y_test_tf = tf.placeholder(tf.float32, shape=(1, nclass)) sigma_tf = tf.placeholder(tf.float32, shape=(1)) input_shape = X_test.shape[1:] n = 3 version = 1 if version == 1: depth = n * 6 + 2 elif version == 2: depth = n * 9 + 2 # Model parameters with tf.variable_scope('test_model', reuse=False): if version == 1: model = resnet_v1(input_shape=input_shape, depth=depth) elif version == 2: model = resnet_v2(input_shape=input_shape, depth=depth) var_cls = model.trainable_weights saver_model = tf.train.Saver(var_cls, max_to_keep=None) x_test_tf_reshaped = tf.reshape(x_test_tf, [-1, height * width * nch]) repeated_x_test_tf = tf.tile(x_test_tf_reshaped, [1, batch_size]) repeated_x_test_tf = tf.reshape(repeated_x_test_tf, [-1, height * width * nch]) repeated_x_test_tf = tf.reshape(repeated_x_test_tf, [-1, height, width, nch]) noise = tf.random.normal(repeated_x_test_tf.shape) * sigma_tf noisy_inputs = repeated_x_test_tf + noise cls_test = KerasModelWrapper(model).get_logits(noisy_inputs)
x_poisoned_tf = tf.placeholder(tf.float32, shape=(None, height, width, nch)) y_poisoned_tf = tf.placeholder(tf.float32, shape=(None, nclass)) x_original_tf = tf.placeholder(tf.float32, shape=(None, height, width, nch)) input_shape = X_test.shape[1:] n = 3 version = 1 if version == 1: depth = n * 6 + 2 elif version == 2: depth = n * 9 + 2 # Model parameters with tf.variable_scope('test_model', reuse=False): test_model = resnet_v1(input_shape=input_shape, depth=depth) var_test = test_model.trainable_weights saver_model_test = tf.train.Saver(var_test, max_to_keep=None) with tf.variable_scope('train_model', reuse=False): train_model = resnet_v1(input_shape=input_shape, depth=depth) var_train = train_model.trainable_weights saver_model_train = tf.train.Saver(var_train, max_to_keep=None) bl_poisoning = bilevel_poisoning( sess, x_train_tf, x_val_tf, x_test_tf, x_poisoned_tf, x_original_tf, y_train_tf, y_val_tf, y_val_class_tf, y_test_tf, y_poisoned_tf, batch_size_poisoned, height, width, nch, nclass, val_batch_size, batch_size_clean, k_macer, sigma_macer, beta_macer, train_model,