From a58b1cc7b1f973e21a38bbbe045b5cf91753fafa Mon Sep 17 00:00:00 2001 From: 13o-bbr-bbq Date: Tue, 2 May 2017 06:22:41 +0900 Subject: [PATCH] Add files via upload --- CNN_test/finetuning.py | 95 +++++++++++++++++++++++++++ CNN_test/predict.py | 58 ++++++++++++++++ CNN_test/search_adv.py | 146 +++++++++++++++++++++++++++++++++++++++++ 3 files changed, 299 insertions(+) create mode 100644 CNN_test/finetuning.py create mode 100644 CNN_test/predict.py create mode 100644 CNN_test/search_adv.py diff --git a/CNN_test/finetuning.py b/CNN_test/finetuning.py new file mode 100644 index 00000000..a0ed3639 --- /dev/null +++ b/CNN_test/finetuning.py @@ -0,0 +1,95 @@ +# -*- coding: utf-8 -*- +import os +from keras.applications.vgg16 import VGG16 +from keras.layers import Input, Dropout, Flatten, Dense +from keras.preprocessing.image import ImageDataGenerator +from keras.models import Sequential, Model +from keras import optimizers + + +# CIFAR10のクラス一覧 +classes = ['airplane', 'automobile', 'bird', 'cat', 'deer', + 'dog', 'frog', 'horse', 'ship', 'truck'] + +batch_size = 32 +nb_classes = len(classes) + +# CIFAR10の画像情報(x:32pic, y:32pic, channel:3) +img_rows, img_cols = 32, 32 +channels = 3 + +# 訓練用画像とテスト用画像のPath +train_data_dir = '.\\cifar10\\train_image' +validation_data_dir = '.\\cifar10\\test_image' + +# 訓練用画像(各クラス5000枚)、テスト用画像(各クラス1000枚) +# epochは30くらいでも十分か？ +nb_train_samples = 50000 +nb_val_samples = 10000 +nb_epoch = 50 + +# 学習済み重みの保存Path +result_dir = '.\\cifar10\\results' +if not os.path.exists(result_dir): + os.mkdir(result_dir) + + +if __name__ == '__main__': + # VGG16モデルと学習済み重み(ImageNet)をロード + input_tensor = Input(shape=(img_rows, img_cols, 3)) + vgg16 = VGG16(include_top=False, weights='imagenet', input_tensor=input_tensor) + + # FC + top_model = Sequential() + top_model.add(Flatten(input_shape=vgg16.output_shape[1:])) + top_model.add(Dense(256, activation='relu')) + top_model.add(Dropout(0.5)) + top_model.add(Dense(nb_classes, activation='softmax')) + + # VGG16とFCを接続 + model = Model(input=vgg16.input, output=top_model(vgg16.output)) + + # 最後の畳み込み層の手前までFreeze + for layer in model.layers[:15]: + layer.trainable = False + + # 多クラス(=10)分類なのでloss=categorical_crossentropyとする + model.compile(loss='categorical_crossentropy', + optimizer=optimizers.SGD(lr=1e-4, momentum=0.9), + metrics=['accuracy']) + + train_datagen = ImageDataGenerator( + rescale=1.0 / 255, + shear_range=0.2, + zoom_range=0.2, + horizontal_flip=True) + + test_datagen = ImageDataGenerator(rescale=1.0 / 255) + + train_generator = train_datagen.flow_from_directory( + train_data_dir, + target_size=(img_rows, img_cols), + color_mode='rgb', + classes=classes, + class_mode='categorical', + batch_size=batch_size, + shuffle=True) + + validation_generator = test_datagen.flow_from_directory( + validation_data_dir, + target_size=(img_rows, img_cols), + color_mode='rgb', + classes=classes, + class_mode='categorical', + batch_size=batch_size, + shuffle=True) + + # Fine-tuning + history = model.fit_generator( + train_generator, + samples_per_epoch=nb_train_samples, + nb_epoch=nb_epoch, + validation_data=validation_generator, + nb_val_samples=nb_val_samples) + + model.save_weights(os.path.join(result_dir, 'finetuning.h5')) diff --git a/CNN_test/predict.py b/CNN_test/predict.py new file mode 100644 index 00000000..d34a9e0c --- /dev/null +++ b/CNN_test/predict.py @@ -0,0 +1,58 @@ +# -*- coding: utf-8 -*- +import os +import sys +import numpy as np +from keras.applications.vgg16 import VGG16 +from keras.models import Sequential, Model +from keras.layers import Input, Dropout, Flatten, Dense +from keras.preprocessing import image + +# CIFAR10のクラス一覧 +classes = ['airplane', 'automobile', 'bird', 'cat', 'deer', + 'dog', 'frog', 'horse', 'ship', 'truck'] +nb_classes = len(classes) + +# CIFAR10の画像情報(x:32pic, y:32pic, channel:3) +img_height, img_width = 32, 32 +channels = 3 + +# VGG16 +input_tensor = Input(shape=(img_height, img_width, channels)) +vgg16 = VGG16(include_top=False, weights='imagenet', input_tensor=input_tensor) + +# FC +fc = Sequential() +fc.add(Flatten(input_shape=vgg16.output_shape[1:])) +fc.add(Dense(256, activation='relu')) +fc.add(Dropout(0.5)) +fc.add(Dense(nb_classes, activation='softmax')) + +# VGG16とFCを接続 +model = Model(input=vgg16.input, output=fc(vgg16.output)) + +# 学習済みの重み(finetuning.pyでFine-tuningした重み)をロード +model.load_weights(os.path.join('cifar10\\results', 'finetuning.h5')) +# model.load_weights(os.path.join('cifar10\\results', 'finetuning_noise.h5')) + +# 多クラス(=10)分類なのでloss=categorical_crossentropyとする +model.compile(loss='categorical_crossentropy', + optimizer='adam', + metrics=['accuracy']) + +if sys.argv[1] == '': + print("usage: Test file is not found.") + sys.exit(1) + +# 画像を4次元テンソルに変換 +img = image.load_img(sys.argv[1], target_size=(img_height, img_width)) +x = image.img_to_array(img) +x = np.expand_dims(x, axis=0) +x = x / 255.0 + +# クラスの予測(top3を出力) +pred = model.predict(x)[0] +top = 3 +top_indices = pred.argsort()[-top:][::-1] +result = [(classes[i], pred[i]) for i in top_indices] +for x in result: + print(x) diff --git a/CNN_test/search_adv.py b/CNN_test/search_adv.py new file mode 100644 index 00000000..b3dc7690 --- /dev/null +++ b/CNN_test/search_adv.py @@ -0,0 +1,146 @@ +# -*- coding: utf-8 -*- +import os +import random +import datetime +import numpy as np +import cv2 +from progressbar import ProgressBar +from keras.applications.vgg16 import VGG16 +from keras.models import Sequential, Model +from keras.layers import Input, Dropout, Flatten, Dense +from keras.preprocessing import image + +# 検証対象のクラス +CLASS = 'cat' +TEST_DIR = 'cifar10\\test_image\\' + CLASS + '\\' +ADV_DIR = 'results\\' + CLASS + '\\' +TEST_IMAGES = 1000 + + +# ランダム選択したピクセルに細工を加える +def random_adv(perturbation_filename, p): + # 画像読み込み + img = cv2.imread(TEST_DIR + perturbation_filename, cv2.IMREAD_UNCHANGED) + + # 画像サイズの取得 + if len(img) == 3: + height, width, channel = img.shape[:3] + else: + height, width = img.shape[:2] + + for i in range(p): + # 細工するピクセルをランダムに選択 + x = random.randint(0, width - 1) + y = random.randint(0, height - 1) + + # ピクセル値の細工 + pixel = img[y, x] + average = sum(pixel) / len(pixel) + + if average < 128: + img[y, x] = [0, 0, 0] + else: + img[y, x] = [255, 255, 255] + + # 細工画像の保存 + adv_filename = 'adv_' + filename + cv2.imwrite(ADV_DIR + adv_filename, img) + + return adv_filename + + +# クラスの予測 +def predict(target_filename, height, width): + # 画像を4次元テンソルに変換 + img = image.load_img(target_filename, target_size=(height, width)) + x = image.img_to_array(img) + x = np.expand_dims(x, axis=0) + x = x / 255.0 + + # クラスの予測 + pred = model.predict(x)[0] + top = 1 + top_indices = pred.argsort()[-top:][::-1] + result = [(classes[i], pred[i]) for i in top_indices] + + ''' + result[0][0] : predict label + result[0][1] : prediction probability + ''' + return result + +if __name__ == "__main__": + # CIFAR10のクラス一覧 + classes = ['airplane', 'automobile', 'bird', 'cat', 'deer', + 'dog', 'frog', 'horse', 'ship', 'truck'] + nb_classes = len(classes) + + # CIFAR10の画像情報(x:32pixel, y:32pixel, channel:3) + img_height, img_width = 32, 32 + channels = 3 + + # VGG16 + input_tensor = Input(shape=(img_height, img_width, channels)) + vgg16 = VGG16(include_top=False, weights='imagenet', input_tensor=input_tensor) + + # FC + fc = Sequential() + fc.add(Flatten(input_shape=vgg16.output_shape[1:])) + fc.add(Dense(256, activation='relu')) + fc.add(Dropout(0.5)) + fc.add(Dense(nb_classes, activation='softmax')) + + # VGG16とFCを接続 + model = Model(input=vgg16.input, output=fc(vgg16.output)) + + # 学習済みの重み(finetuning.pyでFine-tuningした重み)をロード + model.load_weights(os.path.join('cifar10\\results', 'finetuning.h5')) + # model.load_weights(os.path.join('cifar10\\results', 'finetuning_noise.h5')) + + # 多クラス(=10)分類なのでloss=categorical_crossentropyとする + model.compile(loss='categorical_crossentropy', + optimizer='adam', + metrics=['accuracy']) + + # 細工するピクセル数(p)、1画像たりの細工試行最大回数(p_max_num) + p = 1 + p_max_num = 100 + + # 検証対象の画像一覧 + test_files = os.listdir(TEST_DIR) + + # 検証結果保存用のファイル + today_detail = datetime.datetime.today() + test_result = str(today_detail.strftime("%Y%m%d_%H%M%S")) + '.txt' + f = open(test_result, 'w') + f.write('Idx\tnormal image\tadversarial image\tmisclassify\tpredict\n') + + # 細工処理 + count = 0 + progress_count = 1 + progress = ProgressBar(min_value=1, max_value=TEST_IMAGES) + for filename in test_files: + # 進捗表示 + progress.update(progress_count) + progress_count += 1 + + # 細工前画像のクラス分類予測 + result = predict(TEST_DIR + filename, img_height, img_width) + + # 正しく分類できない画像は検証しない + if CLASS != result[0][0]: + continue + + for i in range(p_max_num): + p_filename = random_adv(filename, p) + + # 細工画像のクラス分類予測 + result = predict(ADV_DIR + p_filename, img_height, img_width) + + # 誤分類した場合はログに書き込んで次の画像の検証を行う + if CLASS != result[0][0]: + count += 1 + f.write(str(count)+'\t'+filename+'\t'+p_filename+'\t'+result[0][0]+'\t'+str(result[0][1])+'\n') + break + print('\n') + f.close()