概要

Keras学習済みモデルのXceptionをUTKFaceデータセットでFine-tuningさせ、人種分類モデルを構築する
白人、黒人、アジア、インド、その他（ヒスパニック、ラテン、中東など）の5つに分類

UTKFace
https://susanqq.github.io/UTKFace/

20,000以上の顔画像データセット
性別、年齢、人種のラベリング

実行環境
Google Colaboratory（GPU）

import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
%config InlineBackend.figure_formats = {'png', 'retina'}

import os, zipfile, io, re
from PIL import Image
from sklearn.model_selection import train_test_split
from keras.applications.xception import Xception
from keras.models import Model, load_model
from keras.layers.core import Dense
from keras.layers.pooling import GlobalAveragePooling2D
from keras.optimizers import Adam, RMSprop, SGD
from keras.utils.np_utils import to_categorical
from keras.callbacks import ModelCheckpoint, EarlyStopping, TensorBoard, ReduceLROnPlateau
from keras.preprocessing.image import ImageDataGenerator

データ取得

画像は100にリサイズ
ZIPファイルからデータセットを取得し、配列に変換
クラスはWhite、Black、Asian、Indian、Othersの5クラス（各画像ファイル名にラベリングされている）
取得したデータセットはtrain、valid、testに分割

classes = ["White", "Black", "Asian", "Indian", "Others"]
num_classes = len(classes)
image_size = 100

%%time
# ZIP読み込み
z = zipfile.ZipFile('../dataset/UTKFace.zip')
# 画像ファイルパスのみ取得
imgfiles = [ x for x in z.namelist() if re.search(r"^UTKFace.*jpg$", x)]

X = []
Y = []
for imgfile in imgfiles:
    # ZIPから画像読み込み
    image = Image.open(io.BytesIO(z.read(imgfile)))
    # RGB変換
    image = image.convert('RGB')
    # リサイズ
    image = image.resize((image_size, image_size))
    # 画像から配列に変換
    data = np.asarray(image)
    file = os.path.basename(imgfile)
    file_split = [i for i in file.split('_')]
    X.append(data)
    Y.append(file_split[2])
z.close()
del z, imgfiles

X = np.array(X)
Y = np.array(Y)
print(X.shape, Y.shape)

(23708, 100, 100, 3) (23708,)
CPU times: user 23.9 s, sys: 387 ms, total: 24.3 s
Wall time: 24.3 s

人種がラベリングされていない3つのデータを削除

# df_Y=pd.DataFrame(Y)
# pd.DataFrame(df_Y[0].value_counts()).sort_index().rename(columns={0 :'num'}).T

index = np.where(
    (Y == "20170109142408075.jpg.chip.jpg") | 
    (Y == "20170109150557335.jpg.chip.jpg") | 
    (Y == "20170116174525125.jpg.chip.jpg")
)
X=np.delete(X, index, axis=0)
Y=np.delete(Y, index, axis=0)
print(X.shape, Y.shape)

(23705, 100, 100, 3) (23705,)

# trainデータとtestデータに分割
X_train, X_test, y_train, y_test = train_test_split(
    X, Y,
    random_state = 0,
    stratify = Y,
    test_size = 0.2
)
del X,Y
print(X_train.shape, y_train.shape, X_test.shape, y_test.shape)

(18964, 100, 100, 3) (18964,) (4741, 100, 100, 3) (4741,)

# データ型の変換＆正規化
X_train = X_train.astype('float32') / 255
X_test = X_test.astype('float32') / 255
# one-hot変換
y_train = to_categorical(y_train, num_classes = num_classes)
y_test = to_categorical(y_test, num_classes = num_classes)

# trainデータからvalidデータを分割
X_train, X_valid, y_train, y_valid = train_test_split(
    X_train,
    y_train,
    random_state = 0,
    stratify = y_train,
    test_size = 0.2
)
print(X_train.shape, y_train.shape, X_valid.shape, y_valid.shape)

(15171, 100, 100, 3) (15171, 5) (3793, 100, 100, 3) (3793, 5)

モデル構築

Xception読み込み
Keras学習済みモデルXceptionを読み込む
その際、ネットワーク出力層側にある全結合層を除去

base_model = Xception(
    include_top = False,
    weights = "imagenet",
    input_shape = None
)

全結合層の新規構築

x = base_model.output
x = GlobalAveragePooling2D()(x)
x = Dense(1024, activation = 'relu')(x)
predictions = Dense(num_classes, activation = 'softmax')(x)

Data Augmentation

datagen = ImageDataGenerator(
    featurewise_center = False,
    samplewise_center = False,
    featurewise_std_normalization = False,
    samplewise_std_normalization = False,
    zca_whitening = False,
    rotation_range = 0,
    width_shift_range = 0.1,
    height_shift_range = 0.1,
    horizontal_flip = True,
    vertical_flip = False
)

Callback

# EarlyStopping
early_stopping = EarlyStopping(
    monitor = 'val_loss',
    patience = 10,
    verbose = 1
)

# ModelCheckpoint
weights_dir = './weights/'
if os.path.exists(weights_dir) == False:os.mkdir(weights_dir)
model_checkpoint = ModelCheckpoint(
    weights_dir + "val_loss{val_loss:.3f}.hdf5",
    monitor = 'val_loss',
    verbose = 1,
    save_best_only = True,
    save_weights_only = True,
    period = 3
)

# reduce learning rate
reduce_lr = ReduceLROnPlateau(
    monitor = 'val_loss',
    factor = 0.1,
    patience = 3,
    verbose = 1
)

# log for TensorBoard
logging = TensorBoard(log_dir = "log/")

モデル学習

XceptionをFine-tuning

# ネットワーク定義
model = Model(inputs = base_model.input, outputs = predictions)

#108層までfreeze
for layer in model.layers[:108]:
    layer.trainable = False

    # Batch Normalizationのfreeze解除
    if layer.name.startswith('batch_normalization'):
        layer.trainable = True
    if layer.name.endswith('bn'):
        layer.trainable = True

#109層以降、学習させる
for layer in model.layers[108:]:
    layer.trainable = True
    
# layer.trainableの設定後にcompile
model.compile(
    optimizer = Adam(),
    loss = 'categorical_crossentropy',
    metrics = ["accuracy"]
)

%%time
hist = model.fit_generator(
    datagen.flow(X_train, y_train, batch_size = 32),
    steps_per_epoch = X_train.shape[0] // 32,
    epochs = 50,
    validation_data = (X_valid, y_valid),
    callbacks = [early_stopping, reduce_lr],
    shuffle = True,
    verbose = 1
)

Epoch 1/50
474/474 [==============================] - 121s 256ms/step - loss: 1.3566 - acc: 0.4722 - val_loss: 1.4610 - val_acc: 0.6011
Epoch 2/50
474/474 [==============================] - 112s 237ms/step - loss: 0.8982 - acc: 0.6872 - val_loss: 0.8991 - val_acc: 0.7037
Epoch 3/50
474/474 [==============================] - 112s 236ms/step - loss: 0.7573 - acc: 0.7388 - val_loss: 0.7338 - val_acc: 0.7611

〜省略〜

Epoch 00018: ReduceLROnPlateau reducing learning rate to 1.0000000656873453e-06.
Epoch 19/50
474/474 [==============================] - 111s 235ms/step - loss: 0.2864 - acc: 0.9008 - val_loss: 0.6332 - val_acc: 0.8070
Epoch 20/50
474/474 [==============================] - 111s 235ms/step - loss: 0.2869 - acc: 0.9000 - val_loss: 0.6365 - val_acc: 0.8089
Epoch 21/50
474/474 [==============================] - 111s 235ms/step - loss: 0.2845 - acc: 0.9029 - val_loss: 0.6342 - val_acc: 0.8081

Epoch 00021: ReduceLROnPlateau reducing learning rate to 1.0000001111620805e-07.
Epoch 22/50
474/474 [==============================] - 111s 235ms/step - loss: 0.2831 - acc: 0.9007 - val_loss: 0.6333 - val_acc: 0.8073
Epoch 00022: early stopping
CPU times: user 52min 7s, sys: 9min 12s, total: 1h 1min 19s
Wall time: 41min 7s

学習曲線をプロット

plt.figure(figsize = (18,6))

# accuracy
plt.subplot(1, 2, 1)
plt.plot(hist.history["acc"], label = "acc", marker = "o")
plt.plot(hist.history["val_acc"], label = "val_acc", marker = "o")
#plt.xticks(np.arange())
#plt.yticks(np.arange())
plt.xlabel("epoch")
plt.ylabel("accuracy")
#plt.title("")
plt.legend(loc = "best")
plt.grid(color = 'gray', alpha = 0.2)

# loss
plt.subplot(1, 2, 2)
plt.plot(hist.history["loss"], label = "loss", marker = "o")
plt.plot(hist.history["val_loss"], label = "val_loss", marker = "o")
#plt.xticks(np.arange())
#plt.yticks(np.arange())
plt.xlabel("epoch")
plt.ylabel("loss")
#plt.title("")
plt.legend(loc = "best")
plt.grid(color = 'gray', alpha = 0.2)

plt.show()

モデル評価

score = model.evaluate(X_test, y_test, verbose = 1)
print("evaluate loss: {[0]:.4f}".format(score))
print("evaluate acc: {[1]:.1%}".format(score))

4741/4741 [==============================] - 13s 3ms/step
evaluate loss: 0.6409054181666798
evaluate acc: 0.8089010757475658

モデル保存

model_dir = './model/'
if os.path.exists(model_dir) == False:os.mkdir(model_dir)

model.save(model_dir + 'model.hdf5')

# optimizerのない軽量モデルを保存（学習や評価は不可だが、予測は可能）
model.save(model_dir + 'model-opt.hdf5', include_optimizer = False)

モデル予測

testデータ30件の画像と正解ラベルを出力

# testデータ30件の正解ラベル
true_classes = np.argmax(y_test[0:30], axis = 1)

# testデータ30件の画像と正解ラベルを出力
plt.figure(figsize = (16, 6))
for i in range(30):
    plt.subplot(3, 10, i + 1)
    plt.axis("off")
    plt.title(classes[true_classes[i]])
    plt.imshow(X_test[i])
plt.show()

testデータ30件の画像と予測ラベル・予測確率を出力

# testデータ30件の予測ラベル
pred_classes = np.argmax(model.predict(X_test[0:30]), axis = 1)

# testデータ30件の予測確率
pred_probs = np.max(model.predict(X_test[0:30]), axis = 1)
pred_probs = ['{:.4f}'.format(i) for i in pred_probs]

# testデータ30件の画像と予測ラベル・予測確率を出力
plt.figure(figsize = (16, 6))
for i in range(30):
    plt.subplot(3, 10, i + 1)
    plt.axis("off")
    if pred_classes[i] == true_classes[i]:
        plt.title(classes[pred_classes[i]]+'\n'+pred_probs[i])
    else:
        plt.title(classes[pred_classes[i]]+'\n'+pred_probs[i], color = "red")
    plt.imshow(X_test[i])
plt.show()

モデル評価でaccuracy 80.9%、loss 0.6409を計測
決して高い精度とは言えず、いくつか誤りが見られる

予測検証

testデータ1000件を抽出して、低い確率で正解した画像と高い確率で間違えた画像を目視で確認

testデータ1000件抽出

# testデータ1000件の正解ラベル
true_classes = np.argmax(y_test[0:1000], axis = 1)

# testデータ1000件の予測ラベル
pred_classes = np.argmax(model.predict(X_test[0:1000]), axis = 1)

# testデータ1000件の予測確率
pred_probs = np.max(model.predict(X_test[0:1000]), axis = 1)
pred_probs = np.round(pred_probs, 2)
pred_probs = ['{:.4f}'.format(i) for i in pred_probs]

正解・不正解リスト作成
（画像データ, 予測ラベル, 予測確率）の正解・不正解それぞれのリストを作成

correct=[]
incorrect=[]

for i in range(1000):
    if pred_classes[i] == true_classes[i]:
        correct.append((X_test[i], classes[pred_classes[i]], pred_probs[i]))
    else:
        incorrect.append((X_test[i], classes[pred_classes[i]], pred_probs[i]))
        
print("number of correct:",len(correct))
print("number of incorrect:",len(incorrect))

number of correct : 803
number of incorrect : 197

低確率の正解画像
確率の低い順に正解画像を出力

# 正解画像を確率の低い順に並び替え
correct.sort(key = lambda x:x[2])

# ワースト30件の画像と予測ラベル・予測確率を出力
plt.figure(figsize = (16, 6))
for i in range(30):
    plt.subplot(3, 10, i + 1)
    plt.axis("off")
    plt.title(correct[i][1]+'\n'+correct[i][2])
    plt.imshow(correct[i][0])
plt.show()

高確率の不正解画像
確率の高い順に不正解画像を出力

# 不正解画像を確率の高い順に並び替え
incorrect.sort(key = lambda x:x[2], reverse = True)

# ベスト30件の画像と予測ラベル・予測確率を出力
plt.figure(figsize = (16, 6))
for i in range(30):
    plt.subplot(3, 10, i + 1)
    plt.axis("off")
    plt.title(incorrect[i][1]+'\n'+incorrect[i][2], color = 'red')
    plt.imshow(incorrect[i][0])
plt.show()

今回は目視での検証だが、今後は正誤の人種比や年齢・性別による正誤の違いを検証したい。