Keras/Tensorflow : CIFAR-10のMobileNetV2-likeなアーキテクチャを作った
0. 前提
MobileNetV2は、2018/04にGoogleのMark Sandlerらによって発表された論文："MobileNetV2: Inverted Residuals and Linear Bottlenecks"にて導入された、ニューラルネットワークモデルです。その名の通り、モバイル機器向けの高速なフォワード処理に適しており、精度を保持した上でパラメータの削減に成功しています。これは、同じくGoogleから発表された論文："Mobilenets: Efficient Convolutional Neural Networks for Mobile Vision Applications"で提唱されたMobileNetアーキテクチャのアップグレード版という位置付けで、出力チャネル数を明に増加させる「Expand-layer」やResNet系で馴染み深い「skip-connection」の導入などが特徴的です。
- MobileNetV2(原論文) : https://arxiv.org/pdf/1801.04381.pdf
- MobileNetV2(解説Blog) : http://machinethink.net/blog/mobilenet-v2
- Github - keras_applications :
1. 動作環境
OS: Ubuntu 16.04Package Version ------------------- ------- python 3.5.0 tensorboard 1.9.0 tensorflow 1.9.0 h5py 2.8.0 Keras 2.2.2 Keras-Applications 1.0.4 Keras-Preprocessing 1.0.2
2. プログラム
import_CIFAR-10.py
CIFAR-10データセットをnumpy配列（ndarray）形式で保存しておく実行ファイルです. CIFAR-10の画像は一枚あたり「32w(pixel) × 32h(pixel) × 3ch(RGB)」個のpixelからできています. 今回は, 画像データに対する前処理としてRGB-channelに応じて標準化（平均0, 標準偏差1）を行います.
データセット用のディレクトリと実行ファイルを用意しておくとバージョン管理が楽です.
import numpy as np from keras.datasets import cifar10 from keras.utils import np_utils nb_classes = 10 argmax_ch = 255.0 if __name__=='__main__': # load CIFAR-10 data (X_train, y_train), (X_test, y_test) = cifar10.load_data() # set data type as 'float32' X_train = X_train.astype('float32') #argmax_ch X_test = X_test.astype('float32') #argmax_ch def ch_wise_normalization(X_type, ch): mean_ch = X_type[:, :, :, ch].mean() std_ch = X_type[:, :, :, ch].std() X_type[:, :, :, ch] = (X_type[:, :, :, ch] - mean_ch) / std_ch return X_type[:, :, :, ch] # normalize data for each R-G-B(0, 1, 2) channel X_train[:, :, :, 0] = ch_wise_normalization(X_train, 0) X_train[:, :, :, 1] = ch_wise_normalization(X_train, 1) X_train[:, :, :, 2] = ch_wise_normalization(X_train, 2) X_test[:, :, :, 0] = ch_wise_normalization(X_test, 0) X_test[:, :, :, 1] = ch_wise_normalization(X_test, 1) X_test[:, :, :, 2] = ch_wise_normalization(X_test, 2) # convert class label (0-9) to one-hot encoding format y_train = np_utils.to_categorical(y_train, nb_classes) y_test = np_utils.to_categorical(y_test, nb_classes) # save datasets as np.ndarray class format files np.save('X_train', X_train) np.save('y_train', y_train) np.save('X_test' , X_test) np.save('y_test' , y_test)
mobilenetv2.py
MobileNetV2のアーキテクチャを定義してそのインスタンスを返すためのファイルです.
Kerasではkeras.applications.mobilenetv2.MobileNetV2で、定義ずみアーキテクチャの利用が可能なのですが, CIFAR-10, CIFAR-100の画像データは一片が32 pixelと非常に小さく、一辺が224 pixelで構成されるImageNet用に書かれている原論文のモデルでは, うまく学習ができません. そのため, Githubの実装を参考にして, アーキテクチャを作りました.
- Applications - Keras Documentation
- GitHub - keras-team/keras-applications: Reference implementations of popular deep learning models.
import os import warnings import numpy as np from keras.layers import Input, Activation, Conv2D, Dense, Dropout, BatchNormalization, ReLU, DepthwiseConv2D, GlobalAveragePooling2D, GlobalMaxPooling2D, Add from keras.models import Model from keras import regularizers # define the filter size def _make_divisible(v, divisor, min_value=None): if min_value is None: min_value = divisor new_v = max(min_value, int(v + divisor / 2) // divisor * divisor) # Make sure that round down does not go down by more than 10%. if new_v < 0.9 * v: new_v += divisor return new_v # define the calcuration of each 'Res_Block' def _inverted_res_block(inputs, expansion, stride, alpha, filters, block_id): prefix = 'block_{}_'.format(block_id) in_channels = inputs._keras_shape[-1] pointwise_conv_filters = int(filters * alpha) pointwise_filters = _make_divisible(pointwise_conv_filters, 8) x = inputs # Expand if block_id: x = Conv2D(expansion * in_channels, kernel_size=1, strides=1, padding='same', use_bias=False, activation=None, kernel_initializer="he_normal", kernel_regularizer=regularizers.l2(4e-5), name=prefix + 'expand')(x) x = BatchNormalization(epsilon=1e-3, momentum=0.999, name=prefix + 'expand_BN')(x) x = ReLU(6., name=prefix + 'expand_relu')(x) else: prefix = 'expanded_conv_' # Depthwise x = DepthwiseConv2D(kernel_size=3, strides=stride, activation=None, use_bias=False, padding='same', kernel_initializer="he_normal", depthwise_regularizer=regularizers.l2(4e-5), name=prefix + 'depthwise')(x) x = BatchNormalization(epsilon=1e-3, momentum=0.999, name=prefix + 'depthwise_BN')(x) x = ReLU(6., name=prefix + 'depthwise_relu')(x) # Project x = Conv2D(pointwise_filters, kernel_size=1, strides=1, padding='same', use_bias=False, activation=None, kernel_initializer="he_normal", kernel_regularizer=regularizers.l2(4e-5), name=prefix + 'project')(x) x = BatchNormalization(epsilon=1e-3, momentum=0.999, name=prefix + 'project_BN')(x) if in_channels == pointwise_filters and stride == 1: return Add(name=prefix + 'add')([inputs, x]) return x # build MobileNetV2 models def MobileNetV2(input_shape=(32, 32, 3), alpha=1.0, depth_multiplier=1, include_top=True, pooling=None, classes=10): # fileter size (first block) first_block_filters = _make_divisible(32 * alpha, 8) # input shape (first block) img_input = Input(shape=input_shape) # model architechture x = Conv2D(first_block_filters, kernel_size=3, strides=1, padding='same', use_bias=False, kernel_initializer="he_normal", kernel_regularizer=regularizers.l2(4e-5), name='Conv1')(img_input) #x = BatchNormalization(epsilon=1e-3, momentum=0.999, name='bn_Conv1')(x) #x = ReLU(6., name='Conv1_relu')(x) x = _inverted_res_block(x, filters=16, alpha=alpha, stride=1, expansion=1, block_id=0 ) x = _inverted_res_block(x, filters=24, alpha=alpha, stride=1, expansion=6, block_id=1 ) x = _inverted_res_block(x, filters=24, alpha=alpha, stride=1, expansion=6, block_id=2 ) x = _inverted_res_block(x, filters=32, alpha=alpha, stride=2, expansion=6, block_id=3 ) x = _inverted_res_block(x, filters=32, alpha=alpha, stride=1, expansion=6, block_id=4 ) x = _inverted_res_block(x, filters=32, alpha=alpha, stride=1, expansion=6, block_id=5 ) x = _inverted_res_block(x, filters=64, alpha=alpha, stride=2, expansion=6, block_id=6 ) x = _inverted_res_block(x, filters=64, alpha=alpha, stride=1, expansion=6, block_id=7 ) x = _inverted_res_block(x, filters=64, alpha=alpha, stride=1, expansion=6, block_id=8 ) x = _inverted_res_block(x, filters=64, alpha=alpha, stride=1, expansion=6, block_id=9 ) x = _inverted_res_block(x, filters=96, alpha=alpha, stride=1, expansion=6, block_id=10) x = _inverted_res_block(x, filters=96, alpha=alpha, stride=1, expansion=6, block_id=11) x = _inverted_res_block(x, filters=96, alpha=alpha, stride=1, expansion=6, block_id=12) x = _inverted_res_block(x, filters=160, alpha=alpha, stride=2, expansion=6, block_id=13) x = _inverted_res_block(x, filters=160, alpha=alpha, stride=1, expansion=6, block_id=14) x = _inverted_res_block(x, filters=160, alpha=alpha, stride=1, expansion=6, block_id=15) x = Dropout(rate=0.25)(x) x = _inverted_res_block(x, filters=320, alpha=alpha, stride=1, expansion=6, block_id=16) x = Dropout(rate=0.25)(x) # define fileter size (last block) if alpha > 1.0: last_block_filters = _make_divisible(1280 * alpha, 8) else: last_block_filters = 1280 x = Conv2D(last_block_filters, kernel_size=1, use_bias=False, kernel_initializer="he_normal", kernel_regularizer=regularizers.l2(4e-5), name='Conv_1')(x) x = BatchNormalization(epsilon=1e-3, momentum=0.999, name='Conv_1_bn')(x) x = ReLU(6., name='out_relu')(x) # top layer ("use" or "not use" FC) if include_top: x = GlobalAveragePooling2D(name='global_average_pool')(x) x = Dense(classes, activation='softmax', use_bias=True, name='Logits')(x) else: if pooling == 'avg': x = GlobalAveragePooling2D()(x) elif pooling == 'max': x = GlobalMaxPooling2D()(x) # create model of MobileNetV2 (for CIFAR-10) model = Model(inputs=img_input, outputs=x, name='mobilenetv2_cifar10') return model
main.py
各ファイルを統合して, 学習を実行するメインファイルです. このファイルによる実行がトリガーとなって, 各処理が進みます.
import os import h5py import numpy as np import keras.backend.tensorflow_backend as KTF import tensorflow as tf from keras.models import Sequential, Model from keras.applications.mobilenetv2 import MobileNetV2 from keras.engine.topology import Input from keras.layers.core import Dense, Dropout, Activation, Flatten from keras.layers.normalization import BatchNormalization from keras.layers import Conv2D, SeparableConv2D, MaxPooling2D from keras.optimizers import SGD, Adam, RMSprop from keras.preprocessing.image import ImageDataGenerator from keras.callbacks import EarlyStopping, ModelCheckpoint, TensorBoard, CSVLogger, LearningRateScheduler from datetime import datetime from datafeeders.datafeeder import DataFeeder from app.mobilenet_v2 import MobileNetV2 # set meta params batch_size = 128 nb_classes = 10 nb_epoch = 300 nb_data = 32*32 # set meta info log_dir = '../train_log/mobilenet_v2-like5_log' dataset_dir = '../../CIFAR-10/datasets/dataset_norm' model_name = 'mobilenet_v2-like5__' + datetime.now().strftime("%Y-%m-%d-%H-%M-%S") model_arch_path = os.path.join(log_dir, (model_name + '_arch.png')) model_cp_path = os.path.join(log_dir, (model_name + '_checkpoint.h5')) model_csv_path = os.path.join(log_dir, (model_name + '_csv.csv')) # load data DF = DataFeeder(dataset_dir) X_train, y_train, X_test, y_test = DF.X_train, DF.y_train, DF.X_test, DF.y_test # data augumatation datagen = ImageDataGenerator( width_shift_range=0.2, height_shift_range=0.2, vertical_flip=False, horizontal_flip=True ) datagen.fit(X_train) # build model model = MobileNetV2(input_shape=X_train.shape[1:], include_top=True, alpha=1.0) model.summary() print('Model Name: ', model_name) # save model architechture plot (.png) from keras.utils import plot_model plot_model(model, to_file=model_arch_path, show_shapes=True) # set learning rate learning_rates=[] for i in range(5): learning_rates.append(2e-2) for i in range(50-5): learning_rates.append(1e-2) for i in range(100-50): learning_rates.append(8e-3) for i in range(150-100): learning_rates.append(4e-3) for i in range(200-150): learning_rates.append(2e-3) for i in range(300-200): learning_rates.append(1e-3) # set callbacks callbacks = [] callbacks.append(TensorBoard(log_dir=log_dir, histogram_freq=1)) callbacks.append(ModelCheckpoint(model_cp_path, monitor='val_loss', save_best_only=True)) callbacks.append(LearningRateScheduler(lambda epoch: float(learning_rates[epoch]))) callbacks.append(CSVLogger(model_csv_path)) # compile & learning model model.compile(loss='categorical_crossentropy', optimizer=SGD(lr=2e-2, momentum=0.9, decay=0.0, nesterov=False), metrics=['accuracy']) history = model.fit_generator( datagen.flow(X_train, y_train, batch_size=batch_size), steps_per_epoch=len(X_train) / batch_size, epochs=nb_epoch, verbose=1, callbacks=callbacks, validation_data=(X_test, y_test)) # validation val_loss, val_acc = model.evaluate(X_test, y_test, verbose=0) print('Model Name: ', model_name) print('Test loss : {:.5f}'.format(val_loss)) print('Test accuracy : {:.5f}'.format(val_acc)) # save model "INSTANCE" f1_name = model_name + '_instance' f1_path = os.path.join(log_dir, f1_name) + '.h5' model.save(f1_path) # save model "WEIGHTs" f2_name = model_name + '_weights' f2_path = os.path.join(log_dir, f2_name) + '.h5' model.save_weights(f2_path) # save model "ARCHITECHTURE" f3_name = model_name + '_architechture' f3_path = os.path.join(log_dir, f3_name) + '.json' json_string = model.to_json() with open(f3_path, 'w') as f: f.write(json_string) # save plot figure from utils.plot_log import save_plot_log save_plot_log(log_dir, model_csv_path, index='acc') save_plot_log(log_dir, model_csv_path, index='loss')
3. 学習結果
最終的な Validation Accuracyは 91% くらいでした。パラメータ数が250万弱なのでなかなかの性能。学習率のスケジューリングによって精度が±5％動くので、難しいです。
※lossグラフの縦軸に「crossentropy」とありますが、多変量なので正しくは「categorical crossentropy」です。さーせん。
