def preprocess_data(X, Y, model, GPU_ID):
X = np.array(keras.applications.resnet.preprocess_input(X))
x_test_o = np.copy(X)
loss = keras.losses.CategoricalCrossentropy()
epsilon = .01 * 255
pert = .001 * 255
for i in range(int(len(X) / 100)):
Z = tf.cast(X[i * 100:(100 * i + 100)], tf.float32)
Y_next = Y[i * 100:(100 * i + 100)]
for j in range((int)(math.floor(epsilon / pert))):
with tf.GradientTape() as grad_tracker:
grad_tracker.watch(Z)
prediction = model(Z)
next_loss = loss(Y_next, prediction)
grad = grad_tracker.gradient(next_loss, Z)
signed_grad = tf.sign(grad)
Z = Z + pert * signed_grad
Z1 = tf.clip_by_value(Z[:, :, :, 0], -103.939, 151.061)
Z2 = tf.clip_by_value(Z[:, :, :, 1], -116.779, 138.221)
Z3 = tf.clip_by_value(Z[:, :, :, 2], -123.68, 131.32)
Z = tf.stack([Z1, Z2, Z3], axis=3)
for k in range(100):
X[100 * i + k, :, :, 0] = tf.math.add(Z[k, :, :, 0], 103.939) / 255
X[100 * i + k, :, :, 1] = tf.math.add(Z[k, :, :, 1], 116.779) / 255
X[100 * i + k, :, :, 2] = tf.math.add(Z[k, :, :, 2], 123.68) / 255
X_p = X
first, second, third, fourth = get_processed_data(X_p, img_size2, GPU_ID)
first[:, :, :, 0] = tf.math.add(first[:, :, :, 0] * 255, -103.939)
first[:, :, :, 1] = tf.math.add(first[:, :, :, 1] * 255, -116.779)
first[:, :, :, 2] = tf.math.add(first[:, :, :, 2] * 255, -123.68)
second[:, :, :, 0] = tf.math.add(second[:, :, :, 0] * 255, -103.939)
second[:, :, :, 1] = tf.math.add(second[:, :, :, 1] * 255, -116.779)
second[:, :, :, 2] = tf.math.add(second[:, :, :, 2] * 255, -123.68)
third[:, :, :, 0] = tf.math.add(third[:, :, :, 0] * 255, -103.939)
third[:, :, :, 1] = tf.math.add(third[:, :, :, 1] * 255, -116.779)
third[:, :, :, 2] = tf.math.add(third[:, :, :, 2] * 255, -123.68)
fourth[:, :, :, 0] = tf.math.add(fourth[:, :, :, 0] * 255, -103.939)
fourth[:, :, :, 1] = tf.math.add(fourth[:, :, :, 1] * 255, -116.779)
fourth[:, :, :, 2] = tf.math.add(fourth[:, :, :, 2] * 255, -123.68)
X[:, :, :, 0] = tf.math.add(X[:, :, :, 0] * 255, -103.939)
X[:, :, :, 1] = tf.math.add(X[:, :, :, 1] * 255, -116.779)
X[:, :, :, 2] = tf.math.add(X[:, :, :, 2] * 255, -123.68)
# encode to one-hot
return x_test_o, X, first, second, third, fourth
def preprocess_data(X, Y, model, epsilon, pert):
epsilon = epsilon * 255
pert = pert * 255
X = np.array(keras.applications.resnet.preprocess_input(X))
loss = keras.losses.CategoricalCrossentropy()
for i in range(int(len(X) / 100)):
Z = tf.cast(X[i * 100:(100 * i + 100)], tf.float32)
Y_next = Y[i * 100:(100 * i + 100)]
if (epsilon != 0):
for j in range((int)(math.floor(epsilon / pert))):
with tf.GradientTape() as grad_tracker:
grad_tracker.watch(Z)
prediction = model(Z)
next_loss = loss(Y_next, prediction)
grad = grad_tracker.gradient(next_loss, Z)
signed_grad = tf.sign(grad)
Z = Z + pert * signed_grad
Z1 = tf.clip_by_value(Z[:, :, :, 0], -103.939, 151.061)
Z2 = tf.clip_by_value(Z[:, :, :, 1], -116.779, 138.221)
Z3 = tf.clip_by_value(Z[:, :, :, 2], -123.68, 131.32)
Z = tf.stack([Z1, Z2, Z3], axis=3)
for k in range(100):
X[100 * i + k, :, :, 0] = tf.math.add(Z[k, :, :, 0], 103.939) / 255
X[100 * i + k, :, :, 1] = tf.math.add(Z[k, :, :, 1], 116.779) / 255
X[100 * i + k, :, :, 2] = tf.math.add(Z[k, :, :, 2], 123.68) / 255
X_p = X
X_p = get_processed_data(X_p, img_size2)
X_p[:, :, :, 0] = tf.math.add(X_p[:, :, :, 0] * 255, -103.939)
X_p[:, :, :, 1] = tf.math.add(X_p[:, :, :, 1] * 255, -116.779)
X_p[:, :, :, 2] = tf.math.add(X_p[:, :, :, 2] * 255, -123.68)
X[:, :, :, 0] = tf.math.add(X[:, :, :, 0] * 255, -103.939)
X[:, :, :, 1] = tf.math.add(X[:, :, :, 1] * 255, -116.779)
X[:, :, :, 2] = tf.math.add(X[:, :, :, 2] * 255, -123.68)
# encode to one-hot
return X_p, X