［Keras］CIFAR10でAutoEncoderを実装してみる

こんにちは。Toruです。

最近、MNISTというデータセットを使ってAutoEncoderを実装しているのですが、

他の人もたくさんやっているっていうことと、いまいち画像データが白黒ということもあり面白く無くなってしまいました。

なので今回はCIFAR10というデータセットとKerasというディープラーニングライブラリを使ってAutoEncoderを実装しました。

まあまあいい感じの結果が出たので、これを紹介していきたいと思います。

なるべく初心者の方にもわかりやすく説明していきますので、皆さんの参考にしていただけたら幸いです。

CAFAR10についての紹介

50000の学習データと10000の評価データからなる32×32×3のデータセット
中身は飛行機、車、鳥などの画像があり、それぞれを10種類のクラスに分かれている。
Kerasのデータセットに付属しているためmnistと同様、簡単に呼び出すことができる。

それではいきましょう

【結論】実装したAutoEncoderのプログラム
- その前に、私の開発環境
- 作成したAuto Encoderプログラム
プログラムの詳細
- importライブラリとdefine関数
- main()関数内
おまけ
まとめ

【結論】実装したAutoEncoderのプログラム

その前に、私の開発環境

AutoEncoderを実装していくにあたって、まずは私の開発環境について紹介しておきます。

今回はこの様な開発環境で進めていきます。

私の開発環境

CPU：Intel Core i7 9700
GPU：NVIDIA GeForce RTX 2070 Super
OS：Windows 10 Home
Python：3.6.1
TensorFlow-GPU：1.12.0
Keras：2.2.4

GPU環境の整え方

TensorFlow-GPUの環境の整え方については「CUDA cuDNN TensorFlow GPU版をWindowsにインストールした話」で紹介してますので、こちらを参照してください。

作成したAuto Encoderプログラム

早速結果となりますが、作成したプログラムはこちらになります。

import numpy as np
import tensorflow as tf
from keras import backend as K
from keras import callbacks, layers, models
from keras.datasets import cifar10
from keras.layers import (BatchNormalization, Conv2D, Dense, LeakyReLU,
                          MaxPooling2D, UpSampling2D)
from keras.layers.core import Activation, Dense
from keras.optimizers import Adam
from sklearn.model_selection import train_test_split

def normalizer(x, axis = None):
        min = x.min(axis = axis, keepdims = True)
        max = x.max(axis = axis, keepdims = True)
        result = (x - min) / (max - min)

        return result

def main():

        (x_train, y_train), (x_test, y_test) = cifar10.load_data()

        image_height, image_width = 32, 32
        epochs = 150
        learn_rate = 0.0002
        batch_size = 8

        train_data = np.concatenate([x_train, x_test])
        label_data = np.concatenate([y_train, y_test])

        x_train, x_test, y_train, y_test = train_test_split(train_data, label_data, \
                                                      test_size = 0.2, shuffle = True)
        model = models.Sequential()

        model.add(Conv2D(64, kernel_size = 6, padding = 'same', input_shape = (32, 32, 3),\
                           kernel_initializer = 'random_uniform', bias_initializer = 'zeros'))
        model.add(Activation(LeakyReLU(alpha = 0.5623815057565213)))
        model.add(BatchNormalization())
        model.add(MaxPooling2D(pool_size = (2, 2)))

        model.add(Conv2D(128, kernel_size = 6, padding = 'same'))
        model.add(Activation(LeakyReLU(alpha = 0.3949936266626689)))
        model.add(BatchNormalization())
        model.add(MaxPooling2D(pool_size = (2, 2)))

        model.add(Conv2D(64, kernel_size = 6, padding = "same"))
        model.add(Activation(LeakyReLU(alpha = 0.9758185183456943)))
        model.add(BatchNormalization())
        model.add(UpSampling2D(size = (2, 2)))

        model.add(Conv2D(32, kernel_size = 6, padding = "same"))
        model.add(Activation(LeakyReLU(alpha = 0.288662535902546)))
        model.add(BatchNormalization())
        model.add(Conv2D(16, kernel_size = 6, padding = "same"))
        model.add(Activation(LeakyReLU(alpha = 0.3376263809413643)))
        model.add(BatchNormalization())
        model.add(UpSampling2D(size = (2, 2)))

        model.add(Conv2D(3, kernel_size = 6, padding = "same", activation = "sigmoid"))
        
        model.compile(optimizer = Adam(lr = learn_rate), loss = 'mse', metrics = ['accuracy'])

        model.summary()

        fit_callbacs = [callbacks.EarlyStopping(monitor = 'val_loss', patience = 5, \
                                                                                mode = 'min')]

        hist = model.fit(x_train, x_train, epochs = epochs, batch_size = batch_size, \
                                           shuffle = True, validation_data = (x_test, x_test), \
                                                         callbacks = fit_callbacs, verbose = 1)

        val_loss, val_acc = model.evaluate(x_test, x_test, verbose = 1)

if __name__ == '__main__':
        main()

import numpy as np

import tensorflow as tf

from keras import backend as K

from keras import callbacks, layers, models

from keras.datasets import cifar10

from keras.layers import (BatchNormalization, Conv2D, Dense, LeakyReLU,

MaxPooling2D, UpSampling2D)

from keras.layers.core import Activation, Dense

from keras.optimizers import Adam

from sklearn.model_selection import train_test_split

def normalizer(x, axis = None):

min = x.min(axis = axis, keepdims = True)

max = x.max(axis = axis, keepdims = True)

result = (x - min) / (max - min)

return result

def main():

(x_train, y_train), (x_test, y_test) = cifar10.load_data()

image_height, image_width = 32, 32

epochs = 150

learn_rate = 0.0002

batch_size = 8

train_data = np.concatenate([x_train, x_test])

label_data = np.concatenate([y_train, y_test])

x_train, x_test, y_train, y_test = train_test_split(train_data, label_data, \

test_size = 0.2, shuffle = True)

model = models.Sequential()

model.add(Conv2D(64, kernel_size = 6, padding = 'same', input_shape = (32, 32, 3),\

kernel_initializer = 'random_uniform', bias_initializer = 'zeros'))

model.add(Activation(LeakyReLU(alpha = 0.5623815057565213)))

model.add(BatchNormalization())

model.add(MaxPooling2D(pool_size = (2, 2)))

model.add(Conv2D(128, kernel_size = 6, padding = 'same'))

model.add(Activation(LeakyReLU(alpha = 0.3949936266626689)))

model.add(BatchNormalization())

model.add(MaxPooling2D(pool_size = (2, 2)))

model.add(Conv2D(64, kernel_size = 6, padding = "same"))

model.add(Activation(LeakyReLU(alpha = 0.9758185183456943)))

model.add(BatchNormalization())

model.add(UpSampling2D(size = (2, 2)))

model.add(Conv2D(32, kernel_size = 6, padding = "same"))

model.add(Activation(LeakyReLU(alpha = 0.288662535902546)))

model.add(BatchNormalization())

model.add(Conv2D(16, kernel_size = 6, padding = "same"))

model.add(Activation(LeakyReLU(alpha = 0.3376263809413643)))

model.add(BatchNormalization())

model.add(UpSampling2D(size = (2, 2)))

model.add(Conv2D(3, kernel_size = 6, padding = "same", activation = "sigmoid"))

model.compile(optimizer = Adam(lr = learn_rate), loss = 'mse', metrics = ['accuracy'])

model.summary()

fit_callbacs = [callbacks.EarlyStopping(monitor = 'val_loss', patience = 5, \

mode = 'min')]

hist = model.fit(x_train, x_train, epochs = epochs, batch_size = batch_size, \

shuffle = True, validation_data = (x_test, x_test), \

callbacks = fit_callbacs, verbose = 1)

val_loss, val_acc = model.evaluate(x_test, x_test, verbose = 1)

if __name__ == '__main__':

main()

そして上記プログラムを実行すると、結果はこうなりました。

Using TensorFlow backend.
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv2d_1 (Conv2D)            (None, 32, 32, 64)        6976
_________________________________________________________________
activation_1 (Activation)    (None, 32, 32, 64)        0
_________________________________________________________________
batch_normalization_1 (Batch (None, 32, 32, 64)        256
_________________________________________________________________
max_pooling2d_1 (MaxPooling2 (None, 16, 16, 64)        0
_________________________________________________________________
conv2d_2 (Conv2D)            (None, 16, 16, 128)       295040
_________________________________________________________________
activation_2 (Activation)    (None, 16, 16, 128)       0
_________________________________________________________________
batch_normalization_2 (Batch (None, 16, 16, 128)       512
_________________________________________________________________
max_pooling2d_2 (MaxPooling2 (None, 8, 8, 128)         0
_________________________________________________________________
conv2d_3 (Conv2D)            (None, 8, 8, 64)          294976
_________________________________________________________________
activation_3 (Activation)    (None, 8, 8, 64)          0
_________________________________________________________________
batch_normalization_3 (Batch (None, 8, 8, 64)          256
_________________________________________________________________
up_sampling2d_1 (UpSampling2 (None, 16, 16, 64)        0
_________________________________________________________________
conv2d_4 (Conv2D)            (None, 16, 16, 32)        73760
_________________________________________________________________
activation_4 (Activation)    (None, 16, 16, 32)        0
_________________________________________________________________
batch_normalization_4 (Batch (None, 16, 16, 32)        128
_________________________________________________________________
conv2d_5 (Conv2D)            (None, 16, 16, 16)        18448
_________________________________________________________________
activation_5 (Activation)    (None, 16, 16, 16)        0
_________________________________________________________________
batch_normalization_5 (Batch (None, 16, 16, 16)        64
_________________________________________________________________
up_sampling2d_2 (UpSampling2 (None, 32, 32, 16)        0
_________________________________________________________________
conv2d_6 (Conv2D)            (None, 32, 32, 3)         1731
=================================================================
Total params: 692,147
Trainable params: 691,539
Non-trainable params: 608
_________________________________________________________________
Train on 48000 samples, validate on 12000 samples

Epoch 39/150
48000/48000 [==============================] - 47s 984us/step - loss: 7.9963e-04 - acc: 0.8497 - val_loss: 7.8588e-04 - val_acc: 0.8687
12000/12000 [==============================] - 1s 123us/step
val loss : 0.0007858769219989578
val_acc : 0.868697509765625

10/10 [==============================] - 0s 11ms/step

Using TensorFlow backend.

_________________________________________________________________

Layer (type) Output Shape Param #

=================================================================

conv2d_1 (Conv2D) (None, 32, 32, 64) 6976

_________________________________________________________________

activation_1 (Activation) (None, 32, 32, 64) 0

_________________________________________________________________

batch_normalization_1 (Batch (None, 32, 32, 64) 256

_________________________________________________________________

max_pooling2d_1 (MaxPooling2 (None, 16, 16, 64) 0

_________________________________________________________________

conv2d_2 (Conv2D) (None, 16, 16, 128) 295040

_________________________________________________________________

activation_2 (Activation) (None, 16, 16, 128) 0

_________________________________________________________________

batch_normalization_2 (Batch (None, 16, 16, 128) 512

_________________________________________________________________

max_pooling2d_2 (MaxPooling2 (None, 8, 8, 128) 0

_________________________________________________________________

conv2d_3 (Conv2D) (None, 8, 8, 64) 294976

_________________________________________________________________

activation_3 (Activation) (None, 8, 8, 64) 0

_________________________________________________________________

batch_normalization_3 (Batch (None, 8, 8, 64) 256

_________________________________________________________________

up_sampling2d_1 (UpSampling2 (None, 16, 16, 64) 0

_________________________________________________________________

conv2d_4 (Conv2D) (None, 16, 16, 32) 73760

_________________________________________________________________

activation_4 (Activation) (None, 16, 16, 32) 0

_________________________________________________________________

batch_normalization_4 (Batch (None, 16, 16, 32) 128

_________________________________________________________________

conv2d_5 (Conv2D) (None, 16, 16, 16) 18448

_________________________________________________________________

activation_5 (Activation) (None, 16, 16, 16) 0

_________________________________________________________________

batch_normalization_5 (Batch (None, 16, 16, 16) 64

_________________________________________________________________

up_sampling2d_2 (UpSampling2 (None, 32, 32, 16) 0

_________________________________________________________________

conv2d_6 (Conv2D) (None, 32, 32, 3) 1731

=================================================================

Total params: 692,147

Trainable params: 691,539

Non-trainable params: 608

_________________________________________________________________

Train on 48000 samples, validate on 12000 samples

Epoch 39/150

48000/48000 [==============================] - 47s 984us/step - loss: 7.9963e-04 - acc: 0.8497 - val_loss: 7.8588e-04 - val_acc: 0.8687

12000/12000 [==============================] - 1s 123us/step

val loss : 0.0007858769219989578

val_acc : 0.868697509765625

10/10 [==============================] - 0s 11ms/step

それでは詳しく説明していきます。

プログラムの詳細

importライブラリとdefine関数

まずは定義したライブラリと関数について説明します。

import numpy as np
import tensorflow as tf
from keras import backend as K
from keras import callbacks, layers, models
from keras.datasets import cifar10
from keras.layers import (BatchNormalization, Conv2D, Dense, LeakyReLU,
                          MaxPooling2D, UpSampling2D)
from keras.layers.core import Activation, Dense
from keras.optimizers import Adam
from sklearn.model_selection import train_test_split

def normalizer(x, axis = None):
        min = x.min(axis = axis, keepdims = True)
        max = x.max(axis = axis, keepdims = True)
        result = (x - min) / (max - min)

        return result

import numpy as np

import tensorflow as tf

from keras import backend as K

from keras import callbacks, layers, models

from keras.datasets import cifar10

from keras.layers import (BatchNormalization, Conv2D, Dense, LeakyReLU,

MaxPooling2D, UpSampling2D)

from keras.layers.core import Activation, Dense

from keras.optimizers import Adam

from sklearn.model_selection import train_test_split

def normalizer(x, axis = None):

min = x.min(axis = axis, keepdims = True)

max = x.max(axis = axis, keepdims = True)

result = (x - min) / (max - min)

return result

まず初めにこれらのライブラリをインポートします。

これについては詳しく説明する必要はないと思います。

その下のnormalizer関数では画像のデータを0~1の間にノーマライズ（正規化）するために定義します。

なぜノーマライズが必要なのかというと、機械が学習するときに扱いやすい数字にするために変換します。

扱いやすい数字にするとどうなるかというと、重みの更新が効率化されたり、誤差が収束しやすくなります。

機械学習・ディープラーニングをする上では重要になってきますので正規化については理解できるようにしておきましょう。

画像データを0~1の間にノーマライズ（正規化）する。
正規化することで学習のスピードを上げることができる。
正規化することで損失関数の誤差が収束しやすくなる。

main()関数内

main()関数内はネットワークのモデルを定義しています。

def main():

        (x_train, y_train), (x_test, y_test) = cifar10.load_data()

        image_height, image_width = 32, 32
        epochs = 150
        learn_rate = 0.0002
        batch_size = 8

        train_data = np.concatenate([x_train, x_test])
        label_data = np.concatenate([y_train, y_test])

        x_train, x_test, y_train, y_test = train_test_split(train_data, label_data, \
                                                      test_size = 0.2, shuffle = True)
        model = models.Sequential()

        model.add(Conv2D(64, kernel_size = 6, padding = 'same', input_shape = (32, 32, 3),\
                           kernel_initializer = 'random_uniform', bias_initializer = 'zeros'))
        model.add(Activation(LeakyReLU(alpha = 0.5623815057565213)))
        model.add(BatchNormalization())
        model.add(MaxPooling2D(pool_size = (2, 2)))

        model.add(Conv2D(128, kernel_size = 6, padding = 'same'))
        model.add(Activation(LeakyReLU(alpha = 0.3949936266626689)))
        model.add(BatchNormalization())
        model.add(MaxPooling2D(pool_size = (2, 2)))

        model.add(Conv2D(64, kernel_size = 6, padding = "same"))
        model.add(Activation(LeakyReLU(alpha = 0.9758185183456943)))
        model.add(BatchNormalization())
        model.add(UpSampling2D(size = (2, 2)))

        model.add(Conv2D(32, kernel_size = 6, padding = "same"))
        model.add(Activation(LeakyReLU(alpha = 0.288662535902546)))
        model.add(BatchNormalization())
        model.add(Conv2D(16, kernel_size = 6, padding = "same"))
        model.add(Activation(LeakyReLU(alpha = 0.3376263809413643)))
        model.add(BatchNormalization())
        model.add(UpSampling2D(size = (2, 2)))

        model.add(Conv2D(3, kernel_size = 6, padding = "same", activation = "sigmoid"))
        
        model.compile(optimizer = Adam(lr = learn_rate), loss = 'mse', metrics = ['accuracy'])

        model.summary()

        fit_callbacs = [callbacks.EarlyStopping(monitor = 'val_loss', patience = 5, \
                                                                                mode = 'min')]

        hist = model.fit(x_train, x_train, epochs = epochs, batch_size = batch_size, \
                                           shuffle = True, validation_data = (x_test, x_test), \
                                                         callbacks = fit_callbacs, verbose = 1)

        val_loss, val_acc = model.evaluate(x_test, x_test, verbose = 1)

def main():

(x_train, y_train), (x_test, y_test) = cifar10.load_data()

image_height, image_width = 32, 32

epochs = 150

learn_rate = 0.0002

batch_size = 8

train_data = np.concatenate([x_train, x_test])

label_data = np.concatenate([y_train, y_test])

x_train, x_test, y_train, y_test = train_test_split(train_data, label_data, \

test_size = 0.2, shuffle = True)

model = models.Sequential()

model.add(Conv2D(64, kernel_size = 6, padding = 'same', input_shape = (32, 32, 3),\

kernel_initializer = 'random_uniform', bias_initializer = 'zeros'))

model.add(Activation(LeakyReLU(alpha = 0.5623815057565213)))

model.add(BatchNormalization())

model.add(MaxPooling2D(pool_size = (2, 2)))

model.add(Conv2D(128, kernel_size = 6, padding = 'same'))

model.add(Activation(LeakyReLU(alpha = 0.3949936266626689)))

model.add(BatchNormalization())

model.add(MaxPooling2D(pool_size = (2, 2)))

model.add(Conv2D(64, kernel_size = 6, padding = "same"))

model.add(Activation(LeakyReLU(alpha = 0.9758185183456943)))

model.add(BatchNormalization())

model.add(UpSampling2D(size = (2, 2)))

model.add(Conv2D(32, kernel_size = 6, padding = "same"))

model.add(Activation(LeakyReLU(alpha = 0.288662535902546)))

model.add(BatchNormalization())

model.add(Conv2D(16, kernel_size = 6, padding = "same"))

model.add(Activation(LeakyReLU(alpha = 0.3376263809413643)))

model.add(BatchNormalization())

model.add(UpSampling2D(size = (2, 2)))

model.add(Conv2D(3, kernel_size = 6, padding = "same", activation = "sigmoid"))

model.compile(optimizer = Adam(lr = learn_rate), loss = 'mse', metrics = ['accuracy'])

model.summary()

fit_callbacs = [callbacks.EarlyStopping(monitor = 'val_loss', patience = 5, \

mode = 'min')]

hist = model.fit(x_train, x_train, epochs = epochs, batch_size = batch_size, \

shuffle = True, validation_data = (x_test, x_test), \

callbacks = fit_callbacs, verbose = 1)

val_loss, val_acc = model.evaluate(x_test, x_test, verbose = 1)

3行目でデータセットの読み込みをしています。

5〜8行目ではそれぞれのパラメータの設定をしています。

10,11行目では、28,29行目でデータを8:2に割り振るために、学習データと評価データをそれぞれ結合しています。

もともとの（学習データ：評価データ）＝（50000：10000）で学習させていたのですが、これでも評価データが少なかったようで、損失関数の値が安定しなかったのでこの様にしました。

33行目以下はモデルの定義になります。

モデル内ではCNNを使い、過学習防止のために BatchNormalizetion() を使っています。

過学習防止策は BatchNormalizetion() の他に Dropout() が有名ですがBatchNormalizetionはDropoutよりも有効であると言われています。

Batch Normalizeについての詳しい説明は「Deep LearningにおけるBatch Normalizationの理解メモと、実際にその効果を見てみる」で紹介されていますのでこちらを参照してください。

35行目はInput層です。

少し工夫して重み、バイアスベクトルを初期化しています。

説明すると長くなるのでここでは割愛します。

後は普通のCNNですが、 BatchNormalizetion() は Activation(LeakyReLU) の次に定義する様にしています。

Convolutionのフィルターは64→128→64→32→16→3というふうに増減させています。

学習したときになかなか学習率が上がらなかったらフィルター数を増やすのがセオリーです。

また、このモデル内のパラメータは全てhyperasを使ってパラメータチューニングをし求めました。

おまけ

ちなみに、実行した結果などをグラフ化したり、学習した画像などを保存したりするプログラムを作成したのでそちらも合わせてご覧ください。

画像を保存するフォルダーを日時と時間から作成

def into_dir():
        dt_now = datetime.datetime.now()
        date = str(dt_now.year) + str(dt_now.month) + str(dt_now.day) + "-" + str(dt_now.hour) + "-" + str(dt_now.minute) + "-" + str(dt_now.second)
        os.mkdir('./cifar_model_result/{}'.format(date))

        return dt_now, date

def into_dir():

dt_now = datetime.datetime.now()

date = str(dt_now.year) + str(dt_now.month) + str(dt_now.day) + "-" + str(dt_now.hour) + "-" + str(dt_now.minute) + "-" + str(dt_now.second)

os.mkdir('./cifar_model_result/{}'.format(date))

return dt_now, date

学習した結果を保存するとき、ファイル名が被ってしまったりするのめんどくさいので学習した時間をフォルダー名にして学習ごとの結果をそのフォルダーに入れる様にします。

学習にかかった時間を表示

finish_time = time.time()


processing_time = (finish_time - start_time) / 60

print("===============================")
print("Processing time:{:.2f}".format(processing_time), "min")
print("===============================")

finish_time = time.time()

processing_time = (finish_time - start_time) / 60

print("===============================")

print("Processing time:{:.2f}".format(processing_time), "min")

print("===============================")

これは有名ですが、処理にかかった時間を表示します。

学習した画像や損失、正解率を画像、グラフで表示

p = np.random.random_integers(0, len(x_test), 10)
x_test_sampled = x_test[p]
y_test_sample_lable = y_test[p]
x_test_sampled_pred = model.predict_proba(x_test_sampled, verbose = 1)

fig, axes = plt.subplots(2, 10)

for i, label in enumerate(y_test[p]):

        img = x_test_sampled[i].reshape(image_height, image_width, 3)
        axes[0][i].imshow(img, cmap = cm.gray_r)
        axes[0][i].axis('off')
        axes[0][i].set_title(label, color = 'red')

        pred_img = x_test_sampled_pred[i].reshape(image_height, image_width, 3)
        axes[1][i].imshow(pred_img, cmap = cm.gray_r)
        axes[1][i].axis('off')

plt.savefig('./cifar_model_result/{}/Probacation_Image.png'.format(date))
#plt.show()

#Accuracy Graph
plt.rc('font', family = 'serif')
fig = plt.figure(figsize = (13, 6))
plt.subplot(1, 2, 1)
plt.plot(hist.history['acc'], label = 'Train Acc', color = 'red')
plt.plot(hist.history['val_acc'], label = 'Test Acc', color = 'm')
plt.title('Model Accuracy')
plt.xlabel('Epochs')
plt.ylabel('accuracy')
plt.legend(bbox_to_anchor = (0.7, 0.5), loc = 'center left', borderaxespad = 0, fontsize = 8)

#Loss Graph
plt.rc('font', family = 'serif')
plt.subplot(1, 2, 2)
plt.plot(hist.history['loss'], label = 'Train Loss', color = 'blue')
plt.plot(hist.history['val_loss'], label = 'Test Loss', color = 'c')
plt.title('Model Loss')
plt.xlabel('Epochs')
plt.ylabel('loss')
plt.legend(bbox_to_anchor = (0.7, 0.5), loc = 'center left', borderaxespad = 0, fontsize = 8)
plt.savefig ('./cifar_model_result/{}/Accuracy_Graph.png'.format(date))
#plt.show()

p = np.random.random_integers(0, len(x_test), 10)

x_test_sampled = x_test[p]

y_test_sample_lable = y_test[p]

x_test_sampled_pred = model.predict_proba(x_test_sampled, verbose = 1)

fig, axes = plt.subplots(2, 10)

for i, label in enumerate(y_test[p]):

img = x_test_sampled[i].reshape(image_height, image_width, 3)

axes[0][i].imshow(img, cmap = cm.gray_r)

axes[0][i].axis('off')

axes[0][i].set_title(label, color = 'red')

pred_img = x_test_sampled_pred[i].reshape(image_height, image_width, 3)

axes[1][i].imshow(pred_img, cmap = cm.gray_r)

axes[1][i].axis('off')

plt.savefig('./cifar_model_result/{}/Probacation_Image.png'.format(date))

#plt.show()

#Accuracy Graph

plt.rc('font', family = 'serif')

fig = plt.figure(figsize = (13, 6))

plt.subplot(1, 2, 1)

plt.plot(hist.history['acc'], label = 'Train Acc', color = 'red')

plt.plot(hist.history['val_acc'], label = 'Test Acc', color = 'm')

plt.title('Model Accuracy')

plt.xlabel('Epochs')

plt.ylabel('accuracy')

plt.legend(bbox_to_anchor = (0.7, 0.5), loc = 'center left', borderaxespad = 0, fontsize = 8)

#Loss Graph

plt.rc('font', family = 'serif')

plt.subplot(1, 2, 2)

plt.plot(hist.history['loss'], label = 'Train Loss', color = 'blue')

plt.plot(hist.history['val_loss'], label = 'Test Loss', color = 'c')

plt.title('Model Loss')

plt.xlabel('Epochs')

plt.ylabel('loss')

plt.legend(bbox_to_anchor = (0.7, 0.5), loc = 'center left', borderaxespad = 0, fontsize = 8)

plt.savefig ('./cifar_model_result/{}/Accuracy_Graph.png'.format(date))

#plt.show()

上記のプログラム1行目~18行目では、ランダムに選んだ画像10枚の推論前、推論後を出力します。

出力された結果

こんな感じで出力されます。

上の段が入力された画像で、下の段がAutoEncoderされた画像です。

22行目以下は損失関数のグラフと正解率のグラフを出力するプログラムです。

こんな感じで出力されます。

モデルや学習結果、エポック数、学習率などの情報をテキストファイルで保存

dt_now, date = into_dir()

with open("./cifar_model_result/{}/Summary.txt".format(date), "w") as fp:
        model.summary(print_fn = lambda x: fp.write(x + "\r\n"))

write_date = str(dt_now) + "\n\n" + "Data_Name:CIFAR10" + "\n\n" + "x_train:{}".format(x_train.shape) + "\n\n" + "x_test:{}".format(x_test.shape) + "\n\n" + "y_train:{}".format(y_train.shape) + "\n\n" + "y_test:{}".format(y_test.shape) + "\n\n" + "Epochs:{}".format(epochs) + "epochs" + "\n\n" + "Learn Rate:{}".format(learn_rate) + "%" + "\n\n" + "Batch_Size:{}".format(batch_size) + "\n\n" + "Loss Value:"+ str(f'{val_loss:.10f}') + "\n\n" + "Acc Value:" + str(f'{val_acc:.4f}') + "%" + "\n\n" + "Processing Time:{:.2} minuites".format(processing_time)

with open("./cifar_model_result/{}/Result.txt".format(date), "w") as f:
         f.write(write_date)

dt_now, date = into_dir()

with open("./cifar_model_result/{}/Summary.txt".format(date), "w") as fp:

model.summary(print_fn = lambda x: fp.write(x + "\r\n"))

write_date = str(dt_now) + "\n\n" + "Data_Name:CIFAR10" + "\n\n" + "x_train:{}".format(x_train.shape) + "\n\n" + "x_test:{}".format(x_test.shape) + "\n\n" + "y_train:{}".format(y_train.shape) + "\n\n" + "y_test:{}".format(y_test.shape) + "\n\n" + "Epochs:{}".format(epochs) + "epochs" + "\n\n" + "Learn Rate:{}".format(learn_rate) + "%" + "\n\n" + "Batch_Size:{}".format(batch_size) + "\n\n" + "Loss Value:"+ str(f'{val_loss:.10f}') + "\n\n" + "Acc Value:" + str(f'{val_acc:.4f}') + "%" + "\n\n" + "Processing Time:{:.2} minuites".format(processing_time)