編集履歴

質問編集履歴

コードの追加

2020/05/24 05:33

投稿

tomoki_fab

スコア25

test CHANGED Viewed

File without changes

test CHANGED Viewed

@@ -95,3 +95,379 @@
 インターネットで手当たり次第に調べたり、書籍で参考になりそうな情報を探したり、手を尽くしましたがなぜ全くうまくいかないのかわかりません。問題がレイヤーの組み方なのか、何か決定的に重要な処理を飛ばしてしまっているのか、損失関数などの評価の仕方が悪いのかも見当もつきません。
 なにか見落としているポイントなどがありますでしょうか？
+コード
+```Python
+import numpy as np
+from PIL import Image
+from keras.datasets import mnist
+from keras.layers import *
+from keras.models import *
+from keras.optimizers import *
+# --- プログレスバーを表示するクラス(学習自体には関係ありません) ---------
+class ProgressBar:
+    def __init__(self, entireJob):
+        self.job = entireJob
+        self.width = 40
+    def draw(self, progress):
+        print( ("\r["+"#"*int((progress+1)*self.width/self.job)+" "*(self.width-int((progress+1)*self.width/self.job) ) +"]  %d/%d")%(progress+1,self.job), end="")
+# --- Generatorモデルの定義 -----------
+class Generator:
+    def __init__(self):
+        layer0 = Input(shape=(1,1,100))
+        layer1 = UpSampling2D(size=(3,3))(layer0)
+        layer1 = Conv2D(
+            filters=100,
+            kernel_size=(2,2),
+            strides=(1,1),
+            padding='same',
+            activation='relu' )(layer1)
+        layer1 = BatchNormalization()(layer1)
+        layer2 = UpSampling2D(size=(3,3))(layer1)
+        layer2 = Conv2D(
+            filters=100,
+            kernel_size=(2,2),
+            strides=(1,1),
+            padding='same',
+            activation='relu' )(layer2)
+        layer2 = BatchNormalization()(layer2)
+        layer3 = UpSampling2D(size=(2,2))(layer2)
+        layer3 = Conv2D(
+            filters=80,
+            kernel_size=(3,3),
+            strides=(1,1),
+            padding='valid',
+            activation='elu' )(layer3)
+        layer3 = BatchNormalization()(layer3)
+        layer4 = UpSampling2D(size=(2,2))(layer3)
+        layer4 = Conv2D(
+            filters=50,
+            kernel_size=(3,3),
+            strides=(1,1),
+            padding='same',
+            activation='elu' )(layer4)
+        layer4 = BatchNormalization()(layer4)
+        layer5 = UpSampling2D(size=(2,2))(layer4)
+        layer5 = Conv2D(
+            filters=20,
+            kernel_size=(4,4),
+            strides=(2,2),
+            padding='valid',
+            activation='elu' )(layer5)
+        layer5 = BatchNormalization()(layer5)
+        layer6 = Conv2D(
+            filters=1,
+            kernel_size=(4,4),
+            strides=(1,1),
+            padding='valid',
+            activation='tanh' )(layer5)
+        self.model = Model(layer0, layer6)
+        self.model.summary()
+# --- Discriminatorモデルの定義 -------
+class Discriminator:
+    def __init__(self):
+        layer0 = Input(shape=(28,28,1))
+        layer1 = Conv2D(
+            filters=5,
+            kernel_size=(3,3),
+            strides=(2,2),
+            padding='valid',
+            activation='elu' )(layer0)
+        layer1 = BatchNormalization()(layer1)
+        layer2 = Conv2D(
+            filters=10,
+            kernel_size=(3,3),
+            strides=(2,2),
+            padding='valid',
+            activation='elu' )(layer1)
+        layer2 = BatchNormalization()(layer2)
+        layer3 = Conv2D(
+            filters=5,
+            kernel_size=(3,3),
+            strides=(1,1),
+            padding='valid',
+            activation='relu' )(layer2)
+        layer3 = BatchNormalization()(layer3)
+        layer4 = Flatten()(layer3)
+        layer4 = Dense(units=30, activation='tanh')(layer4)
+        layer4 = BatchNormalization()(layer4)
+        layer5 = Dense(units=1, activation='sigmoid' )(layer4)
+        self.model = Model(layer0, layer5)
+        self.model.summary()
+class Main:
+    def __init__(self):
+        # --- Discriminatorの定義 -----------------
+        self.discriminator = Discriminator().model
+        self.discriminator.compile(
+            optimizer=SGD(learning_rate=1e-4),
+            loss='binary_crossentropy',
+            metrics=['accuracy'] )
+        # --- GeneratorとDiscriminatorを連結したモデルの定義 ---
+        self.generator = Generator().model
+        z = Input(shape=(1,1,100))
+        img = self.generator(z)
+        self.discriminator.trainable = False    # Discriminatorを更新しないよう設定
+        valid = self.discriminator(img)
+        self.combined = Model(z, valid)
+        self.combined.compile(
+            optimizer=Adam(learning_rate=1e-6),
+            loss='binary_crossentropy',
+            metrics=['accuracy'] )
+        # --- MNISTデータセットの用意 ---------------
+        (x_train, t_train), (x_test, t_test) = mnist.load_data()
+        x_train = x_train.reshape(60000, 28, 28, 1)
+        x_test = x_test.reshape(10000, 28, 28, 1)
+        self.x_train = x_train.astype('float32')
+        self.x_test = x_test.astype('float32')
+    # --- 学習 -------------------------------------
+    def _train(self, iteration, batch_size):
+        progress = ProgressBar(iteration)       # プログレスバーを用意
+        for i in range(iteration):
+            z = np.random.uniform(-1,1,(batch_size//2,1,1,100))     # ノイズベクトルの生成
+            f_img = self.generator.predict(z)                       # f_img(fake_img)の生成
+            r_img = self.x_train[np.random.randint(0, 60000, batch_size//2)]        # r_img(real_img)を読み込み
+            loss_d, acc_d = self.discriminator.train_on_batch(f_img, np.zeros((batch_size//2,1)))   # Discriminatorの学習
+            loss_d_, acc_d_ = self.discriminator.train_on_batch(r_img, np.ones( (batch_size//2,1))) # acc_d = Discriminatorのaccuracy
+            acc_d += acc_d_
+            z = np.random.uniform(-1,1,(batch_size,1,1,100))        # ノイズベクトルの生成
+            loss_g, acc_g = self.combined.train_on_batch(z, np.ones((batch_size,1)))    # Generatorの学習
+            progress.draw(i)        # プログレスバーの表示
+            print("   Accuracy=(%f,%f)"%(acc_g, acc_d/2), end="")
+    def train(self, iteration, batch_size, epoch):
+        for i in range(epoch):
+            print("Epoch %d/%d\n"%(i+1, epoch))
+            self._train(iteration, batch_size)      # _train()をepoch回繰り返します
+    # --- 学習が終わった時の確認用に一枚だけ画像を作ります -------
+    def create_image(self):
+        z = np.random.uniform(-1,1,(1,1,1,100))
+        img = self.generator.predict(z)
+        return img.reshape(1,28,28)
+if __name__ == "__main__":
+    main = Main()
+    main.train(iteration=1875, batch_size=32, epoch=1)
+    # --- 画像を表示　-----------------------
+    img = main.create_image()
+    img = Image.fromarray(np.uint8(img.reshape(28,28) * 255))
+    img.show()
+    img.save("gan_generated_img.png")
+```