Beberapa bulan yang lalu penulis dan teman - teman yang tergabung dalam sebuah tim mengikuti lomba Big Data Competition di acara Satria Data 2021 yang diselenggarakan oleh PUSPRESNAS yang bekerjasama dengan IPB University. Ini merupakan kali kedua penulis mengikuti lomba BDC - Satria Data, dengan susunan tim yang sedikit berbeda dan task yang berbeda juga penulis mendapatkan banyak pengalaman dan ilmu dari lomba tersebut. Meskipun belum diberikan rezeki untuk menjadi juara Alhamdulillah tim kami berhasil melaju sampai babak final.

Berikut ini adalah kode dan langkah - langkah yang tim kami gunakan pada saat mengikuti lomba BDC Satria Data 2021.

Author

• Naufal Zhafran Albaqi
• Muhammad Amanda
• Wahyu Setianto

Running Environment

• Platform : Kaggle
• Accelerator : GPU NVIDIA TESLA P100

Main Dataset

Dataset hasil preprocessing dan augmentasi dapat diakses lewat platform berikut.

Gender Classification

Pada tahap penyisihan kami ditugaskan untuk melakukan pengklasifikasian terhadap gender seseorang didalam foto (binary classification). Data yang disediakan merupakan data gambar dengan ukuran dan resolusi yang berbeda sehingga merupakan tantangan tersendiri untuk melakukan data preprocessing. Pada data yang disediakan juga terdapat beberapa gambar yang memiliki lebih dari satu orang di dalamnya sehingga menurut tim kami cukup sulit jika membersihkan data secara otomatis dan perlu dilakukan manual filtering untuk meningkatkan kualitas data sebelum dilatih kedalam model. Kami juga melakukan proses augmentasi sehingga data semakin banyak dan lebih banyak yang dapat dipelajari oleh model. Model yang digunakan oleh tim kami dalam menyelesaikan masalah ini adalah tensorflow custom model dengan arsitektur 5 convolution layer yang dipadukan dengan max pooling layer dan ditutup dengan dense layer berkativasi sigmoid. Ditahap akhir untuk meningkatkan kualitas prediksi kami melakukan ensembel terhadap hasil prediksi 5 model yang dilatih dengan data yang berbeda (menggunakan StratifiedKfold).

First Thing First

Menginstall library yang diperlukan dan mengimport library - library yang akan digunakan serta menseting variable config yang akan digunakan di dalam notebook ini.

1. Menginstal library MTCNN dan Albumentations

Library MTCNN dan Albumentations adalah library yang digunakan untuk preprocessing data gambar pada notebook ini

$ python
!pip -q install mtcnn --upgrade
!pip -q install albumentations --upgrade

WARNING: Running pip as the 'root' user can result in broken permissions and conflicting behaviour with the system package manager. It is recommended to use a virtual environment instead: https://pip.pypa.io/warnings/venv
WARNING: Running pip as the 'root' user can result in broken permissions and conflicting behaviour with the system package manager. It is recommended to use a virtual environment instead: https://pip.pypa.io/warnings/venv

2. Importing library

Mengimport library yang akan digunakan dalam notebook ini.

$ python
# Umum
import os, random, re
from tqdm.notebook import tqdm
import numpy as np
import pandas as pd
from PIL import Image

# Tensorflow
import tensorflow as tf
from tensorflow.keras.preprocessing.image import load_img, img_to_array

# Metrics & Splitting data
from sklearn.metrics import *
from sklearn.model_selection import *

# Plotting
import matplotlib.pyplot as plt
import seaborn as sns

# Preprocessing
import cv2
from mtcnn import MTCNN
import albumentations as A

print("Tensorflow :", tf.__version__)

Tensorflow : 2.6.0

3. Setup CONFIG

Mensetup varible - variable yang digunakan sebagai config pada notebook ini

$ python
SEED = 2021
SIZE = (200, 200)
BATCH_SIZE = 32
FACE_THRESHOLD = 0.95
FACE_DETECTOR = MTCNN()

2022-03-25 05:01:04.404789: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2022-03-25 05:01:04.507729: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2022-03-25 05:01:04.508532: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2022-03-25 05:01:04.510232: I tensorflow/core/platform/cpu_feature_guard.cc:142] This TensorFlow binary is optimized with oneAPI Deep Neural Network Library (oneDNN) to use the following CPU instructions in performance-critical operations:  AVX2 AVX512F FMA
To enable them in other operations, rebuild TensorFlow with the appropriate compiler flags.
2022-03-25 05:01:04.511454: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2022-03-25 05:01:04.512136: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2022-03-25 05:01:04.512828: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2022-03-25 05:01:06.301278: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2022-03-25 05:01:06.302130: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2022-03-25 05:01:06.302815: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2022-03-25 05:01:06.303445: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1510] Created device /job:localhost/replica:0/task:0/device:GPU:0 with 15403 MB memory:  -> device: 0, name: Tesla P100-PCIE-16GB, pci bus id: 0000:00:04.0, compute capability: 6.0

Dataset

Menggunakan dataset yang telah digunakan oleh tim kami dan melakukan loading agar mendapatkan informasi path dari data gambar

Dataset : Data SD20210000722

$ python
train = pd.read_csv("../input/bdc-2021/train.csv")
test = pd.read_csv("../input/bdc-2021/submission.csv")
train.head()

memperjelas path ke setiap data gambar

$ python
images = []
labels = []
test_images = []

TRAIN_DIR = "../input/bdc-2021/Training"
TEST_DIR = "../input/bdc-2021/Testing"

for no, label in train[["nomor", "jenis kelamin"]].values:
    TEMP_DIR = os.path.join(TRAIN_DIR, str(no))
    for file in os.listdir(TEMP_DIR):
        file_dir = os.path.join(TEMP_DIR, file)
        if ".ini" not in file_dir:
            images.append(file_dir)
            labels.append(label)

for no in test.id.values:
    file_dir = os.path.join(TEST_DIR, f"{no}.jpg")
    if os.path.isfile(file_dir):
        test_images.append(file_dir)
    else:
        test_images.append(None)
        print(file_dir)

menampilkan dan mengecek beberapa gambar pada data train

$ python
def read(path):
    """
    Read data gambar
    """
    img = Image.open(path)
    return img

def show_images(list_dir, label = None, load_image = read, seed = SEED):
    """
    Menampilkan Gambar Secara acak sebanyak 5 buah.
    """
    random.seed(seed)
    unique = ["init"]
    if label:
        unique = list(set(label))
    fig, axes = plt.subplots(len(unique), 5, figsize = (20, 5 * len(unique)))
    for i in range(len(unique)):
        if i == 0 and unique[i] == "init":
            data = random.sample(list_dir, 5)
        else:
            data = random.sample([x for x in zip(list_dir, label) if x[1] == unique[i]], 5)
        for j in range(5):
            if unique[0] != "init":
                img = load_image(data[j][0])
                axes[i, j].imshow(img)
                axes[i, j].set_title(f'Label : {data[j][1]}', fontsize = 14)
                axes[i, j].axis('off')
            else:
                img = load_image(data[j])
                axes[j].imshow(img)
                axes[j].axis('off')
    fig.tight_layout()
    plt.show()

$ python
show_images(images, labels, seed=20)

Preprocess Data

Metode yang digunakan:

1. Mengekstrak wajah - wajah yang terdapat pada gambar menjadi gambar - gambar baru dengan label yang sama dengan menggunakan model MTCNN
2. Pada data test jika terdapat dua wajah yang terdeteksi pada satu gambar akan di ambil wajah dengan tingkat confidence terbesar yang diberikan oleh model MTCNN.
3. Jika tidak terdetect wajah pada salah satu gambar maka akan dilakukan crop pada bagian tengah gambar sehingga gambar berbentuk persegi atau jxj pixel.
4. Selanjutnya gambar akan di resize menjadi ukuran 256x256 pixel

berikut adalah contoh hasil preprocess data gambar.

$ python
def get_faces(path):
    image = cv2.cvtColor(cv2.imread(path), cv2.COLOR_BGR2RGB)
    faces = FACE_DETECTOR.detect_faces(image)
    return faces

def load_and_preprocess_image(path: str, size = SIZE):
    """
    Load & Preprocess data gambar
    """
    image = img_to_array(load_img(path))
    faces = [x['box'] for x in get_faces(path) if x['confidence'] > FACE_THRESHOLD]
    if len(faces) > 0:
        x, y, w, h = faces[0]
        image = image[y:y+h, x:x+w]
    img = tf.convert_to_tensor(image, dtype=tf.float32)
    if len(faces) == 0:
        shapes = tf.shape(img)
        h, w = shapes[-3], shapes[-2]
        dim = tf.minimum(h, w)
        img = tf.image.resize_with_crop_or_pad(img, dim, dim)
    img = tf.image.resize(img, size)
    img = tf.cast(img, tf.float32) / 255.0
    return img.numpy()

$ python
show_images(images, labels, load_image = load_and_preprocess_image, seed=20)

2022-03-25 05:01:13.728692: I tensorflow/compiler/mlir/mlir_graph_optimization_pass.cc:185] None of the MLIR Optimization Passes are enabled (registered 2)
2022-03-25 05:01:14.721635: I tensorflow/stream_executor/cuda/cuda_dnn.cc:369] Loaded cuDNN version 8005

Augmentasi Data

Melakukan augmentasi untuk memperbanyak data. Metode augmentasi yang digunakan yaitu:

1. Horizontal flip
2. Donwscale kualitas gambar
3. Random rotate dengan rentang -30 sampai 30 derajad
4. Shift, scale, dan rotate gambar
5. Blur
6. Random brightness

$ python
aug = A.Compose([
    A.HorizontalFlip(p=0.4),
    A.Downscale(scale_min=0.6, scale_max=0.9, p=0.3),
    A.Rotate(limit=(-30,30), p=0.6),
    A.ShiftScaleRotate(shift_limit=(-0.07, 0.07), scale_limit=(-0.05, 0.1), rotate_limit=(-15, 15), p=0.4),
    A.OneOf([
        A.MotionBlur(p=.4),
        A.MedianBlur(blur_limit=3, p=0.4),
        A.Blur(blur_limit=3, p=0.4),
    ], p=0.4),
    A.RandomBrightnessContrast(brightness_limit=(-0.25, 0.15), p=0.4),
])

$ python
def visualize_aug(path):
    fig, axes = plt.subplots(1, 5, figsize = (20, 5))
    image = load_and_preprocess_image(path)
    axes[0].imshow(image)
    axes[0].axis('off')
    for i in range(1, 5):
        augmented = aug(image=image)['image']
        axes[i].imshow(augmented)
        axes[i].axis('off')
    fig.tight_layout()
    plt.show()

$ python
random.seed(SEED)
for i in range(3):
    visualize_aug(images[i])

Running preprocessing pada data gambar secara keseluruhan

$ python
def image_preprocessing(new_dir, images, labels=None):
    if os.path.isdir(new_dir):
        !rm -rf {new_dir}
    os.mkdir(new_dir)

    new_images, new_labels = [], []
    if not labels:
        labels = [None for _ in range(len(images))]

    for path, label in tqdm(zip(images, labels), total=len(images)):
        image = img_to_array(load_img(path))
        if label != None:
            faces = [x['box'] for x in sorted(get_faces(path), key=lambda x: x['confidence'],
                                              reverse=True) if x['confidence'] > FACE_THRESHOLD]
        else:
            faces = [x['box'] for x in sorted(get_faces(path), key=lambda x: x['confidence'], reverse=True)]
        if len(faces) > 0:
            if label != None:
                for j, (x, y, w, h) in enumerate(faces):
                    img = image[y:y+h, x:x+w]
                    img = tf.convert_to_tensor(img, dtype=tf.float32)
                    img = tf.image.resize(img, SIZE)
                    img = tf.cast(img, tf.float32) / 255.0

                    img_dir = os.path.join(new_dir, f'{j}_{path.split("/")[-1]}')
                    new_images.append(img_dir)
                    new_labels.append(label)
                    tf.keras.preprocessing.image.save_img(img_dir, img)

                    for k in range(3):
                        augmented = aug(image=img.numpy())['image']
                        img_dir = os.path.join(new_dir, f'aug-{k}_{j}_{path.split("/")[-1]}')
                        new_images.append(img_dir)
                        new_labels.append(label)
                        tf.keras.preprocessing.image.save_img(img_dir, augmented)
            else:
                x, y, w, h = faces[0]
                img = image[y:y+h, x:x+w]
                img = tf.convert_to_tensor(img, dtype=tf.float32)
                img = tf.image.resize(img, SIZE)
                img = tf.cast(img, tf.float32) / 255.0

                img_dir = os.path.join(new_dir, path.split('/')[-1])
                new_images.append(img_dir)
                new_labels.append(label)
                tf.keras.preprocessing.image.save_img(img_dir, img)
        else :
            img = tf.convert_to_tensor(image, dtype=tf.float32)
            shapes = tf.shape(img)
            h, w = shapes[-3], shapes[-2]
            dim = tf.minimum(h, w)
            img = tf.image.resize_with_crop_or_pad(img, dim, dim)
            img = tf.image.resize(img, SIZE)
            img = tf.cast(img, tf.float32) / 255.0

            img_dir = os.path.join(new_dir, path.split('/')[-1])
            new_images.append(img_dir)
            new_labels.append(label)
            tf.keras.preprocessing.image.save_img(img_dir, img)

            if label != None:
                for k in range(3):
                    augmented = aug(image=img.numpy())['image']
                    img_dir = os.path.join(new_dir,  f'aug-{k}_{path.split("/")[-1]}')
                    new_images.append(img_dir)
                    new_labels.append(label)
                    tf.keras.preprocessing.image.save_img(img_dir, augmented)

    return new_images, new_labels

Untuk menghemat waktu running akan di skip bagian ini dan di ganti dengan meload data hasil preprocess yang sudah di save pada run sebelumnya. Namun jika ingin melakukan preprocess pada run sekarang maka uncomment code di bawah ini.

Peringatan : running block code di bawah memakan waktu sekitar 50 menit dengan GPU Nvidia Tesla P100-PCIE.

$ python
# new_train_dir = "./train"
# new_test_dir = "./test"

# random.seed(SEED)
# new_images, new_labels = image_preprocessing(new_train_dir, images, labels)
# new_test_images, _ = image_preprocessing(new_test_dir, test_images)

Note : Comment dua block kode di bawah jika melakukan preprocess pada run saat ini.

$ python
preprocessed = pd.read_csv("../input/bdc-2021/preprocessed-augmented/preprocessed.csv")
preprocessed.head()

Filterisasi Data

Menyaring data yang akan dilatih ke dalam model dengan melakukan filterisasi terhadap data gambar yang memiliki kualitas gambar yang kurang baik dan misslabel.

1. Mendownload list data gambar yang akan dibuang.

$ python
!wget https://raw.githubusercontent.com/Hyuto/bdc-2021/master/excluded-gender.txt

with open("./excluded-gender.txt") as f:
    excluded = f.read().split("\n")

patterns = fr'{"|".join(excluded)}'

--2022-03-25 05:01:44--  https://raw.githubusercontent.com/Hyuto/bdc-2021/master/excluded-gender.txt
Resolving raw.githubusercontent.com (raw.githubusercontent.com)... 185.199.110.133, 185.199.111.133, 185.199.108.133, ...
Connecting to raw.githubusercontent.com (raw.githubusercontent.com)|185.199.110.133|:443... connected.
HTTP request sent, awaiting response... 200 OK
Length: 3111 (3.0K) [text/plain]
Saving to: ‘excluded-gender.txt’

excluded-gender.txt 100%[===================>]   3.04K  --.-KB/s    in 0s

2022-03-25 05:01:44 (23.6 MB/s) - ‘excluded-gender.txt’ saved [3111/3111]

2. Filterisasi data

$ python
preprocessed_dir = "../input/bdc-2021/preprocessed-augmented"
new_images, new_labels = [], []

for image, label in preprocessed[["image", "label"]].values:
    if not re.search(patterns, image):
        new_images.append(os.path.join(preprocessed_dir, image))
        new_labels.append(label)

new_test_images = np.asarray([os.path.join(preprocessed_dir, "test", f"{x}.jpg") for x in test.id.values])

new_images = np.asarray(new_images)
new_labels = np.asarray(new_labels)

Mengecek distribusi label pada data

$ python
plt.figure(figsize=(5, 5))
sns.countplot(x=new_labels)
plt.show()

Jumlah data yang berlabel 0 dan 1 cenderung sama.

Modelling

Membuat model dengan arsitektur sebagai berikut

$ python
def build_model(kernel_s=(3,3)):
    model = tf.keras.models.Sequential([
        tf.keras.layers.Conv2D(32,kernel_s,activation='relu',input_shape=(200,200,3),
                            kernel_regularizer=tf.keras.regularizers.l2(0.001),padding="VALID"),
        tf.keras.layers.MaxPooling2D((2,2)),
        tf.keras.layers.Conv2D(64,kernel_s,activation='relu'),
        tf.keras.layers.MaxPooling2D((2,2)),
        tf.keras.layers.Conv2D(64,kernel_s,activation='relu'),
        tf.keras.layers.MaxPooling2D((2,2)),
        tf.keras.layers.Conv2D(128,kernel_s,activation='relu'),
        tf.keras.layers.MaxPooling2D((2,2)),
        tf.keras.layers.Conv2D(128,kernel_s,activation='relu'),
        tf.keras.layers.MaxPooling2D((2,2)),
        tf.keras.layers.Flatten(),
        tf.keras.layers.Dense(256, activation='relu',
                              kernel_regularizer=tf.keras.regularizers.l2(5e-4)),
        tf.keras.layers.Dense(1, activation='sigmoid')
    ])
    model.compile(loss="binary_crossentropy", optimizer="adam", metrics=['accuracy'])
    return model

Tensorflow Data

Load data gambar menggunakan Tensorflow Data agar pada saat pelatihan model penggunaan memmori dapat lebih optimal

$ python
def decode_image(filename, label=None, image_size=SIZE):
    """
    Decode Image from String Path Tensor
    """
    bits = tf.io.read_file(filename)
    image = tf.image.decode_jpeg(bits, channels=3)
    image = tf.cast(image, tf.float32) / 255.0
    image = tf.image.resize(image, SIZE)

    if label is None: # if test
        return image
    else:
        return image, label

Training Model

Proses ini dilakukan dengan mensplit data menggunakan StratifiedKFold sebanyak 5 split lalu untuk setiap split akan dibangun sebuah model untuk dilatih. Setiap model ini akan digunakan untuk melakukan peramalan terhadap data test. Pada proses trainingnya setiap model akan diukur tingkat kebaikannya dengan menggunakan metric accuracy lalu dikakukan penyimpanan weight pada model saat val_accuracy berada pada nilai terbesar selama pelatihan model sebanyak 50 epochs.

$ python
split = 5
prediksi = np.zeros((len(new_test_images), 1), dtype=np.float32)
acc_scores, f1_scores = [], []

test_dataset = (
    tf.data.Dataset
    .from_tensor_slices((new_test_images))
    .map(decode_image)
    .batch(BATCH_SIZE)
)

cv = StratifiedKFold(n_splits=split, shuffle=True, random_state=SEED)
for i, (train_index, test_index) in enumerate(cv.split(new_images, new_labels)):
    tf.keras.backend.clear_session()
    x_train, x_valid = new_images[train_index], new_images[test_index]
    y_train, y_valid = new_labels[train_index], new_labels[test_index]

    train_dataset = (
        tf.data.Dataset
        .from_tensor_slices((x_train, y_train))
        .map(decode_image, num_parallel_calls=tf.data.AUTOTUNE)
        .cache()
        .repeat()
        .shuffle(1024)
        .batch(BATCH_SIZE)
        .prefetch(tf.data.AUTOTUNE)
    )

    valid_dataset = (
        tf.data.Dataset
        .from_tensor_slices((x_valid, y_valid))
        .map(decode_image, num_parallel_calls=tf.data.AUTOTUNE)
        .batch(BATCH_SIZE)
        .cache()
        .prefetch(tf.data.AUTOTUNE)
    )

    model = build_model()
    checkpoint = tf.keras.callbacks.ModelCheckpoint(f'{i}_best_model.h5', monitor='val_accuracy',
                                                save_best_only=True, save_weights_only=True,
                                                mode='max')
    print(f"\nCV {i+1}")
    model.fit(train_dataset, epochs=50, validation_data=valid_dataset,
              steps_per_epoch=len(x_train) // BATCH_SIZE,
              callbacks = [checkpoint])
    model.load_weights(f'{i}_best_model.h5')
    val_pred_classes = np.array(model.predict(valid_dataset).flatten() >= .5, dtype = 'int')
    acc, f1 = accuracy_score(y_valid, val_pred_classes), f1_score(y_valid, val_pred_classes)

    acc_scores.append(acc)
    f1_scores.append(f1)
    prediksi += model.predict(test_dataset)

    del train_dataset
    del valid_dataset

CV 1
Epoch 1/50
263/263 [==============================] - 32s 107ms/step - loss: 0.5496 - accuracy: 0.7317 - val_loss: 0.3710 - val_accuracy: 0.8544
Epoch 2/50
263/263 [==============================] - 7s 27ms/step - loss: 0.3210 - accuracy: 0.8819 - val_loss: 0.3272 - val_accuracy: 0.8753
Epoch 3/50
263/263 [==============================] - 7s 26ms/step - loss: 0.2792 - accuracy: 0.8977 - val_loss: 0.2815 - val_accuracy: 0.8957
Epoch 4/50
263/263 [==============================] - 7s 26ms/step - loss: 0.2451 - accuracy: 0.9177 - val_loss: 0.2392 - val_accuracy: 0.9189
Epoch 5/50
263/263 [==============================] - 7s 26ms/step - loss: 0.2160 - accuracy: 0.9268 - val_loss: 0.2512 - val_accuracy: 0.9151
Epoch 6/50
263/263 [==============================] - 7s 28ms/step - loss: 0.1843 - accuracy: 0.9437 - val_loss: 0.2210 - val_accuracy: 0.9265
Epoch 7/50
263/263 [==============================] - 7s 26ms/step - loss: 0.1688 - accuracy: 0.9470 - val_loss: 0.2120 - val_accuracy: 0.9336
Epoch 8/50
263/263 [==============================] - 7s 26ms/step - loss: 0.1480 - accuracy: 0.9578 - val_loss: 0.2075 - val_accuracy: 0.9346
Epoch 9/50
263/263 [==============================] - 7s 26ms/step - loss: 0.1341 - accuracy: 0.9627 - val_loss: 0.2290 - val_accuracy: 0.9294
Epoch 10/50
263/263 [==============================] - 7s 26ms/step - loss: 0.1163 - accuracy: 0.9701 - val_loss: 0.1956 - val_accuracy: 0.9417
Epoch 11/50
263/263 [==============================] - 7s 27ms/step - loss: 0.1062 - accuracy: 0.9718 - val_loss: 0.2243 - val_accuracy: 0.9369
Epoch 12/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0872 - accuracy: 0.9802 - val_loss: 0.2190 - val_accuracy: 0.9403
Epoch 13/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0953 - accuracy: 0.9781 - val_loss: 0.2050 - val_accuracy: 0.9507
Epoch 14/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0794 - accuracy: 0.9841 - val_loss: 0.2156 - val_accuracy: 0.9455
Epoch 15/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0732 - accuracy: 0.9874 - val_loss: 0.2320 - val_accuracy: 0.9426
Epoch 16/50
263/263 [==============================] - 7s 27ms/step - loss: 0.0663 - accuracy: 0.9870 - val_loss: 0.2245 - val_accuracy: 0.9502
Epoch 17/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0748 - accuracy: 0.9843 - val_loss: 0.2315 - val_accuracy: 0.9469
Epoch 18/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0637 - accuracy: 0.9884 - val_loss: 0.2470 - val_accuracy: 0.9497
Epoch 19/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0601 - accuracy: 0.9892 - val_loss: 0.2627 - val_accuracy: 0.9436
Epoch 20/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0570 - accuracy: 0.9901 - val_loss: 0.3084 - val_accuracy: 0.9360
Epoch 21/50
263/263 [==============================] - 7s 27ms/step - loss: 0.0535 - accuracy: 0.9916 - val_loss: 0.2784 - val_accuracy: 0.9393
Epoch 22/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0552 - accuracy: 0.9910 - val_loss: 0.2408 - val_accuracy: 0.9488
Epoch 23/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0517 - accuracy: 0.9925 - val_loss: 0.2810 - val_accuracy: 0.9350
Epoch 24/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0525 - accuracy: 0.9914 - val_loss: 0.2122 - val_accuracy: 0.9459
Epoch 25/50
263/263 [==============================] - 7s 27ms/step - loss: 0.0522 - accuracy: 0.9917 - val_loss: 0.2259 - val_accuracy: 0.9493
Epoch 26/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0522 - accuracy: 0.9919 - val_loss: 0.2751 - val_accuracy: 0.9379
Epoch 27/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0560 - accuracy: 0.9904 - val_loss: 0.2802 - val_accuracy: 0.9422
Epoch 28/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0383 - accuracy: 0.9960 - val_loss: 0.2612 - val_accuracy: 0.9464
Epoch 29/50
263/263 [==============================] - 7s 27ms/step - loss: 0.0558 - accuracy: 0.9905 - val_loss: 0.2870 - val_accuracy: 0.9440
Epoch 30/50
263/263 [==============================] - 7s 27ms/step - loss: 0.0329 - accuracy: 0.9981 - val_loss: 0.2645 - val_accuracy: 0.9559
Epoch 31/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0476 - accuracy: 0.9923 - val_loss: 0.2235 - val_accuracy: 0.9464
Epoch 32/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0371 - accuracy: 0.9961 - val_loss: 0.2934 - val_accuracy: 0.9521
Epoch 33/50
263/263 [==============================] - 7s 27ms/step - loss: 0.0494 - accuracy: 0.9923 - val_loss: 0.2960 - val_accuracy: 0.9393
Epoch 34/50
263/263 [==============================] - 7s 25ms/step - loss: 0.0551 - accuracy: 0.9905 - val_loss: 0.2176 - val_accuracy: 0.9535
Epoch 35/50
263/263 [==============================] - 7s 27ms/step - loss: 0.0355 - accuracy: 0.9958 - val_loss: 0.2783 - val_accuracy: 0.9502
Epoch 36/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0387 - accuracy: 0.9949 - val_loss: 0.2497 - val_accuracy: 0.9493
Epoch 37/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0338 - accuracy: 0.9957 - val_loss: 0.2888 - val_accuracy: 0.9521
Epoch 38/50
263/263 [==============================] - 7s 27ms/step - loss: 0.0592 - accuracy: 0.9891 - val_loss: 0.2465 - val_accuracy: 0.9422
Epoch 39/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0314 - accuracy: 0.9980 - val_loss: 0.2826 - val_accuracy: 0.9478
Epoch 40/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0322 - accuracy: 0.9966 - val_loss: 0.3041 - val_accuracy: 0.9474
Epoch 41/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0365 - accuracy: 0.9954 - val_loss: 0.4735 - val_accuracy: 0.9241
Epoch 42/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0423 - accuracy: 0.9935 - val_loss: 0.2528 - val_accuracy: 0.9578
Epoch 43/50
263/263 [==============================] - 7s 27ms/step - loss: 0.0274 - accuracy: 0.9982 - val_loss: 0.3552 - val_accuracy: 0.9459
Epoch 44/50
263/263 [==============================] - 7s 27ms/step - loss: 0.0300 - accuracy: 0.9973 - val_loss: 0.2836 - val_accuracy: 0.9398
Epoch 45/50
263/263 [==============================] - 7s 25ms/step - loss: 0.0340 - accuracy: 0.9952 - val_loss: 0.2927 - val_accuracy: 0.9497
Epoch 46/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0442 - accuracy: 0.9917 - val_loss: 0.3046 - val_accuracy: 0.9388
Epoch 47/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0264 - accuracy: 0.9986 - val_loss: 0.2410 - val_accuracy: 0.9540
Epoch 48/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0287 - accuracy: 0.9973 - val_loss: 0.3600 - val_accuracy: 0.9374
Epoch 49/50
263/263 [==============================] - 7s 27ms/step - loss: 0.0464 - accuracy: 0.9907 - val_loss: 0.2628 - val_accuracy: 0.9469
Epoch 50/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0275 - accuracy: 0.9976 - val_loss: 0.2887 - val_accuracy: 0.9578

CV 2
Epoch 1/50
263/263 [==============================] - 17s 57ms/step - loss: 0.6296 - accuracy: 0.6513 - val_loss: 0.5415 - val_accuracy: 0.7240
Epoch 2/50
263/263 [==============================] - 7s 26ms/step - loss: 0.3590 - accuracy: 0.8524 - val_loss: 0.3806 - val_accuracy: 0.8350
Epoch 3/50
263/263 [==============================] - 7s 26ms/step - loss: 0.3032 - accuracy: 0.8843 - val_loss: 0.3152 - val_accuracy: 0.8630
Epoch 4/50
263/263 [==============================] - 7s 26ms/step - loss: 0.2561 - accuracy: 0.9057 - val_loss: 0.2714 - val_accuracy: 0.8928
Epoch 5/50
263/263 [==============================] - 7s 27ms/step - loss: 0.2361 - accuracy: 0.9123 - val_loss: 0.2537 - val_accuracy: 0.9090
Epoch 6/50
263/263 [==============================] - 7s 26ms/step - loss: 0.2088 - accuracy: 0.9269 - val_loss: 0.2638 - val_accuracy: 0.8985
Epoch 7/50
263/263 [==============================] - 7s 26ms/step - loss: 0.2099 - accuracy: 0.9269 - val_loss: 0.2278 - val_accuracy: 0.9156
Epoch 8/50
263/263 [==============================] - 7s 26ms/step - loss: 0.1721 - accuracy: 0.9433 - val_loss: 0.2418 - val_accuracy: 0.9042
Epoch 9/50
263/263 [==============================] - 7s 25ms/step - loss: 0.1546 - accuracy: 0.9474 - val_loss: 0.2095 - val_accuracy: 0.9298
Epoch 10/50
263/263 [==============================] - 7s 27ms/step - loss: 0.1497 - accuracy: 0.9527 - val_loss: 0.2097 - val_accuracy: 0.9213
Epoch 11/50
263/263 [==============================] - 7s 26ms/step - loss: 0.1343 - accuracy: 0.9588 - val_loss: 0.2332 - val_accuracy: 0.9075
Epoch 12/50
263/263 [==============================] - 7s 26ms/step - loss: 0.1173 - accuracy: 0.9630 - val_loss: 0.1970 - val_accuracy: 0.9350
Epoch 13/50
263/263 [==============================] - 7s 26ms/step - loss: 0.1042 - accuracy: 0.9686 - val_loss: 0.1894 - val_accuracy: 0.9407
Epoch 14/50
263/263 [==============================] - 7s 26ms/step - loss: 0.1053 - accuracy: 0.9697 - val_loss: 0.2082 - val_accuracy: 0.9350
Epoch 15/50
263/263 [==============================] - 7s 27ms/step - loss: 0.0924 - accuracy: 0.9745 - val_loss: 0.2383 - val_accuracy: 0.9241
Epoch 16/50
263/263 [==============================] - 7s 27ms/step - loss: 0.0796 - accuracy: 0.9806 - val_loss: 0.2311 - val_accuracy: 0.9398
Epoch 17/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0795 - accuracy: 0.9787 - val_loss: 0.1954 - val_accuracy: 0.9478
Epoch 18/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0676 - accuracy: 0.9837 - val_loss: 0.2104 - val_accuracy: 0.9488
Epoch 19/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0645 - accuracy: 0.9849 - val_loss: 0.2294 - val_accuracy: 0.9464
Epoch 20/50
263/263 [==============================] - 7s 28ms/step - loss: 0.0607 - accuracy: 0.9861 - val_loss: 0.2366 - val_accuracy: 0.9398
Epoch 21/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0668 - accuracy: 0.9822 - val_loss: 0.2420 - val_accuracy: 0.9360
Epoch 22/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0617 - accuracy: 0.9869 - val_loss: 0.2115 - val_accuracy: 0.9478
Epoch 23/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0499 - accuracy: 0.9898 - val_loss: 0.3090 - val_accuracy: 0.9284
Epoch 24/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0541 - accuracy: 0.9895 - val_loss: 0.2563 - val_accuracy: 0.9403
Epoch 25/50
263/263 [==============================] - 7s 28ms/step - loss: 0.0453 - accuracy: 0.9913 - val_loss: 0.2587 - val_accuracy: 0.9483
Epoch 26/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0576 - accuracy: 0.9881 - val_loss: 0.2429 - val_accuracy: 0.9488
Epoch 27/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0355 - accuracy: 0.9955 - val_loss: 0.2969 - val_accuracy: 0.9384
Epoch 28/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0564 - accuracy: 0.9875 - val_loss: 0.2725 - val_accuracy: 0.9388
Epoch 29/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0396 - accuracy: 0.9949 - val_loss: 0.2618 - val_accuracy: 0.9526
Epoch 30/50
263/263 [==============================] - 7s 27ms/step - loss: 0.0440 - accuracy: 0.9928 - val_loss: 0.3184 - val_accuracy: 0.9398
Epoch 31/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0753 - accuracy: 0.9816 - val_loss: 0.2277 - val_accuracy: 0.9388
Epoch 32/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0435 - accuracy: 0.9922 - val_loss: 0.2447 - val_accuracy: 0.9502
Epoch 33/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0357 - accuracy: 0.9955 - val_loss: 0.2904 - val_accuracy: 0.9455
Epoch 34/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0654 - accuracy: 0.9850 - val_loss: 0.2455 - val_accuracy: 0.9436
Epoch 35/50
263/263 [==============================] - 7s 27ms/step - loss: 0.0321 - accuracy: 0.9966 - val_loss: 0.2937 - val_accuracy: 0.9516
Epoch 36/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0222 - accuracy: 0.9999 - val_loss: 0.2844 - val_accuracy: 0.9559
Epoch 37/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0193 - accuracy: 1.0000 - val_loss: 0.2873 - val_accuracy: 0.9602
Epoch 38/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0174 - accuracy: 1.0000 - val_loss: 0.2869 - val_accuracy: 0.9602
Epoch 39/50
263/263 [==============================] - 7s 27ms/step - loss: 0.0157 - accuracy: 1.0000 - val_loss: 0.2909 - val_accuracy: 0.9602
Epoch 40/50
263/263 [==============================] - 7s 27ms/step - loss: 0.0141 - accuracy: 1.0000 - val_loss: 0.2880 - val_accuracy: 0.9592
Epoch 41/50
263/263 [==============================] - 7s 25ms/step - loss: 0.0127 - accuracy: 1.0000 - val_loss: 0.2887 - val_accuracy: 0.9616
Epoch 42/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0113 - accuracy: 1.0000 - val_loss: 0.2911 - val_accuracy: 0.9625
Epoch 43/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0100 - accuracy: 1.0000 - val_loss: 0.2906 - val_accuracy: 0.9611
Epoch 44/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0088 - accuracy: 1.0000 - val_loss: 0.2956 - val_accuracy: 0.9545
Epoch 45/50
263/263 [==============================] - 7s 28ms/step - loss: 0.0251 - accuracy: 0.9945 - val_loss: 0.3765 - val_accuracy: 0.8853
Epoch 46/50
263/263 [==============================] - 7s 26ms/step - loss: 0.1159 - accuracy: 0.9606 - val_loss: 0.2241 - val_accuracy: 0.9360
Epoch 47/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0551 - accuracy: 0.9849 - val_loss: 0.2401 - val_accuracy: 0.9426
Epoch 48/50
263/263 [==============================] - 7s 27ms/step - loss: 0.0517 - accuracy: 0.9873 - val_loss: 0.2366 - val_accuracy: 0.9388
Epoch 49/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0354 - accuracy: 0.9933 - val_loss: 0.2496 - val_accuracy: 0.9393
Epoch 50/50
263/263 [==============================] - 7s 28ms/step - loss: 0.0344 - accuracy: 0.9917 - val_loss: 0.2557 - val_accuracy: 0.9379

CV 3
Epoch 1/50
263/263 [==============================] - 18s 60ms/step - loss: 0.7068 - accuracy: 0.5899 - val_loss: 0.6871 - val_accuracy: 0.5965
Epoch 2/50
263/263 [==============================] - 7s 25ms/step - loss: 0.5515 - accuracy: 0.6947 - val_loss: 0.3532 - val_accuracy: 0.8525
Epoch 3/50
263/263 [==============================] - 7s 27ms/step - loss: 0.3284 - accuracy: 0.8645 - val_loss: 0.3120 - val_accuracy: 0.8853
Epoch 4/50
263/263 [==============================] - 7s 26ms/step - loss: 0.2807 - accuracy: 0.8931 - val_loss: 0.3025 - val_accuracy: 0.8862
Epoch 5/50
263/263 [==============================] - 7s 26ms/step - loss: 0.2362 - accuracy: 0.9144 - val_loss: 0.2469 - val_accuracy: 0.9099
Epoch 6/50
263/263 [==============================] - 7s 27ms/step - loss: 0.2126 - accuracy: 0.9247 - val_loss: 0.2345 - val_accuracy: 0.9194
Epoch 7/50
263/263 [==============================] - 7s 26ms/step - loss: 0.1923 - accuracy: 0.9344 - val_loss: 0.2541 - val_accuracy: 0.9203
Epoch 8/50
263/263 [==============================] - 7s 26ms/step - loss: 0.1788 - accuracy: 0.9387 - val_loss: 0.2092 - val_accuracy: 0.9355
Epoch 9/50
263/263 [==============================] - 7s 27ms/step - loss: 0.1509 - accuracy: 0.9491 - val_loss: 0.2047 - val_accuracy: 0.9398
Epoch 10/50
263/263 [==============================] - 7s 26ms/step - loss: 0.1353 - accuracy: 0.9569 - val_loss: 0.2152 - val_accuracy: 0.9369
Epoch 11/50
263/263 [==============================] - 7s 27ms/step - loss: 0.1220 - accuracy: 0.9664 - val_loss: 0.2115 - val_accuracy: 0.9417
Epoch 12/50
263/263 [==============================] - 7s 26ms/step - loss: 0.1036 - accuracy: 0.9709 - val_loss: 0.2067 - val_accuracy: 0.9440
Epoch 13/50
263/263 [==============================] - 7s 26ms/step - loss: 0.1018 - accuracy: 0.9712 - val_loss: 0.2111 - val_accuracy: 0.9369
Epoch 14/50
263/263 [==============================] - 7s 27ms/step - loss: 0.0852 - accuracy: 0.9766 - val_loss: 0.2244 - val_accuracy: 0.9412
Epoch 15/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0773 - accuracy: 0.9812 - val_loss: 0.2498 - val_accuracy: 0.9260
Epoch 16/50
263/263 [==============================] - 7s 25ms/step - loss: 0.0757 - accuracy: 0.9805 - val_loss: 0.2084 - val_accuracy: 0.9440
Epoch 17/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0645 - accuracy: 0.9844 - val_loss: 0.2288 - val_accuracy: 0.9426
Epoch 18/50
263/263 [==============================] - 7s 25ms/step - loss: 0.0628 - accuracy: 0.9840 - val_loss: 0.2407 - val_accuracy: 0.9422
Epoch 19/50
263/263 [==============================] - 7s 28ms/step - loss: 0.0561 - accuracy: 0.9873 - val_loss: 0.2402 - val_accuracy: 0.9488
Epoch 20/50
263/263 [==============================] - 7s 27ms/step - loss: 0.0521 - accuracy: 0.9894 - val_loss: 0.3115 - val_accuracy: 0.9379
Epoch 21/50
263/263 [==============================] - 7s 25ms/step - loss: 0.0628 - accuracy: 0.9847 - val_loss: 0.2724 - val_accuracy: 0.9426
Epoch 22/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0561 - accuracy: 0.9861 - val_loss: 0.2290 - val_accuracy: 0.9512
Epoch 23/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0502 - accuracy: 0.9904 - val_loss: 0.2277 - val_accuracy: 0.9474
Epoch 24/50
263/263 [==============================] - 7s 27ms/step - loss: 0.0501 - accuracy: 0.9898 - val_loss: 0.2613 - val_accuracy: 0.9464
Epoch 25/50
263/263 [==============================] - 7s 27ms/step - loss: 0.0316 - accuracy: 0.9963 - val_loss: 0.2643 - val_accuracy: 0.9488
Epoch 26/50
263/263 [==============================] - 7s 25ms/step - loss: 0.0223 - accuracy: 0.9998 - val_loss: 0.3362 - val_accuracy: 0.9502
Epoch 27/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0540 - accuracy: 0.9872 - val_loss: 0.3159 - val_accuracy: 0.9161
Epoch 28/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0542 - accuracy: 0.9882 - val_loss: 0.2793 - val_accuracy: 0.9431
Epoch 29/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0355 - accuracy: 0.9947 - val_loss: 0.2806 - val_accuracy: 0.9417
Epoch 30/50
263/263 [==============================] - 7s 28ms/step - loss: 0.0498 - accuracy: 0.9887 - val_loss: 0.2897 - val_accuracy: 0.9464
Epoch 31/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0443 - accuracy: 0.9918 - val_loss: 0.2955 - val_accuracy: 0.9341
Epoch 32/50
263/263 [==============================] - 7s 25ms/step - loss: 0.0433 - accuracy: 0.9923 - val_loss: 0.3032 - val_accuracy: 0.9497
Epoch 33/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0397 - accuracy: 0.9932 - val_loss: 0.3064 - val_accuracy: 0.9483
Epoch 34/50
263/263 [==============================] - 7s 25ms/step - loss: 0.0307 - accuracy: 0.9963 - val_loss: 0.3349 - val_accuracy: 0.9436
Epoch 35/50
263/263 [==============================] - 7s 27ms/step - loss: 0.0324 - accuracy: 0.9964 - val_loss: 0.4585 - val_accuracy: 0.9256
Epoch 36/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0497 - accuracy: 0.9900 - val_loss: 0.3056 - val_accuracy: 0.9317
Epoch 37/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0489 - accuracy: 0.9888 - val_loss: 0.2959 - val_accuracy: 0.9469
Epoch 38/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0313 - accuracy: 0.9962 - val_loss: 0.3199 - val_accuracy: 0.9469
Epoch 39/50
263/263 [==============================] - 7s 27ms/step - loss: 0.0278 - accuracy: 0.9973 - val_loss: 0.3277 - val_accuracy: 0.9445
Epoch 40/50
263/263 [==============================] - 7s 28ms/step - loss: 0.0473 - accuracy: 0.9904 - val_loss: 0.2563 - val_accuracy: 0.9412
Epoch 41/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0343 - accuracy: 0.9944 - val_loss: 0.2936 - val_accuracy: 0.9474
Epoch 42/50
263/263 [==============================] - 7s 25ms/step - loss: 0.0344 - accuracy: 0.9941 - val_loss: 0.3198 - val_accuracy: 0.9450
Epoch 43/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0365 - accuracy: 0.9945 - val_loss: 0.2744 - val_accuracy: 0.9426
Epoch 44/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0310 - accuracy: 0.9962 - val_loss: 0.4059 - val_accuracy: 0.9360
Epoch 45/50
263/263 [==============================] - 7s 25ms/step - loss: 0.0439 - accuracy: 0.9913 - val_loss: 0.3263 - val_accuracy: 0.9450
Epoch 46/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0281 - accuracy: 0.9970 - val_loss: 0.3318 - val_accuracy: 0.9483
Epoch 47/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0315 - accuracy: 0.9956 - val_loss: 0.4163 - val_accuracy: 0.9317
Epoch 48/50
263/263 [==============================] - 7s 25ms/step - loss: 0.0377 - accuracy: 0.9932 - val_loss: 0.3119 - val_accuracy: 0.9417
Epoch 49/50
263/263 [==============================] - 7s 28ms/step - loss: 0.0329 - accuracy: 0.9952 - val_loss: 0.3192 - val_accuracy: 0.9455
Epoch 50/50
263/263 [==============================] - 7s 27ms/step - loss: 0.0200 - accuracy: 0.9993 - val_loss: 0.3720 - val_accuracy: 0.9497

CV 4
Epoch 1/50
263/263 [==============================] - 16s 54ms/step - loss: 0.5668 - accuracy: 0.7114 - val_loss: 0.3847 - val_accuracy: 0.8412
Epoch 2/50
263/263 [==============================] - 7s 27ms/step - loss: 0.3313 - accuracy: 0.8683 - val_loss: 0.3414 - val_accuracy: 0.8644
Epoch 3/50
263/263 [==============================] - 7s 25ms/step - loss: 0.2821 - accuracy: 0.8963 - val_loss: 0.2769 - val_accuracy: 0.8933
Epoch 4/50
263/263 [==============================] - 7s 26ms/step - loss: 0.2404 - accuracy: 0.9144 - val_loss: 0.2703 - val_accuracy: 0.9000
Epoch 5/50
263/263 [==============================] - 7s 26ms/step - loss: 0.2224 - accuracy: 0.9198 - val_loss: 0.3158 - val_accuracy: 0.8985
Epoch 6/50
263/263 [==============================] - 7s 26ms/step - loss: 0.2027 - accuracy: 0.9324 - val_loss: 0.2464 - val_accuracy: 0.9161
Epoch 7/50
263/263 [==============================] - 7s 27ms/step - loss: 0.1660 - accuracy: 0.9482 - val_loss: 0.2103 - val_accuracy: 0.9279
Epoch 8/50
263/263 [==============================] - 7s 28ms/step - loss: 0.1506 - accuracy: 0.9544 - val_loss: 0.2314 - val_accuracy: 0.9227
Epoch 9/50
263/263 [==============================] - 7s 26ms/step - loss: 0.1538 - accuracy: 0.9526 - val_loss: 0.1944 - val_accuracy: 0.9426
Epoch 10/50
263/263 [==============================] - 7s 26ms/step - loss: 0.1245 - accuracy: 0.9639 - val_loss: 0.2137 - val_accuracy: 0.9327
Epoch 11/50
263/263 [==============================] - 7s 25ms/step - loss: 0.1091 - accuracy: 0.9697 - val_loss: 0.2203 - val_accuracy: 0.9327
Epoch 12/50
263/263 [==============================] - 7s 26ms/step - loss: 0.1108 - accuracy: 0.9727 - val_loss: 0.2389 - val_accuracy: 0.9322
Epoch 13/50
263/263 [==============================] - 7s 25ms/step - loss: 0.0898 - accuracy: 0.9766 - val_loss: 0.2162 - val_accuracy: 0.9440
Epoch 14/50
263/263 [==============================] - 6s 25ms/step - loss: 0.0893 - accuracy: 0.9790 - val_loss: 0.2326 - val_accuracy: 0.9393
Epoch 15/50
263/263 [==============================] - 7s 25ms/step - loss: 0.0767 - accuracy: 0.9824 - val_loss: 0.2705 - val_accuracy: 0.9279
Epoch 16/50
263/263 [==============================] - 7s 25ms/step - loss: 0.0746 - accuracy: 0.9843 - val_loss: 0.2469 - val_accuracy: 0.9327
Epoch 17/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0844 - accuracy: 0.9803 - val_loss: 0.2399 - val_accuracy: 0.9440
Epoch 18/50
263/263 [==============================] - 7s 27ms/step - loss: 0.0657 - accuracy: 0.9865 - val_loss: 0.2712 - val_accuracy: 0.9303
Epoch 19/50
263/263 [==============================] - 7s 25ms/step - loss: 0.0463 - accuracy: 0.9938 - val_loss: 0.2771 - val_accuracy: 0.9403
Epoch 20/50
263/263 [==============================] - 7s 25ms/step - loss: 0.0694 - accuracy: 0.9842 - val_loss: 0.2531 - val_accuracy: 0.9440
Epoch 21/50
263/263 [==============================] - 7s 25ms/step - loss: 0.0579 - accuracy: 0.9887 - val_loss: 0.2975 - val_accuracy: 0.9331
Epoch 22/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0626 - accuracy: 0.9873 - val_loss: 0.2944 - val_accuracy: 0.9312
Epoch 23/50
263/263 [==============================] - 7s 25ms/step - loss: 0.0560 - accuracy: 0.9899 - val_loss: 0.3258 - val_accuracy: 0.9270
Epoch 24/50
263/263 [==============================] - 7s 25ms/step - loss: 0.0398 - accuracy: 0.9961 - val_loss: 0.3283 - val_accuracy: 0.9426
Epoch 25/50
263/263 [==============================] - 7s 25ms/step - loss: 0.0512 - accuracy: 0.9912 - val_loss: 0.2765 - val_accuracy: 0.9384
Epoch 26/50
263/263 [==============================] - 6s 25ms/step - loss: 0.0443 - accuracy: 0.9936 - val_loss: 0.3364 - val_accuracy: 0.9469
Epoch 27/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0699 - accuracy: 0.9843 - val_loss: 0.3219 - val_accuracy: 0.9336
Epoch 28/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0470 - accuracy: 0.9923 - val_loss: 0.2976 - val_accuracy: 0.9436
Epoch 29/50
263/263 [==============================] - 7s 27ms/step - loss: 0.0506 - accuracy: 0.9907 - val_loss: 0.3426 - val_accuracy: 0.9294
Epoch 30/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0512 - accuracy: 0.9904 - val_loss: 0.3651 - val_accuracy: 0.9322
Epoch 31/50
263/263 [==============================] - 7s 25ms/step - loss: 0.0451 - accuracy: 0.9928 - val_loss: 0.3058 - val_accuracy: 0.9360
Epoch 32/50
263/263 [==============================] - 7s 25ms/step - loss: 0.0482 - accuracy: 0.9928 - val_loss: 0.2869 - val_accuracy: 0.9412
Epoch 33/50
263/263 [==============================] - 7s 25ms/step - loss: 0.0499 - accuracy: 0.9916 - val_loss: 0.3573 - val_accuracy: 0.9298
Epoch 34/50
263/263 [==============================] - 7s 25ms/step - loss: 0.0346 - accuracy: 0.9967 - val_loss: 0.3462 - val_accuracy: 0.9417
Epoch 35/50
263/263 [==============================] - 7s 25ms/step - loss: 0.0316 - accuracy: 0.9971 - val_loss: 0.3925 - val_accuracy: 0.9317
Epoch 36/50
263/263 [==============================] - 7s 25ms/step - loss: 0.0537 - accuracy: 0.9895 - val_loss: 0.2978 - val_accuracy: 0.9440
Epoch 37/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0398 - accuracy: 0.9943 - val_loss: 0.3594 - val_accuracy: 0.9275
Epoch 38/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0467 - accuracy: 0.9920 - val_loss: 0.3128 - val_accuracy: 0.9445
Epoch 39/50
263/263 [==============================] - 7s 27ms/step - loss: 0.0313 - accuracy: 0.9974 - val_loss: 0.4037 - val_accuracy: 0.9398
Epoch 40/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0485 - accuracy: 0.9909 - val_loss: 0.3852 - val_accuracy: 0.9369
Epoch 41/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0391 - accuracy: 0.9950 - val_loss: 0.3830 - val_accuracy: 0.9331
Epoch 42/50
263/263 [==============================] - 7s 25ms/step - loss: 0.0416 - accuracy: 0.9930 - val_loss: 0.3671 - val_accuracy: 0.9374
Epoch 43/50
263/263 [==============================] - 7s 25ms/step - loss: 0.0364 - accuracy: 0.9950 - val_loss: 0.3237 - val_accuracy: 0.9474
Epoch 44/50
263/263 [==============================] - 7s 25ms/step - loss: 0.0264 - accuracy: 0.9985 - val_loss: 0.3693 - val_accuracy: 0.9417
Epoch 45/50
263/263 [==============================] - 7s 25ms/step - loss: 0.0433 - accuracy: 0.9932 - val_loss: 0.3646 - val_accuracy: 0.9279
Epoch 46/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0444 - accuracy: 0.9928 - val_loss: 0.3437 - val_accuracy: 0.9431
Epoch 47/50
263/263 [==============================] - 7s 25ms/step - loss: 0.0250 - accuracy: 0.9986 - val_loss: 0.3882 - val_accuracy: 0.9388
Epoch 48/50
263/263 [==============================] - 7s 25ms/step - loss: 0.0246 - accuracy: 0.9980 - val_loss: 0.3886 - val_accuracy: 0.9426
Epoch 49/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0425 - accuracy: 0.9918 - val_loss: 0.3517 - val_accuracy: 0.9507
Epoch 50/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0303 - accuracy: 0.9969 - val_loss: 0.3057 - val_accuracy: 0.9478

CV 5
Epoch 1/50
263/263 [==============================] - 17s 55ms/step - loss: 0.5526 - accuracy: 0.7123 - val_loss: 0.4276 - val_accuracy: 0.8112
Epoch 2/50
263/263 [==============================] - 7s 25ms/step - loss: 0.3423 - accuracy: 0.8639 - val_loss: 0.3358 - val_accuracy: 0.8657
Epoch 3/50
263/263 [==============================] - 7s 25ms/step - loss: 0.2658 - accuracy: 0.8976 - val_loss: 0.2966 - val_accuracy: 0.8824
Epoch 4/50
263/263 [==============================] - 7s 26ms/step - loss: 0.2423 - accuracy: 0.9117 - val_loss: 0.2690 - val_accuracy: 0.9009
Epoch 5/50
263/263 [==============================] - 7s 25ms/step - loss: 0.2077 - accuracy: 0.9260 - val_loss: 0.2571 - val_accuracy: 0.9023
Epoch 6/50
263/263 [==============================] - 7s 25ms/step - loss: 0.1903 - accuracy: 0.9318 - val_loss: 0.2284 - val_accuracy: 0.9222
Epoch 7/50
263/263 [==============================] - 7s 25ms/step - loss: 0.1657 - accuracy: 0.9445 - val_loss: 0.2258 - val_accuracy: 0.9213
Epoch 8/50
263/263 [==============================] - 7s 26ms/step - loss: 0.1652 - accuracy: 0.9432 - val_loss: 0.2013 - val_accuracy: 0.9355
Epoch 9/50
263/263 [==============================] - 7s 28ms/step - loss: 0.1431 - accuracy: 0.9538 - val_loss: 0.2040 - val_accuracy: 0.9331
Epoch 10/50
263/263 [==============================] - 7s 25ms/step - loss: 0.1299 - accuracy: 0.9596 - val_loss: 0.2155 - val_accuracy: 0.9341
Epoch 11/50
263/263 [==============================] - 7s 25ms/step - loss: 0.1153 - accuracy: 0.9677 - val_loss: 0.2186 - val_accuracy: 0.9317
Epoch 12/50
263/263 [==============================] - 7s 25ms/step - loss: 0.1065 - accuracy: 0.9691 - val_loss: 0.2047 - val_accuracy: 0.9360
Epoch 13/50
263/263 [==============================] - 7s 25ms/step - loss: 0.0938 - accuracy: 0.9743 - val_loss: 0.2146 - val_accuracy: 0.9369
Epoch 14/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0882 - accuracy: 0.9766 - val_loss: 0.2535 - val_accuracy: 0.9341
Epoch 15/50
263/263 [==============================] - 7s 25ms/step - loss: 0.0912 - accuracy: 0.9754 - val_loss: 0.2184 - val_accuracy: 0.9369
Epoch 16/50
263/263 [==============================] - 7s 25ms/step - loss: 0.0677 - accuracy: 0.9835 - val_loss: 0.2354 - val_accuracy: 0.9417
Epoch 17/50
263/263 [==============================] - 7s 25ms/step - loss: 0.0715 - accuracy: 0.9815 - val_loss: 0.2516 - val_accuracy: 0.9464
Epoch 18/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0579 - accuracy: 0.9867 - val_loss: 0.2567 - val_accuracy: 0.9412
Epoch 19/50
263/263 [==============================] - 7s 28ms/step - loss: 0.0695 - accuracy: 0.9823 - val_loss: 0.2783 - val_accuracy: 0.9369
Epoch 20/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0583 - accuracy: 0.9846 - val_loss: 0.3258 - val_accuracy: 0.9231
Epoch 21/50
263/263 [==============================] - 7s 25ms/step - loss: 0.0544 - accuracy: 0.9884 - val_loss: 0.2713 - val_accuracy: 0.9431
Epoch 22/50
263/263 [==============================] - 7s 25ms/step - loss: 0.0559 - accuracy: 0.9876 - val_loss: 0.2624 - val_accuracy: 0.9454
Epoch 23/50
263/263 [==============================] - 7s 25ms/step - loss: 0.0464 - accuracy: 0.9901 - val_loss: 0.2678 - val_accuracy: 0.9417
Epoch 24/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0545 - accuracy: 0.9887 - val_loss: 0.2304 - val_accuracy: 0.9435
Epoch 25/50
263/263 [==============================] - 7s 25ms/step - loss: 0.0389 - accuracy: 0.9932 - val_loss: 0.2915 - val_accuracy: 0.9450
Epoch 26/50
263/263 [==============================] - 7s 25ms/step - loss: 0.0637 - accuracy: 0.9851 - val_loss: 0.2919 - val_accuracy: 0.9255
Epoch 27/50
263/263 [==============================] - 7s 25ms/step - loss: 0.0538 - accuracy: 0.9894 - val_loss: 0.2124 - val_accuracy: 0.9435
Epoch 28/50
263/263 [==============================] - 7s 25ms/step - loss: 0.0286 - accuracy: 0.9975 - val_loss: 0.2864 - val_accuracy: 0.9507
Epoch 29/50
263/263 [==============================] - 7s 27ms/step - loss: 0.0464 - accuracy: 0.9887 - val_loss: 0.3532 - val_accuracy: 0.9398
Epoch 30/50
263/263 [==============================] - 7s 28ms/step - loss: 0.0390 - accuracy: 0.9923 - val_loss: 0.2825 - val_accuracy: 0.9492
Epoch 31/50
263/263 [==============================] - 7s 25ms/step - loss: 0.0379 - accuracy: 0.9943 - val_loss: 0.2991 - val_accuracy: 0.9454
Epoch 32/50
263/263 [==============================] - 7s 25ms/step - loss: 0.0388 - accuracy: 0.9931 - val_loss: 0.3238 - val_accuracy: 0.9407
Epoch 33/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0400 - accuracy: 0.9930 - val_loss: 0.2848 - val_accuracy: 0.9488
Epoch 34/50
263/263 [==============================] - 7s 25ms/step - loss: 0.0354 - accuracy: 0.9947 - val_loss: 0.3607 - val_accuracy: 0.9350
Epoch 35/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0479 - accuracy: 0.9909 - val_loss: 0.2716 - val_accuracy: 0.9417
Epoch 36/50
263/263 [==============================] - 7s 25ms/step - loss: 0.0425 - accuracy: 0.9930 - val_loss: 0.2982 - val_accuracy: 0.9412
Epoch 37/50
263/263 [==============================] - 7s 25ms/step - loss: 0.0332 - accuracy: 0.9957 - val_loss: 0.3064 - val_accuracy: 0.9421
Epoch 38/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0391 - accuracy: 0.9930 - val_loss: 0.3159 - val_accuracy: 0.9488
Epoch 39/50
263/263 [==============================] - 7s 25ms/step - loss: 0.0214 - accuracy: 0.9988 - val_loss: 0.3776 - val_accuracy: 0.9473
Epoch 40/50
263/263 [==============================] - 7s 27ms/step - loss: 0.0334 - accuracy: 0.9939 - val_loss: 0.3528 - val_accuracy: 0.9241
Epoch 41/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0408 - accuracy: 0.9907 - val_loss: 0.3231 - val_accuracy: 0.9440
Epoch 42/50
263/263 [==============================] - 7s 25ms/step - loss: 0.0196 - accuracy: 0.9994 - val_loss: 0.3313 - val_accuracy: 0.9464
Epoch 43/50
263/263 [==============================] - 7s 26ms/step - loss: 0.0296 - accuracy: 0.9961 - val_loss: 0.3367 - val_accuracy: 0.9445
Epoch 44/50
263/263 [==============================] - 7s 25ms/step - loss: 0.0429 - accuracy: 0.9903 - val_loss: 0.3703 - val_accuracy: 0.9398
Epoch 45/50
263/263 [==============================] - 7s 25ms/step - loss: 0.0447 - accuracy: 0.9888 - val_loss: 0.3111 - val_accuracy: 0.9417
Epoch 46/50
263/263 [==============================] - 7s 25ms/step - loss: 0.0230 - accuracy: 0.9976 - val_loss: 0.4020 - val_accuracy: 0.9435
Epoch 47/50
263/263 [==============================] - 7s 25ms/step - loss: 0.0436 - accuracy: 0.9903 - val_loss: 0.3020 - val_accuracy: 0.9393
Epoch 48/50
263/263 [==============================] - 7s 25ms/step - loss: 0.0235 - accuracy: 0.9981 - val_loss: 0.3608 - val_accuracy: 0.9507
Epoch 49/50
263/263 [==============================] - 7s 25ms/step - loss: 0.0374 - accuracy: 0.9929 - val_loss: 0.3193 - val_accuracy: 0.9469
Epoch 50/50
263/263 [==============================] - 7s 25ms/step - loss: 0.0223 - accuracy: 0.9980 - val_loss: 0.3957 - val_accuracy: 0.9388

Evaluasi Model

Melihat kualitas model pada setiap split terhadap data validasi dengan akurasi dan f1-scorenya

$ python
for i in range(split):
    print(f"Split {i + 1} : {acc_scores[i]} acc - {f1_scores[i]} f1")

print("\nMean Acc", sum(acc_scores) / split)
print("Mean F1 ", sum(f1_scores) / split)

Split 1 : 0.9577999051683262 acc - 0.9470553242117787 f1
Split 2 : 0.9625414888572783 acc - 0.9533923303834808 f1
Split 3 : 0.9511616880037933 acc - 0.939161252215003 f1
Split 4 : 0.9506875296348981 acc - 0.938751472320377 f1
Split 5 : 0.9506641366223909 acc - 0.937799043062201 f1

Mean Acc 0.9545709496573374
Mean F1  0.943231884438568

Membuat Submission

Membuat submission

$ python
submission = pd.DataFrame({'id' : [x.split('/')[-1].split('.')[0] for x in new_test_images],
                           'jenis kelamin': np.array((prediksi / split).flatten() >= .5, dtype = 'int')})
test = test.merge(submission, on="id")
test.head()

Export csv

$ python
test.to_csv("submission-gender.csv", index=False)

Age Detection

Pada tahap semifinal kami ditugaskan untuk melakukan pendeteksian terhadap umur seseorang didalam foto (regression). Data yang disediakan merupakan data gambar yang sama pada tahap penyisihan. Pada tahap data preprocessing kami melakukan hal yang cukup sama seperti pada tahap penyisihan namun pada tahap maual filtering sedikit lebih diketatkan. Model yang digunakan oleh tim kami dalam menyelesaikan masalah ini adalah tensorflow custom model dengan arsitektur 6 convolution layer yang dipadukan dengan max pooling layer dan batch normalization, kami menginkan model yang lebih besar dan lebih dalam dari model yang digunakan pada tahap penyisihan. Kemudian, ditutup dengan dense layer berkativasi linear. Sama seperti tahap penyisihan kami melakukan ensembel terhadap hasil prediksi 3 model yang dilatih dengan data yang berbeda (menggunakan StratifiedKfold) untuk meningkatkan kualitas prediksi.

First Thing First

Menginstall library yang diperlukan dan mengimport library - library yang akan digunakan serta menseting variable config yang akan digunakan di dalam notebook ini.

1. Menginstal library MTCNN dan Albumentations

Library MTCNN dan Albumentations adalah library yang digunakan untuk preprocessing data gambar pada notebook ini

$ python
!pip -q install mtcnn
!pip -q install albumentations --upgrade

WARNING: Running pip as the 'root' user can result in broken permissions and conflicting behaviour with the system package manager. It is recommended to use a virtual environment instead: https://pip.pypa.io/warnings/venv
WARNING: Running pip as the 'root' user can result in broken permissions and conflicting behaviour with the system package manager. It is recommended to use a virtual environment instead: https://pip.pypa.io/warnings/venv

2. Importing library

Mengimport library yang akan digunakan dalam notebook ini.

$ python
# Umum
import os, random, re
from tqdm.notebook import tqdm
import numpy as np
import pandas as pd
from PIL import Image

# Tensorflow
import tensorflow as tf
from tensorflow.keras.preprocessing.image import load_img, img_to_array

# Metrics & Splitting data
from sklearn.metrics import *
from sklearn.model_selection import *

# Plotting
import matplotlib.pyplot as plt
import seaborn as sns

# Preprocessing
import cv2
from mtcnn import MTCNN
import albumentations as A

print("Tensorflow :", tf.__version__)

Tensorflow : 2.6.0

3. Setup CONFIG

Mensetup varible - variable yang digunakan sebagai config pada notebook ini

$ python
SEED = 2021
SIZE = (200, 200)
BATCH_SIZE = 32
FACE_THRESHOLD = 0.95
FACE_DETECTOR = MTCNN()

2022-03-25 05:56:07.122802: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2022-03-25 05:56:07.216440: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2022-03-25 05:56:07.217282: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2022-03-25 05:56:07.218923: I tensorflow/core/platform/cpu_feature_guard.cc:142] This TensorFlow binary is optimized with oneAPI Deep Neural Network Library (oneDNN) to use the following CPU instructions in performance-critical operations:  AVX2 AVX512F FMA
To enable them in other operations, rebuild TensorFlow with the appropriate compiler flags.
2022-03-25 05:56:07.219966: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2022-03-25 05:56:07.220994: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2022-03-25 05:56:07.221947: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2022-03-25 05:56:08.885974: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2022-03-25 05:56:08.886927: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2022-03-25 05:56:08.887719: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2022-03-25 05:56:08.888401: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1510] Created device /job:localhost/replica:0/task:0/device:GPU:0 with 15403 MB memory:  -> device: 0, name: Tesla P100-PCIE-16GB, pci bus id: 0000:00:04.0, compute capability: 6.0

Dataset

Menggunakan dataset yang telah digunakan oleh tim kami dan melakukan loading agar mendapatkan informasi path dari data gambar

Dataset : Data SD20210000722

$ python
train = pd.read_csv("../input/bdc-2021/train.csv")
test = pd.read_csv("../input/bdc-2021/submission.csv")
train.head()

memperjelas path ke setiap data gambar

$ python
images = []
labels = []
ages = []
test_images = []

TRAIN_DIR = "../input/bdc-2021/Training"
TEST_DIR = "../input/bdc-2021/Testing"

for no, label, usia in train[["nomor", "jenis kelamin", "usia"]].values:
    TEMP_DIR = os.path.join(TRAIN_DIR, str(no))
    for file in os.listdir(TEMP_DIR):
        file_dir = os.path.join(TEMP_DIR, file)
        if ".ini" not in file_dir:
            images.append(file_dir)
            labels.append(label)
            ages.append(usia)

for no in test.id.values:
    file_dir = os.path.join(TEST_DIR, f"{no}.jpg")
    if os.path.isfile(file_dir):
        test_images.append(file_dir)
    else:
        test_images.append(None)
        print(file_dir)

menampilkan dan mengecek beberapa gambar pada data train

$ python
def read(path):
    """
    Read data gambar
    """
    img = Image.open(path)
    return img

def show_images(list_dir, label = None, age = None, load_image = read, seed = SEED):
    """
    Menampilkan Gambar Secara acak sebanyak 5 buah.
    """
    random.seed(seed)
    unique = ["init"]
    if label:
        unique = list(set(label))
    fig, axes = plt.subplots(len(unique), 5, figsize = (20, 5 * len(unique)))
    for i in range(len(unique)):
        if i == 0 and unique[i] == "init":
            data = random.sample(list_dir, 5)
        elif age != None:
            data = random.sample([x for x in zip(list_dir, label, age) if x[1] == unique[i]], 5)
        else:
            data = random.sample([x for x in zip(list_dir, label) if x[1] == unique[i]], 5)
        for j in range(5):
            if unique[0] != "init":
                img = load_image(data[j][0])
                text = f'Label : {data[j][1]}'
                if age != None:
                    text += f'\nAge : {data[j][2]}'
                axes[i, j].imshow(img)
                axes[i, j].set_title(text, fontsize = 14)
                axes[i, j].axis('off')
            else:
                img = load_image(data[j])
                axes[j].imshow(img)
                axes[j].axis('off')
    fig.tight_layout()
    plt.show()

$ python
show_images(images, labels, ages, seed=20)

Preprocess Data

Metode yang digunakan:

1. Mengekstrak wajah - wajah yang terdapat pada gambar menjadi gambar - gambar baru dengan label yang sama dengan menggunakan model MTCNN
2. Pada data test jika terdapat dua wajah yang terdeteksi pada satu gambar akan di ambil wajah dengan tingkat confidence terbesar yang diberikan oleh model MTCNN.
3. Jika tidak terdetect wajah pada salah satu gambar maka akan dilakukan crop pada bagian tengah gambar sehingga gambar berbentuk persegi atau jxj pixel.
4. Selanjutnya gambar akan di resize menjadi ukuran 256x256 pixel

berikut adalah contoh hasil preprocess data gambar.

$ python
def get_faces(path):
    image = cv2.cvtColor(cv2.imread(path), cv2.COLOR_BGR2RGB)
    faces = FACE_DETECTOR.detect_faces(image)
    return faces

def load_and_preprocess_image(path: str, size = SIZE):
    """
    Load & Preprocess data gambar
    """
    image = img_to_array(load_img(path))
    faces = [x['box'] for x in get_faces(path) if x['confidence'] > FACE_THRESHOLD]
    if len(faces) > 0:
        x, y, w, h = faces[0]
        image = image[y:y+h, x:x+w]
    img = tf.convert_to_tensor(image, dtype=tf.float32)
    if len(faces) == 0:
        shapes = tf.shape(img)
        h, w = shapes[-3], shapes[-2]
        dim = tf.minimum(h, w)
        img = tf.image.resize_with_crop_or_pad(img, dim, dim)
    img = tf.image.resize(img, size)
    img = tf.cast(img, tf.float32) / 255.0
    return img.numpy()

$ python
show_images(images, labels, ages, load_image = load_and_preprocess_image, seed=20)

2022-03-25 05:56:16.153559: I tensorflow/compiler/mlir/mlir_graph_optimization_pass.cc:185] None of the MLIR Optimization Passes are enabled (registered 2)
2022-03-25 05:56:17.086578: I tensorflow/stream_executor/cuda/cuda_dnn.cc:369] Loaded cuDNN version 8005

Augmentasi Data

Melakukan augmentasi untuk memperbanyak data. Metode augmentasi yang digunakan yaitu:

1. Horizontal flip
2. Donwscale kualitas gambar
3. Random rotate dengan rentang -30 sampai 30 derajad
4. Shift, scale, dan rotate gambar
5. Blur
6. Random brightness

$ python
aug = A.Compose([
    A.HorizontalFlip(p=0.4),
    A.Downscale(scale_min=0.6, scale_max=0.9, p=0.3),
    A.Rotate(limit=(-30,30), p=0.6),
    A.ShiftScaleRotate(shift_limit=(-0.07, 0.07), scale_limit=(-0.05, 0.1), rotate_limit=(-15, 15), p=0.4),
    A.OneOf([
        A.MotionBlur(p=.4),
        A.MedianBlur(blur_limit=3, p=0.4),
        A.Blur(blur_limit=3, p=0.4),
    ], p=0.4),
    A.RandomBrightnessContrast(brightness_limit=(-0.25, 0.15), p=0.4),
])

$ python
def visualize_aug(path):
    fig, axes = plt.subplots(1, 5, figsize = (20, 5))
    image = load_and_preprocess_image(path)
    axes[0].imshow(image)
    axes[0].axis('off')
    for i in range(1, 5):
        augmented = aug(image=image)['image']
        axes[i].imshow(augmented)
        axes[i].axis('off')
    fig.tight_layout()
    plt.show()

$ python
random.seed(SEED)
for i in range(3):
    visualize_aug(images[i])

Running preprocessing pada data gambar secara keseluruhan

$ python
def image_preprocessing(new_dir, images, labels=None):
    if os.path.isdir(new_dir):
        !rm -rf {new_dir}
    os.mkdir(new_dir)

    new_images, new_labels = [], []
    if not labels:
        labels = [None for _ in range(len(images))]

    for path, label in tqdm(zip(images, labels), total=len(images)):
        image = img_to_array(load_img(path))
        if label != None:
            faces = [x['box'] for x in sorted(get_faces(path), key=lambda x: x['confidence'],
                                              reverse=True) if x['confidence'] > FACE_THRESHOLD]
        else:
            faces = [x['box'] for x in sorted(get_faces(path), key=lambda x: x['confidence'], reverse=True)]
        if len(faces) > 0:
            if label != None:
                for j, (x, y, w, h) in enumerate(faces):
                    img = image[y:y+h, x:x+w]
                    img = tf.convert_to_tensor(img, dtype=tf.float32)
                    img = tf.image.resize(img, SIZE)
                    img = tf.cast(img, tf.float32) / 255.0

                    img_dir = os.path.join(new_dir, f'{j}_{path.split("/")[-1]}')
                    new_images.append(img_dir)
                    new_labels.append(label)
                    tf.keras.preprocessing.image.save_img(img_dir, img)

                    for k in range(3):
                        augmented = aug(image=img.numpy())['image']
                        img_dir = os.path.join(new_dir, f'aug-{k}_{j}_{path.split("/")[-1]}')
                        new_images.append(img_dir)
                        new_labels.append(label)
                        tf.keras.preprocessing.image.save_img(img_dir, augmented)
            else:
                x, y, w, h = faces[0]
                img = image[y:y+h, x:x+w]
                img = tf.convert_to_tensor(img, dtype=tf.float32)
                img = tf.image.resize(img, SIZE)
                img = tf.cast(img, tf.float32) / 255.0

                img_dir = os.path.join(new_dir, path.split('/')[-1])
                new_images.append(img_dir)
                new_labels.append(label)
                tf.keras.preprocessing.image.save_img(img_dir, img)
        else :
            img = tf.convert_to_tensor(image, dtype=tf.float32)
            shapes = tf.shape(img)
            h, w = shapes[-3], shapes[-2]
            dim = tf.minimum(h, w)
            img = tf.image.resize_with_crop_or_pad(img, dim, dim)
            img = tf.image.resize(img, SIZE)
            img = tf.cast(img, tf.float32) / 255.0

            img_dir = os.path.join(new_dir, path.split('/')[-1])
            new_images.append(img_dir)
            new_labels.append(label)
            tf.keras.preprocessing.image.save_img(img_dir, img)

            if label != None:
                for k in range(3):
                    augmented = aug(image=img.numpy())['image']
                    img_dir = os.path.join(new_dir,  f'aug-{k}_{path.split("/")[-1]}')
                    new_images.append(img_dir)
                    new_labels.append(label)
                    tf.keras.preprocessing.image.save_img(img_dir, augmented)

    return new_images, new_labels

Untuk menghemat waktu running akan di skip bagian ini dan di ganti dengan meload data hasil preprocess yang sudah di save pada run sebelumnya. Namun jika ingin melakukan preprocess pada run sekarang maka uncomment code di bawah ini.

Peringatan : running block code di bawah memakan waktu sekitar 50 menit dengan GPU Nvidia Tesla P100-PCIE.

$ python
# new_train_dir = "./train"
# new_test_dir = "./test"

# random.seed(SEED)
# new_images, new_labels = image_preprocessing(new_train_dir, images, labels)
# new_test_images, _ = image_preprocessing(new_test_dir, test_images)

Note : Comment dua block kode di bawah jika melakukan preprocess pada run saat ini.

$ python
preprocessed = pd.read_csv("../input/bdc-2021/preprocessed-augmented/preprocessed.csv")
preprocessed.head()