def test_calc_avgpool(self):
image_data = self._image_data()
net = self._network('avgpool')
input_bounds = naive_bounds.input_bounds(image_data.image, delta=.1)
dual_obj, dual_var_lists = self._build_objective(
net, input_bounds, image_data.label)
# Explicitly build the expected TensorFlow graph for calculating objective.
(
conv2d_0,
relu_1, # pylint:disable=unused-variable
avgpool_2,
relu_3, # pylint:disable=unused-variable
linear_obj) = self._verifiable_layer_builder(net).build_layers()
(mu_0, ), (lam_1, ), (mu_2, ), _ = dual_var_lists
# Expected input bounds for each layer.
conv2d_0_lb, conv2d_0_ub = self._expected_input_bounds(
image_data.image, .1)
relu_1_lb, relu_1_ub = ibp.IntervalBounds(
conv2d_0_lb, conv2d_0_ub).apply_conv2d(None, conv2d_0.module.w,
conv2d_0.module.b, 'SAME',
(1, 1))
avgpool_2_lb = tf.nn.relu(relu_1_lb)
avgpool_2_ub = tf.nn.relu(relu_1_ub)
relu_3_lb = tf.nn.avg_pool(avgpool_2_lb,
ksize=[2, 2],
padding='VALID',
strides=(1, 1))
relu_3_ub = tf.nn.avg_pool(avgpool_2_ub,
ksize=[2, 2],
padding='VALID',
strides=(1, 1))
# Expected objective value.
objective = 0
act_coeffs_0 = -common.conv_transpose(
mu_0, conv2d_0.module.w, conv2d_0.input_shape, 'SAME', (1, 1))
obj_0 = -tf.reduce_sum(mu_0 * conv2d_0.module.b, axis=(2, 3, 4))
objective += standard_layer_calcs.linear_dual_objective(
None, act_coeffs_0, obj_0, conv2d_0_lb, conv2d_0_ub)
objective += standard_layer_calcs.activation_layer_dual_objective(
tf.nn.relu, mu_0, lam_1, relu_1_lb, relu_1_ub)
act_coeffs_2 = -common.avgpool_transpose(
mu_2,
result_shape=relu_1.output_shape,
kernel_shape=(2, 2),
strides=(1, 1))
objective += standard_layer_calcs.linear_dual_objective(
lam_1, act_coeffs_2, 0., avgpool_2_lb, avgpool_2_ub)
objective_w, objective_b = common.targeted_objective(
linear_obj.module.w, linear_obj.module.b, image_data.label)
shaped_objective_w = tf.reshape(
objective_w,
[self._num_classes(), self._batch_size()] + avgpool_2.output_shape)
objective += standard_layer_calcs.activation_layer_dual_objective(
tf.nn.relu, mu_2, -shaped_objective_w, relu_3_lb, relu_3_ub)
objective += objective_b
self._assert_dual_objective_close(objective, dual_obj, image_data)
def backward_prop(self, y, w_fn=None):
del w_fn
return common.avgpool_transpose(y,
result_shape=self.input_shape,
kernel_shape=self.kernel_shape,
strides=self.strides)