menggunakan file model baru yang sudah terkuantisasi
+MODEL_FILE_PATH = "best_model_quantized.pth"
+MAPPINGS_FILE_PATH = "mappings.json"
+
+# Inisialisasi server saat aplikasi dimulai
+initialize_server(MODEL_FILE_PATH, MAPPINGS_FILE_PATH)
+
+# Berguna untuk pengujian lokal
+if __name__ == '__main__':
+    app.run(host='0.0.0.0', port=5111, debug=True)

best_model_quantized.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:355509fbab22f2773d1c5ca340feb261e5682f69cfb688b481787ae77da9e4d0
+size 46128300

mappings.json

@@ -0,0 +1,357 @@
+{
+    "backbone": "tf_efficientnet_b3.ns_jft_in1k",
+    "img_height": 416,
+    "img_width": 416,
+    "type_to_idx": {
+        "object": 0,
+        "text": 1
+    },
+    "idx_to_type": {
+        "0": "object",
+        "1": "text"
+    },
+    "object_to_idx": {
+        "2": 0,
+        "Apple": 1,
+        "Bear": 2,
+        "Bee": 3,
+        "Blank": 4,
+        "Bus": 5,
+        "Butterfly": 6,
+        "Car": 7,
+        "Cat": 8,
+        "Cow": 9,
+        "Crab": 10,
+        "Crown": 11,
+        "DOG": 12,
+        "Dinosaurs": 13,
+        "Dog": 14,
+        "Factory": 15,
+        "Guitar": 16,
+        "Helicopter": 17,
+        "Horse": 18,
+        "House": 19,
+        "Koala": 20,
+        "Motorcycle": 21,
+        "Penguin": 22,
+        "Rabbit": 23,
+        "Shark": 24,
+        "Speaker": 25,
+        "Statue": 26,
+        "Watermelon": 27,
+        "Wine": 28,
+        "Wolf": 29,
+        "atv": 30,
+        "bee": 31,
+        "bicycle": 32,
+        "bird": 33,
+        "butterfly": 34,
+        "cart": 35,
+        "chicken": 36,
+        "crocodile": 37,
+        "fish": 38,
+        "flower": 39,
+        "koala": 40,
+        "monster car": 41,
+        "motorcycle": 42,
+        "mouse": 43,
+        "panda": 44,
+        "snowman": 45,
+        "tiger": 46,
+        "tractor": 47
+    },
+    "idx_to_object": {
+        "0": "2",
+        "1": "Apple",
+        "2": "Bear",
+        "3": "Bee",
+        "4": "Blank",
+        "5": "Bus",
+        "6": "Butterfly",
+        "7": "Car",
+        "8": "Cat",
+        "9": "Cow",
+        "10": "Crab",
+        "11": "Crown",
+        "12": "DOG",
+        "13": "Dinosaurs",
+        "14": "Dog",
+        "15": "Factory",
+        "16": "Guitar",
+        "17": "Helicopter",
+        "18": "Horse",
+        "19": "House",
+        "20": "Koala",
+        "21": "Motorcycle",
+        "22": "Penguin",
+        "23": "Rabbit",
+        "24": "Shark",
+        "25": "Speaker",
+        "26": "Statue",
+        "27": "Watermelon",
+        "28": "Wine",
+        "29": "Wolf",
+        "30": "atv",
+        "31": "bee",
+        "32": "bicycle",
+        "33": "bird",
+        "34": "butterfly",
+        "35": "cart",
+        "36": "chicken",
+        "37": "crocodile",
+        "38": "fish",
+        "39": "flower",
+        "40": "koala",
+        "41": "monster car",
+        "42": "motorcycle",
+        "43": "mouse",
+        "44": "panda",
+        "45": "snowman",
+        "46": "tiger",
+        "47": "tractor"
+    },
+    "ctc_char_to_idx": {
+        " ": 1,
+        "!": 2,
+        "#": 3,
+        "$": 4,
+        "%": 5,
+        "&": 6,
+        "(": 7,
+        ")": 8,
+        "*": 9,
+        "+": 10,
+        ",": 11,
+        "-": 12,
+        ".": 13,
+        "0": 14,
+        "1": 15,
+        "2": 16,
+        "3": 17,
+        "4": 18,
+        "5": 19,
+        "6": 20,
+        "7": 21,
+        "8": 22,
+        "9": 23,
+        ":": 24,
+        "=": 25,
+        "?": 26,
+        "@": 27,
+        "A": 28,
+        "B": 29,
+        "C": 30,
+        "D": 31,
+        "E": 32,
+        "F": 33,
+        "G": 34,
+        "H": 35,
+        "I": 36,
+        "J": 37,
+        "K": 38,
+        "L": 39,
+        "M": 40,
+        "N": 41,
+        "O": 42,
+        "P": 43,
+        "Q": 44,
+        "R": 45,
+        "S": 46,
+        "T": 47,
+        "U": 48,
+        "V": 49,
+        "W": 50,
+        "X": 51,
+        "Y": 52,
+        "Z": 53,
+        "[": 54,
+        "\\": 55,
+        "^": 56,
+        "_": 57,
+        "a": 58,
+        "b": 59,
+        "c": 60,
+        "d": 61,
+        "e": 62,
+        "f": 63,
+        "g": 64,
+        "h": 65,
+        "i": 66,
+        "j": 67,
+        "k": 68,
+        "l": 69,
+        "m": 70,
+        "n": 71,
+        "o": 72,
+        "p": 73,
+        "q": 74,
+        "r": 75,
+        "s": 76,
+        "t": 77,
+        "u": 78,
+        "v": 79,
+        "w": 80,
+        "x": 81,
+        "y": 82,
+        "z": 83,
+        "{": 84,
+        "}": 85,
+        "÷": 86,
+        "Б": 87,
+        "Г": 88,
+        "Д": 89,
+        "Ж": 90,
+        "З": 91,
+        "И": 92,
+        "Й": 93,
+        "Л": 94,
+        "О": 95,
+        "Ф": 96,
+        "Ч": 97,
+        "Ш": 98,
+        "Ю": 99,
+        "Я": 100,
+        "д": 101,
+        "ж": 102,
+        "ш": 103,
+        "٣": 104,
+        "٤": 105,
+        "٩": 106,
+        "ข": 107,
+        "ถ": 108,
+        "บ": 109,
+        "ร": 110,
+        "ว": 111,
+        "ห": 112,
+        "与": 113,
+        "养": 114,
+        "決": 115,
+        "海": 116,
+        "的": 117,
+        "窄": 118,
+        "<blank>": 0
+    },
+    "ctc_idx_to_char": {
+        "1": " ",
+        "2": "!",
+        "3": "#",
+        "4": "$",
+        "5": "%",
+        "6": "&",
+        "7": "(",
+        "8": ")",
+        "9": "*",
+        "10": "+",
+        "11": ",",
+        "12": "-",
+        "13": ".",
+        "14": "0",
+        "15": "1",
+        "16": "2",
+        "17": "3",
+        "18": "4",
+        "19": "5",
+        "20": "6",
+        "21": "7",
+        "22": "8",
+        "23": "9",
+        "24": ":",
+        "25": "=",
+        "26": "?",
+        "27": "@",
+        "28": "A",
+        "29": "B",
+        "30": "C",
+        "31": "D",
+        "32": "E",
+        "33": "F",
+        "34": "G",
+        "35": "H",
+        "36": "I",
+        "37": "J",
+        "38": "K",
+        "39": "L",
+        "40": "M",
+        "41": "N",
+        "42": "O",
+        "43": "P",
+        "44": "Q",
+        "45": "R",
+        "46": "S",
+        "47": "T",
+        "48": "U",
+        "49": "V",
+        "50": "W",
+        "51": "X",
+        "52": "Y",
+        "53": "Z",
+        "54": "[",
+        "55": "\\",
+        "56": "^",
+        "57": "_",
+        "58": "a",
+        "59": "b",
+        "60": "c",
+        "61": "d",
+        "62": "e",
+        "63": "f",
+        "64": "g",
+        "65": "h",
+        "66": "i",
+        "67": "j",
+        "68": "k",
+        "69": "l",
+        "70": "m",
+        "71": "n",
+        "72": "o",
+        "73": "p",
+        "74": "q",
+        "75": "r",
+        "76": "s",
+        "77": "t",
+        "78": "u",
+        "79": "v",
+        "80": "w",
+        "81": "x",
+        "82": "y",
+        "83": "z",
+        "84": "{",
+        "85": "}",
+        "86": "÷",
+        "87": "Б",
+        "88": "Г",
+        "89": "Д",
+        "90": "Ж",
+        "91": "З",
+        "92": "И",
+        "93": "Й",
+        "94": "Л",
+        "95": "О",
+        "96": "Ф",
+        "97": "Ч",
+        "98": "Ш",
+        "99": "Ю",
+        "100": "Я",
+        "101": "д",
+        "102": "ж",
+        "103": "ш",
+        "104": "٣",
+        "105": "٤",
+        "106": "٩",
+        "107": "ข",
+        "108": "ถ",
+        "109": "บ",
+        "110": "ร",
+        "111": "ว",
+        "112": "ห",
+        "113": "与",
+        "114": "养",
+        "115": "決",
+        "116": "海",
+        "117": "的",
+        "118": "窄",
+        "0": "<blank>"
+    },
+    "ctc_blank_idx": 0,
+    "max_text_len": 30
+}

requirements.txt

@@ -0,0 +1,12 @@
+# Versi paket dikunci untuk stabilitas build
+flask==2.3.2
+flask-cors
+gunicorn==21.2.0
+numpy==1.24.3
+Pillow==10.0.0
+# Versi PyTorch tanpa flag spesifik, lebih portabel
+torch==2.0.1
+torchvision==0.15.2
+timm
+albumentations
+opencv-python-headless==4.8.0.74