def add_layer(net, module, inputs, flatten=False, batch_norm=None):
if flatten:
reshape_module = snt.BatchFlatten()
outputs = reshape_module(inputs)
net.append(
ibp.BatchReshapeWrapper(reshape_module,
outputs.shape[1:].as_list()), outputs,
inputs)
inputs = outputs
outputs = module(inputs)
if isinstance(module, AvgPool):
module.__name__ = 'avg_pool'
parameters = {
'ksize': [1] + module.kernel_shape + [1],
'padding': module.padding,
'strides': [1] + module.strides + [1]
}
net.append(ibp.IncreasingMonotonicWrapper(module, **parameters),
outputs, inputs)
elif isinstance(module, snt.Conv2D):
net.append(ibp.LinearConv2dWrapper(module), outputs, inputs)
elif isinstance(module, snt.Conv1D):
net.append(ibp.LinearConv1dWrapper(module), outputs, inputs)
elif isinstance(module, snt.Linear):
net.append(ibp.LinearFCWrapper(module), outputs, inputs)
else:
net.append(ibp.IncreasingMonotonicWrapper(module), outputs, inputs)
if batch_norm is not None:
inputs = outputs
outputs = batch_norm(inputs, is_training=False, test_local_stats=False)
net.append(ibp.BatchNormWrapper(batch_norm), outputs, inputs)
return outputs
def testReluIntervalBounds(self):
m = tf.nn.relu
z = tf.constant([[-2, 3]], dtype=tf.float32)
m = ibp.IncreasingMonotonicWrapper(m)
input_bounds = ibp.IntervalBounds(z - 1., z + 1.)
output_bounds = m.propagate_bounds(input_bounds)
with self.test_session() as sess:
l, u = sess.run([output_bounds.lower, output_bounds.upper])
self.assertAlmostEqual([[0., 2.]], l.tolist())
self.assertAlmostEqual([[0., 4.]], u.tolist())
def testReluBackwardBounds(self):
m = tf.nn.relu
z = tf.constant([[-2, 3]], dtype=tf.float32)
m = ibp.IncreasingMonotonicWrapper(m)
input_bounds = ibp.IntervalBounds(z - 1., z + 1.)
m.propagate_bounds(input_bounds) # Create IBP bounds.
crown_init_bounds = _generate_identity_spec([m],
shape=(1, 2, 2),
dimension=2)
output_bounds = m.propagate_bounds(crown_init_bounds)
concrete_bounds = output_bounds.concretize()
with self.test_session() as sess:
l, u = sess.run([concrete_bounds.lower, concrete_bounds.upper])
self.assertAlmostEqual([[0., 2.]], l.tolist())
self.assertAlmostEqual([[0., 4.]], u.tolist())