def model_z(nb_filters=32, nb_classes=10, input_shape=(None, 28, 28, 1)):
layers = [
Conv2D(nb_filters, (3, 3), (1, 1), "SAME"),
ReLU(),
Conv2D(nb_filters, (3, 3), (2, 2), "VALID"),
ReLU(),
Conv2D(2 * nb_filters, (3, 3), (1, 1), "VALID"),
ReLU(),
Conv2D(2 * nb_filters, (3, 3), (2, 2), "VALID"),
ReLU(),
Conv2D(4 * nb_filters, (3, 3), (1, 1), "VALID"),
ReLU(),
Conv2D(4 * nb_filters, (3, 3), (2, 2), "VALID"),
ReLU(),
Flatten(),
Linear(600),
ReLU(),
Dropout(0.5),
Linear(600),
ReLU(),
Dropout(0.5),
Linear(nb_classes),
Softmax()
]
model = DefenseMLP(layers, input_shape)
return model
def model_c(nb_filters=64, nb_classes=10, input_shape=(None, 28, 28, 1)):
layers = [
Conv2D(nb_filters * 2, (3, 3), (1, 1), "SAME", use_bias=True),
ReLU(),
Conv2D(nb_filters, (5, 5), (2, 2), "VALID", use_bias=True),
ReLU(),
Flatten(),
Dropout(0.25),
Linear(128),
ReLU(),
Dropout(0.5),
Linear(nb_classes),
Softmax()
]
model = DefenseMLP(layers, input_shape)
return model
def test_no_drop():
"""test_no_drop: Make sure dropout does nothing by default (so it does not
cause stochasticity at test time)"""
model = MLP(input_shape=[1, 1], layers=[Dropout(name='output')])
x = tf.constant([[1]], dtype=tf.float32)
y = model.get_layer(x, 'output')
sess = tf.Session()
# Do multiple runs because dropout is stochastic
for _ in range(10):
y_value = sess.run(y)
assert y_value == 1.
def model_y(nb_filters=64, nb_classes=10, input_shape=(None, 28, 28, 1)):
layers = [
Conv2D(nb_filters, (3, 3), (1, 1), "SAME"),
ReLU(),
Conv2D(nb_filters, (3, 3), (2, 2), "VALID"),
ReLU(),
Conv2D(2 * nb_filters, (3, 3), (2, 2), "VALID"),
ReLU(),
Conv2D(2 * nb_filters, (3, 3), (2, 2), "VALID"),
ReLU(),
Flatten(),
Linear(256),
ReLU(),
Dropout(0.5),
Linear(256),
ReLU(),
Dropout(0.5),
Linear(nb_classes),
Softmax()
]
model = DefenseMLP(layers, input_shape, feature_layer='ReLU13')
return model
def test_drop():
"""test_drop: Make sure dropout is activated successfully"""
# We would like to configure the test to deterministically drop,
# so that the test does not need to use multiple runs.
# However, tf.nn.dropout divides by include_prob, so zero or
# infinitesimal include_prob causes NaNs.
# 1e-8 does not cause NaNs and shouldn't be a significant source
# of test flakiness relative to dependency downloads failing, etc.
model = MLP(input_shape=[1, 1],
layers=[Dropout(name='output', include_prob=1e-8)])
x = tf.constant([[1]], dtype=tf.float32)
y = model.get_layer(x, 'output', dropout=True)
sess = tf.Session()
y_value = sess.run(y)
# Subject to very rare random failure because include_prob is not exact 0
assert y_value == 0., y_value
def test_override():
# Make sure dropout_dict changes dropout probabilities successful
# We would like to configure the test to deterministically drop,
# so that the test does not need to use multiple runs.
# However, tf.nn.dropout divides by include_prob, so zero or
# infinitesimal include_prob causes NaNs.
# For this test, random failure to drop will not cause the test to fail.
# The stochastic version should not even run if everything is working
# right.
model = MLP(input_shape=[1, 1], layers=[Dropout(name='output',
include_prob=1e-8)])
x = tf.constant([[1]], dtype=tf.float32)
dropout_dict = {'output': 1.}
y = model.get_layer(x, 'output', dropout=True, dropout_dict=dropout_dict)
sess = tf.Session()
y_value = sess.run(y)
assert y_value == 1., y_value
def model_d(input_shape=(None, 28, 28, 1), nb_classes=10):
"""
Defines the model architecture to be used by the substitute. Use
the example model interface.
:param img_rows: number of rows in input
:param img_cols: number of columns in input
:param nb_classes: number of classes in output
:return: tensorflow model
"""
# Define a fully connected model (it's different than the black-box)
layers = [
Flatten(),
Linear(200),
ReLU(),
Dropout(0.5),
Linear(200),
ReLU(),
Linear(nb_classes),
Softmax()
]
return DefenseMLP(layers, input_shape)
