def test_tf_keras_exception():
bounds = (0, 255)
def mean_brightness_net(images):
logits = tf.reduce_mean(images, axis=(1, 2))
return logits
model = mean_brightness_net
with pytest.raises(ValueError):
TensorFlowModel.from_keras(model, bounds=bounds)
TensorFlowModel.from_keras(model, bounds=bounds, input_shape=(5, 5, 3))
def __init__(self, kmodel, method, criterion):
'''
Adversarial Trainingを行うクラス
params
:kmodel: Kerasモデル（学習済み）
:method: foolboxのbatch_attacksモジュール
'''
self.kmodel = kmodel
self.fmodel = TensorFlowModel.from_keras(kmodel, bounds=(0, 1))
self.criterion = criterion
self.method_base = method
def test_tf_keras_constructor():
bounds = (0, 255)
def create_model():
data_format = "channels_last"
input_shape = [28, 28, 1]
l = tf.keras.layers # noqa: E741
max_pool = l.MaxPooling2D((2, 2), (2, 2),
padding="same",
data_format=data_format)
return tf.keras.Sequential([
l.Conv2D(
32,
5,
padding="same",
data_format=data_format,
input_shape=input_shape,
activation=tf.nn.relu,
),
max_pool,
l.Conv2D(
64,
5,
padding="same",
data_format=data_format,
activation=tf.nn.relu,
),
max_pool,
l.Flatten(),
l.Dense(1024, activation=tf.nn.relu),
l.Dropout(0.4),
l.Dense(10),
])
model = create_model()
fmodel = TensorFlowModel.from_keras(model, bounds=bounds)
assert fmodel.num_classes() == 10
fmodel.session.run(tf.global_variables_initializer())
test_images = np.random.rand(2, 28, 28, 1).astype(np.float32)
assert fmodel.forward(test_images).shape == (2, 10)
def fit(self,
x,
y,
epochs,
validation_data=None,
batch_size=32,
verbose=1):
'''
params
:x: オリジナルデータ
:y: オリジナルラベル
:i: イテレーション数
'''
self.x = x
self.y = y
for i in range(epochs):
print(f'\n{i+1} / {epochs} epochs')
# rand_idx = np.random.permutation(np.arange(self.x.shape[0])) # 全データのAdversarial Examplesをランダムな順番で作成
rand_idx = np.random.choice(
np.arange(self.x.shape[0]),
size=batch_size) # 全データ作成している時間がないので、ランダムな10サンプル分を作成
self.fmodel = TensorFlowModel.from_keras(self.kmodel,
bounds=(0, 1))
self.method = self.method_base(self.fmodel,
criterion=self.criterion)
self.adv_imgs, self.orig_cls = [], []
for r_idx in rand_idx:
x_adv = self.method(self.x[r_idx], self.y[r_idx], unpack=False)
if x_adv.image is None:
continue
self.adv_imgs.append(x_adv.image)
self.orig_cls.append(x_adv.original_class)
self.x = np.append(self.x, np.array(self.adv_imgs), axis=0)
self.y = np.append(self.y, self.orig_cls)
self.kmodel.fit(x=self.x,
y=self.y,
epochs=1,
validation_data=validation_data,
verbose=verbose)