TensorFlow + Keras を使って MNIST（手書き数字データ）を入力とした学習を動かしたときのメモです． シンプルなCNNの識別器とDCGANを動かします． {{small:コード全文:[link:https://github.com/sanko-shoko/simplesp_py/tree/master/mnist] }} *インストール tensorflow keras の順にインストールします． {/ pip install tensorflow pip install keras /} gpuを使う場合はこちらです． {/ pip install tensorflow_gpu pip install keras /} インストールしたバージョンを確認します． {/ pip list ... Keras (2.0.6) tensorflow (1.2.1) /} *simple CNN 主な参考元 {{small:[link:https://github.com/fchollet/keras/blob/master/examples/mnist_cnn.py] }} 畳み込み層x2 と 全結合層x2 の比較的簡単なネットワークを構築します． **準備 一通り必要な機能のインポートとデータセットの取得を行います． {# from __future__ import print_function import os import keras from keras.datasets import mnist from keras.models import Sequential from keras.layers import Dense, Conv2D from keras.layers import Dropout, Flatten, MaxPooling2D # load mnist (x_train, y_train), (x_test, y_test) = mnist.load_data() # normalize x_train = x_train.astype('float32') / 255 x_test = x_test.astype('float32') / 255 # one hot y_train = keras.utils.to_categorical(y_train, 10) y_test = keras.utils.to_categorical(y_test, 10) # reshape shape = (28, 28, 1) x_train = x_train.reshape(-1, shape[0], shape[1], shape[2]) x_test = x_test.reshape(-1, shape[0], shape[1], shape[2]) print('x_train :', x_train.shape) print('x_test  :', x_test.shape) print('y_train :', y_train.shape) print('y_test  :', y_test.shape) #} ここまでの出力 {/ Using TensorFlow backend. x_train : (60000, 28, 28, 1) x_test  : (10000, 28, 28, 1) y_train : (60000, 10) y_test  : (10000, 10) /} MNISTは学習用の画像が6万枚，テスト用の画像が1万枚あります． 画像のサイズは28x28で，グレースケールの1チャンネルです． 各画像に設定されているラベルは0～9までの10種類です． **モデルの設定 {# # define model def cnn_model(class_num, input_shape):     model = Sequential()     model.add(Conv2D(32, (3, 3), activation='relu', input_shape=input_shape))     model.add(Conv2D(64, (3, 3), activation='relu'))     model.add(MaxPooling2D(pool_size=(2, 2)))     model.add(Dropout(0.25))     model.add(Flatten())     model.add(Dense(128, activation='relu'))     model.add(Dropout(0.5))     model.add(Dense(class_num, activation='softmax'))     return model model = cnn_model(10, shape) model.summary() opt = keras.optimizers.Adadelta() model.compile(loss=keras.losses.categorical_crossentropy, optimizer=opt, metrics=['accuracy']) #} ここまでの出力 {/ _________________________________________________________________ Layer (type)                 Output Shape              Param # ================================================================= conv2d_1 (Conv2D)            (None, 26, 26, 32)        320 _________________________________________________________________ conv2d_2 (Conv2D)            (None, 24, 24, 64)        18496 _________________________________________________________________ max_pooling2d_1 (MaxPooling2 (None, 12, 12, 64)        0 _________________________________________________________________ dropout_1 (Dropout)          (None, 12, 12, 64)        0 _________________________________________________________________ flatten_1 (Flatten)          (None, 9216)              0 _________________________________________________________________ dense_1 (Dense)              (None, 128)               1179776 _________________________________________________________________ dropout_2 (Dropout)          (None, 128)               0 _________________________________________________________________ dense_2 (Dense)              (None, 10)                1290 ================================================================= Total params: 1,199,882 Trainable params: 1,199,882 Non-trainable params: 0 _________________________________________________________________ /} **学習と評価 {# epochs = 10 batch_size = 128 param_folder = './param' if not os.path.isdir(param_folder):     os.makedirs(param_folder) # call back (for save param) cbk = keras.callbacks.ModelCheckpoint(filepath = os.path.join(param_folder, 'param{epoch:02d}.hdf5')) # train model.fit(x_train, y_train, batch_size, epochs=epochs, verbose = 1, callbacks = [cbk], validation_data = (x_test, y_test))       # test result = model.predict(x_test) score = model.evaluate(x_test, y_test, verbose = 0) print('test result:', result.argmax(axis=1)) print('test loss and accuracy:', score) #} ここまでの出力 {/ Train on 60000 samples, validate on 10000 samples Epoch 1/10 60000/60000 [==============================] - 146s - loss: 0.3395 - acc: 0.8978 - val_loss: 0.0793 - val_acc: 0.9750 Epoch 2/10 60000/60000 [==============================] - 142s - loss: 0.1150 - acc: 0.9667 - val_loss: 0.0529 - val_acc: 0.9828 Epoch 3/10 60000/60000 [==============================] - 148s - loss: 0.0872 - acc: 0.9747 - val_loss: 0.0465 - val_acc: 0.9858 Epoch 4/10 60000/60000 [==============================] - 155s - loss: 0.0720 - acc: 0.9784 - val_loss: 0.0410 - val_acc: 0.9858 Epoch 5/10 60000/60000 [==============================] - 144s - loss: 0.0629 - acc: 0.9813 - val_loss: 0.0357 - val_acc: 0.9873 Epoch 6/10 60000/60000 [==============================] - 140s - loss: 0.0565 - acc: 0.9831 - val_loss: 0.0353 - val_acc: 0.9877 Epoch 7/10 60000/60000 [==============================] - 132s - loss: 0.0522 - acc: 0.9844 - val_loss: 0.0342 - val_acc: 0.9885 Epoch 8/10 60000/60000 [==============================] - 139s - loss: 0.0466 - acc: 0.9859 - val_loss: 0.0301 - val_acc: 0.9896 Epoch 9/10 60000/60000 [==============================] - 140s - loss: 0.0436 - acc: 0.9866 - val_loss: 0.0300 - val_acc: 0.9892 Epoch 10/10 60000/60000 [==============================] - 149s - loss: 0.0417 - acc: 0.9875 - val_loss: 0.0299 - val_acc: 0.9897 test result: [7 2 1 ..., 4 5 6] test loss and accuracy: [0.029858358073154886, 0.98970000000000002] /} *DCGAN 主な参考元 {{small:[link:https://github.com/jacobgil/keras-dcgan] }} **準備 一通り必要な機能のインポートとデータセットの取得を行います． {# from __future__ import print_function import os import numpy as np from PIL import Image import keras from keras.datasets import mnist from keras.models import Sequential from keras.layers import Dense, Conv2D, BatchNormalization from keras.layers import Activation, Flatten, Dropout, UpSampling2D, MaxPooling2D, Reshape # load mnist (x_train, y_train), (x_test, y_test) = mnist.load_data() # normalize x_train = (x_train.astype('float32') - 127.5) / 127.5 # reshape shape = (28, 28, 1) x_train = x_train.reshape(-1, shape[0], shape[1], shape[2]) print('x_train :', x_train.shape) #} ここまでの出力 {/ Using TensorFlow backend. x_train : (60000, 28, 28, 1) /} **モデルの設定 {# # define model def generator_model():     model = Sequential()     model.add(Dense(1024, input_shape = (100, )))     model.add(Activation('tanh'))     model.add(Dense(7 * 7 * 128))     model.add(BatchNormalization())     model.add(Activation('tanh'))     model.add(Reshape((7, 7, 128), input_shape = (7 * 7 * 128,)))     model.add(UpSampling2D(size = (2, 2)))     model.add(Conv2D(64, (5, 5), padding='same'))     model.add(Activation('tanh'))     model.add(UpSampling2D(size = (2, 2)))     model.add(Conv2D(1, (5, 5), padding = 'same'))     model.add(Activation('tanh'))     return model def discriminator_model():     model = Sequential()     model.add(Conv2D(64, (5, 5), padding='same', input_shape=(28, 28, 1)))     model.add(Activation('tanh'))     model.add(MaxPooling2D(pool_size=(2, 2)))     model.add(Conv2D(128, (5, 5)))     model.add(Activation('tanh'))     model.add(MaxPooling2D(pool_size=(2, 2)))     model.add(Flatten())     model.add(Dense(1024))     model.add(Activation('tanh'))     model.add(Dense(1))     model.add(Activation('sigmoid'))     return model def combined_model(generator, discriminator):     model = Sequential()     model.add(generator)     model.add(discriminator)     return model generator = generator_model() generator.summary() discriminator = discriminator_model() discriminator.summary() discriminator.trainable = False combined = combined_model(generator, discriminator) combined.summary() opt = keras.optimizers.SGD(lr=0.0005, momentum=0.9, nesterov=True) discriminator.trainable = True discriminator.compile(loss='binary_crossentropy', optimizer=opt) discriminator.trainable = False combined.compile(loss='binary_crossentropy', optimizer=opt) #} ここまでの出力 {/ _________________________________________________________________ Layer (type)                 Output Shape              Param # ================================================================= dense_1 (Dense)              (None, 1024)              103424 _________________________________________________________________ activation_1 (Activation)    (None, 1024)              0 _________________________________________________________________ dense_2 (Dense)              (None, 6272)              6428800 _________________________________________________________________ batch_normalization_1 (Batch (None, 6272)              25088 _________________________________________________________________ activation_2 (Activation)    (None, 6272)              0 _________________________________________________________________ reshape_1 (Reshape)          (None, 7, 7, 128)         0 _________________________________________________________________ up_sampling2d_1 (UpSampling2 (None, 14, 14, 128)       0 _________________________________________________________________ conv2d_1 (Conv2D)            (None, 14, 14, 64)        204864 _________________________________________________________________ activation_3 (Activation)    (None, 14, 14, 64)        0 _________________________________________________________________ up_sampling2d_2 (UpSampling2 (None, 28, 28, 64)        0 _________________________________________________________________ conv2d_2 (Conv2D)            (None, 28, 28, 1)         1601 _________________________________________________________________ activation_4 (Activation)    (None, 28, 28, 1)         0 ================================================================= Total params: 6,763,777 Trainable params: 6,751,233 Non-trainable params: 12,544 _________________________________________________________________ _________________________________________________________________ Layer (type)                 Output Shape              Param # ================================================================= conv2d_3 (Conv2D)            (None, 28, 28, 64)        1664 _________________________________________________________________ activation_5 (Activation)    (None, 28, 28, 64)        0 _________________________________________________________________ max_pooling2d_1 (MaxPooling2 (None, 14, 14, 64)        0 _________________________________________________________________ conv2d_4 (Conv2D)            (None, 10, 10, 128)       204928 _________________________________________________________________ activation_6 (Activation)    (None, 10, 10, 128)       0 _________________________________________________________________ max_pooling2d_2 (MaxPooling2 (None, 5, 5, 128)         0 _________________________________________________________________ flatten_1 (Flatten)          (None, 3200)              0 _________________________________________________________________ dense_3 (Dense)              (None, 1024)              3277824 _________________________________________________________________ activation_7 (Activation)    (None, 1024)              0 _________________________________________________________________ dense_4 (Dense)              (None, 1)                 1025 _________________________________________________________________ activation_8 (Activation)    (None, 1)                 0 ================================================================= Total params: 3,485,441 Trainable params: 3,485,441 Non-trainable params: 0 _________________________________________________________________ _________________________________________________________________ Layer (type)                 Output Shape              Param # ================================================================= sequential_1 (Sequential)    (None, 28, 28, 1)         6763777 _________________________________________________________________ sequential_2 (Sequential)    (None, 1)                 3485441 ================================================================= Total params: 10,249,218 Trainable params: 6,751,233 Non-trainable params: 3,497,985 _________________________________________________________________ /} GANではdiscriminator単体のモデル①と，generatorとdiscriminatorを連結したモデル②を交互に学習させます． モデル①の学習時は，discriminatorのパラメータを学習させます． モデル②の学習時は，discriminatorのパラメータは固定させ，generatorのパラメータのみを学習させます． **学習 {# epochs = 100 batch_size = 128 param_folder = './param' if not os.path.isdir(param_folder):     os.makedirs(param_folder) for epoch in range(epochs):     print('Epoch {}/{}'.format(epoch + 1, epochs))     itmax = int(x_train.shape[0] / batch_size)     progbar = keras.utils.generic_utils.Progbar(target = itmax)     for i in range(itmax):         # train discriminator         x = x_train[i * batch_size : (i + 1) * batch_size]         n = np.random.uniform(-1, 1, (batch_size, 100))         g = generator.predict(n, verbose=0)         y = [1] * batch_size + [0] * batch_size         d_loss = discriminator.train_on_batch(np.concatenate((x, g)), y)         # train generator         n = np.random.uniform(-1, 1, (batch_size, 100))         y = [1] * batch_size         g_loss = combined.train_on_batch(n, y)         progbar.add(1, values=[("d_loss", d_loss), ("g_loss", g_loss)])         # save image         if i % 20 == 0:             tmp = [r.reshape(-1, 28) for r in np.split(g[:100], 10)]             img = np.concatenate(tmp, axis = 1)             img = (img * 127.5 + 127.5).astype(np.uint8)             Image.fromarray(img).save("{}_{}.png".format(epoch, i))     # save param     generator.save_weights(os.path.join(param_folder, 'generator_{}.hdf5'.format(epoch)))     discriminator.save_weights(os.path.join(param_folder, 'discriminator_{}.hdf5'.format(epoch))) #} ここまでの出力 {/ Epoch 1/100 468/468 [==============================] - 1327s - d_loss: 0.4235 - g_loss: 1.0386 Epoch 2/100 468/468 [==============================] - 1366s - d_loss: 0.4928 - g_loss: 1.0905 Epoch 3/100 468/468 [==============================] - 1393s - d_loss: 0.4240 - g_loss: 1.2327 Epoch 4/100 468/468 [==============================] - 1406s - d_loss: 0.3942 - g_loss: 1.4242 Epoch 5/100 468/468 [==============================] - 1452s - d_loss: 0.2884 - g_loss: 1.9572 Epoch 6/100 468/468 [==============================] - 1465s - d_loss: 0.2263 - g_loss: 2.3618 Epoch 7/100 468/468 [==============================] - 1379s - d_loss: 0.2653 - g_loss: 2.4052 Epoch 8/100 468/468 [==============================] - 1358s - d_loss: 0.3320 - g_loss: 2.1082 Epoch 9/100 468/468 [==============================] - 1371s - d_loss: 0.3575 - g_loss: 2.0694 Epoch 10/100 468/468 [==============================] - 1363s - d_loss: 0.3413 - g_loss: 1.9845 ... /} [img:gdty] {{small:図１ generatorで生成される画像群}}