def test_one_point():
"""Point-wise test case with only one constraint.
This leads to weight bounds that are unbounded-on-one-side.
"""
# Where the weight is too big.
network = Network([
FullyConnectedLayer(np.eye(2), np.zeros(shape=(2,))),
ReluLayer(),
FullyConnectedLayer(np.array([[1.0, 0.0], [1.0, 0.0]]),
np.array([0.0, 0.0])),
])
layer_index = 0
points = [[1.0, 0.5]]
labels = [1]
patcher = NetPatcher(network, layer_index, points, labels)
patched = patcher.compute(steps=1)
assert np.argmax(patched.compute(points)) == 1
# Where the weight is too small.
network = Network([
FullyConnectedLayer(np.eye(2), np.array([0.0, 0.0])),
ReluLayer(),
FullyConnectedLayer(np.array([[1.0, 0.0], [1.0, 0.0]]), np.array([0.0, 2.0])),
])
layer_index = 0
points = [[1.0, 0.0]]
labels = [0]
patcher = NetPatcher(network, layer_index, points, labels)
patched = patcher.compute(steps=1)
assert np.argmax(patched.compute(points)) == 0
def test_from_planes():
"""Test case to load key points from a set of labeled 2D polytopes.
"""
if not Network.has_connection():
pytest.skip("No server connected.")
network = Network([
ReluLayer(),
FullyConnectedLayer(np.eye(2), np.zeros(shape=(2,)))
])
layer_index = 1
planes = [
np.array([[-1.0, -3.0], [-0.5, -3.0], [-0.5, 9.0], [-1.0, 9.0]]),
np.array([[8.0, -2.0], [16.0, -2.0], [16.0, 6.0], [8.0, 6.0]]),
]
labels = [1, 0]
patcher = NetPatcher.from_planes(network, layer_index, planes, labels)
assert patcher.network is network
assert patcher.layer_index is layer_index
assert len(patcher.inputs) == (4 + 2) + (4 + 2)
true_key_points = list(planes[0])
true_key_points += [np.array([-1.0, 0.0]), np.array([-0.5, 0.0])]
true_key_points += list(planes[1])
true_key_points += [np.array([8.0, 0.0]), np.array([16.0, 0.0])]
true_labels = ([1] * 6) + ([0] * 6)
for true_point, true_label in zip(true_key_points, true_labels):
try:
i = next(i for i, point in enumerate(patcher.inputs)
if np.allclose(point, true_point))
except StopIteration:
assert False
assert true_label == patcher.labels[i]
def test_compute_from_syrenn():
"""Tests the it works given an arbitrary network and planes.
"""
if not Network.has_connection():
pytest.skip("No server connected.")
network = Network([ReluLayer()])
planes = [np.array([[1.0, 2.0], [3.0, 4.0], [1.0, 2.0]]),
np.array([[3.0, 2.0], [7.0, 4.0], [5.0, 2.0]])]
transformed = network.transform_planes(planes, True, True)
classifier = PlanesClassifier.from_syrenn(transformed)
assert classifier.partially_computed
classified = classifier.compute()
assert len(classified) == len(planes)
regions, labels = classified[0]
assert np.allclose(regions, planes[0])
assert np.allclose(labels, [1])
regions, labels = classified[1]
assert np.allclose(regions, planes[1])
assert np.allclose(labels, [0])
# Ensure it doesn't re-compute things it already knows.
assert classifier.compute() is classified
def test_from_spec():
"""Test case to load key points from a spec function.
"""
if not Network.has_connection():
pytest.skip("No server connected.")
network = Network([
ReluLayer(),
FullyConnectedLayer(np.eye(2), np.zeros(shape=(2,)))
])
layer_index = 1
region_of_interest = np.array([
[0.5, -3.0],
[1.0, -3.0],
[1.0, 9.0],
[0.5, 9.0],
])
spec_fn = lambda i: np.isclose(i[:, 0], 1.0).astype(np.float32)
patcher = NetPatcher.from_spec_function(network, layer_index,
region_of_interest, spec_fn)
assert patcher.network is network
assert patcher.layer_index is layer_index
assert len(patcher.inputs) == (4 + 2)
true_key_points = list(region_of_interest)
true_key_points += [np.array([0.5, 0.0]), np.array([1.0, 0.0])]
true_labels = [0, 1, 1, 0, 0, 1]
for true_point, true_label in zip(true_key_points, true_labels):
try:
i = next(i for i, point in enumerate(patcher.inputs)
if np.allclose(point, true_point))
except StopIteration:
assert False
assert true_label == patcher.labels[i]
def test_compute_from_network():
"""Tests the it works given an arbitrary network and planes.
"""
if not Network.has_connection():
pytest.skip("No server connected.")
network = Network([ReluLayer()])
planes = [np.array([[1.0, 2.0], [3.0, 4.0], [1.0, 2.0]]),
np.array([[3.0, 2.0], [7.0, 4.0], [5.0, 2.0]])]
classifier = PlanesClassifier(network, planes, preimages=True)
classifier.partial_compute()
assert classifier.partially_computed
transformed = network.transform_planes(planes, True, True)
assert all(all(np.allclose(actual_polytope[0], truth_polytope[0]) and
np.allclose(actual_polytope[1], truth_polytope[1])
for actual_polytope, truth_polytope in zip(actual, truth))
for actual, truth in zip(classifier.transformed_planes,
transformed))
classified = classifier.compute()
assert len(classified) == len(planes)
regions, labels = classified[0]
assert np.allclose(regions, planes[0])
assert np.allclose(labels, [1])
regions, labels = classified[1]
assert np.allclose(regions, planes[1])
assert np.allclose(labels, [0])
# Ensure it doesn't re-compute things it already knows.
assert classifier.compute() is classified
def test_compute_from_network():
"""Tests the it works given an arbitrary network and lines.
"""
if not Network.has_connection():
pytest.skip("No server connected.")
network = Network([ReluLayer()])
lines = [(np.array([0.0, 1.0]), np.array([0.0, -1.0])),
(np.array([2.0, 3.0]), np.array([4.0, 3.0]))]
classifier = LinesClassifier(network, lines, preimages=True)
classifier.partial_compute()
exactlines = network.exactlines(lines, True, True)
assert all(
np.allclose(actual[0], truth[0]) and np.allclose(actual[1], truth[1])
for actual, truth in zip(classifier.transformed_lines, exactlines))
classified = classifier.compute()
assert len(classified) == len(lines)
regions, labels = classified[0]
assert np.allclose(regions,
[[[0.0, 1.0], [0.0, 0.0]], [[0.0, 0.0], [0.0, -1.0]]])
assert np.allclose(labels, [1, 0])
regions, labels = classified[1]
assert np.allclose(regions,
[[[2.0, 3.0], [3.0, 3.0]], [[3.0, 3.0], [4.0, 3.0]]])
assert np.allclose(labels, [1, 0])
# Ensure it doesn't re-compute things it already knows.
assert classifier.compute() is classified
def layer_jacobian(self, points, representatives):
"""Computes the Jacobian of the FULLY CONNECTED layer parameters.
Returns (n_points, n_outputs, [weight_shape])
Basically, we assume WLOG that the layer is the first layer then we get
something like for each point:
B_nB_{n-1}...B_1Ax
For each point we can collapse B_n...B_1 into a matrix B of shape
(n_outputs, n_A_outputs)
then we compute the einsum with x to get
(n_outputs, n_A_outputs, n_A_inputs)
which is what we want.
"""
pre_network = Network(self.network.layers[:self.layer_index])
points = pre_network.compute(points)
n_points, A_in_dims = points.shape
representatives = pre_network.compute(representatives)
representatives = self.network.layers[self.layer_index].compute(representatives)
n_points, A_out_dims = representatives.shape
post_network = Network(self.network.layers[(self.layer_index+1):])
# (n_points, A_out, A_out)
jacobian = np.repeat([np.eye(A_out_dims)], n_points, axis=0)
# (A_out, n_points, A_out)
jacobian = jacobian.transpose((1, 0, 2))
for layer in post_network.layers:
if isinstance(layer, LINEAR_LAYERS):
if isinstance(layer, ConcatLayer):
assert not any(isinstance(input_layer, ConcatLayer)
for input_layer in layer.input_layers)
assert all(isinstance(input_layer, LINEAR_LAYERS)
for input_layer in layer.input_layers)
representatives = layer.compute(representatives)
jacobian = jacobian.reshape((A_out_dims * n_points, -1))
jacobian = layer.compute(jacobian, jacobian=True)
jacobian = jacobian.reshape((A_out_dims, n_points, -1))
elif isinstance(layer, ReluLayer):
# (n_points, running_dims)
zero_indices = (representatives <= 0)
representatives[zero_indices] = 0.
# (A_out, n_points, running_dims)
jacobian[:, zero_indices] = 0.
elif isinstance(layer, HardTanhLayer):
big_indices = (representatives >= 1.)
small_indices = (representatives <= -1.)
np.clip(representatives, -1.0, 1.0, out=representatives)
jacobian[:, big_indices] = 0.
jacobian[:, small_indices] = 0.
else:
raise NotImplementedError
# (A_out, n_points, n_classes) -> (n_points, n_classes, A_out)
B = jacobian.transpose((1, 2, 0))
# (n_points, n_classes, n_A_in, n_A_out)
C = np.einsum("nco,ni->ncio", B, points)
return C, B
def test_compute_from_exactline_error():
"""Tests that it requires the plural exactline*s*(), not the singular.
"""
if not Network.has_connection():
pytest.skip("No server connected.")
network = Network([ReluLayer()])
exactline = network.exactline([-1.0, 1.0], [1.0, 0.0], True, True)
try:
classifier = LinesClassifier.from_exactlines(exactline)
assert False
except TypeError as e:
pass
def linearize_around(self, inputs, network, layer_index):
"""Linearizes a network before/after a certain layer.
We return a 3-tuple, (pre_lins, mid_inputs, post_lins) satisfying:
1) For all x in the same linear region as x_i:
Network(x) = PostLin_i(Layer_l(PreLin_i(x)))
2) For all x_i:
mid_inputs_i = PreLin_i(x_i)
pre_lins and post_lins are lists of FullyConnectedLayers, one per
input, and mid_inputs is a Numpy array.
"""
pre_network = Network(network.layers[:layer_index])
pre_linearized = self.linearize_network(inputs, pre_network)
mid_inputs = pre_network.compute(inputs)
pre_post_inputs = network.layers[layer_index].compute(mid_inputs)
post_network = Network(network.layers[(layer_index + 1):])
post_linearized = self.linearize_network(pre_post_inputs, post_network)
return pre_linearized, mid_inputs, post_linearized
def test_compute_from_network():
"""Tests the it works given an arbitrary network and lines.
"""
if not Network.has_connection():
pytest.skip("No server connected.")
network = Network([ReluLayer()])
lines = [(np.array([0.0, 1.0]), np.array([0.0, -1.0])),
(np.array([2.0, 3.0]), np.array([4.0, 3.0]))]
helper = IntegratedGradients(network, lines)
helper.partial_compute()
assert len(helper.exactlines) == len(lines)
assert np.allclose(helper.exactlines[0], [0.0, 0.5, 1.0])
assert np.allclose(helper.exactlines[1], [0.0, 1.0])
assert helper.n_samples == [2, 1]
attributions_0 = helper.compute_attributions(0)
assert len(attributions_0) == len(lines)
# The second component doesn't affect the 0-label at all, and the first
# component is 0 everywhere, so we have int_0^0 0.0dx = 0.0
assert np.allclose(attributions_0[0], [0.0, 0.0])
# Gradient of the 0-label is (1.0, 0.0) everywhere since its in the first
# orthant, and the partition has a size of (2.0, 0.0), so the IG is (2.0,
# 0.0).
assert np.allclose(attributions_0[1], [2.0, 0.0])
attributions_1 = helper.compute_attributions(1)
assert len(attributions_1) == len(lines)
# The gradient in the first partition is (0.0, 1.0) with a size of (0.0,
# -1.0) -> contribution of (0.0, -1.0). In the second partition, (0.0,
# 0.0)*(0.0, -1.0) = (0.0, 0.0).
assert np.allclose(attributions_1[0], [0.0, -1.0])
# Gradient is (0, 1) and the size is (2, 0) so IG is (0, 0).
assert np.allclose(attributions_1[1], [0.0, 0.0])
attributions_1_re = helper.compute_attributions(1)
# Ensure it doesn't re-compute the attributions.
assert attributions_1 is attributions_1_re
def test_compute_from_exactlines():
"""Tests the it works given pre-transformed lines.
"""
if not Network.has_connection():
pytest.skip("No server connected.")
network = Network([ReluLayer()])
lines = [(np.array([0.0, 1.0]), np.array([0.0, -1.0])),
(np.array([2.0, 3.0]), np.array([4.0, 3.0]))]
exactlines = network.exactlines(lines, True, True)
classifier = LinesClassifier.from_exactlines(exactlines)
classified = classifier.compute()
assert len(classified) == len(lines)
regions, labels = classified[0]
assert np.allclose(regions,
[[[0.0, 1.0], [0.0, 0.0]], [[0.0, 0.0], [0.0, -1.0]]])
assert np.allclose(labels, [1, 0])
regions, labels = classified[1]
assert np.allclose(regions,
[[[2.0, 3.0], [3.0, 3.0]], [[3.0, 3.0], [4.0, 3.0]]])
assert np.allclose(labels, [1, 0])
def test_already_good():
"""Point-wise test case where all constraints are already met.
We want to make sure that it doesn't change any weights unnecessarily.
"""
network = Network([
FullyConnectedLayer(np.eye(2), np.zeros(shape=(2,))),
ReluLayer(),
])
layer_index = 0
points = [[1.0, 0.5], [2.0, -0.5], [4.0, 5.0]]
labels = [0, 0, 1]
patcher = NetPatcher(network, layer_index, points, labels)
patched = patcher.compute(steps=1)
patched_layer = patched.value_layers[0]
assert np.allclose(patched_layer.weights.numpy(), np.eye(2))
assert np.allclose(patched_layer.biases.numpy(), np.zeros(shape=(2,)))
def test_optimal():
"""Point-wise test case that the greedy algorithm can solve in 1 step.
All it needs to do is triple the second component.
"""
network = Network([
FullyConnectedLayer(np.eye(2), np.zeros(shape=(2, ))),
ReluLayer(),
])
layer_index = 0
points = [[1.0, 0.5], [2.0, -0.5], [5.0, 4.0]]
labels = [1, 0, 1]
patcher = ProvableRepair(network, layer_index, points, labels)
patched = patcher.compute()
assert patched.differ_index == 0
assert np.count_nonzero(
np.argmax(patched.compute(points), axis=1) == labels) == 3
def compute(self):
network = Network.deserialize(self.network.serialize())
if self.layer is not None:
for param in self.get_parameters(network):
param.requires_grad = False
parameters = self.get_parameters(network, self.layer)
for param in parameters:
param.requires_grad = True
if self.norm_objective:
original_parameters = [param.detach().clone() for param in parameters]
for param in original_parameters:
# Do not train these, they're just for reference.
param.requires_grad = False
start = timer()
optimizer = torch.optim.SGD(parameters, lr=self.lr, momentum=self.momentum)
indices = list(range(len(self.inputs)))
random.seed(24)
self.epoched_out = None
holdout_n_correct = self.holdout_n_correct(network)
for epoch in range(self.epochs):
# NOTE: In the paper, we checked this _after_ the inner loop. It
# should only make a difference in the case where the network
# already met the specification, so should make no difference to
# the results.
if self.auto_stop and self.is_done(network):
self.maybe_print("100% training accuracy!")
self.epoched_out = False
break
random.shuffle(indices)
losses = []
for batch_start in range(0, len(self.inputs), self.batch_size):
batch = slice(batch_start, batch_start + self.batch_size)
inputs = torch.tensor([self.inputs[i] for i in indices[batch]])
labels = torch.tensor([self.labels[i] for i in indices[batch]])
# representatives = [self.representatives[i] for i in indices[batch]]
optimizer.zero_grad()
output = network.compute(inputs)
loss = torch.nn.functional.cross_entropy(output, labels)
if self.norm_objective:
for curr_param, og_param in zip(parameters, original_parameters):
delta = (curr_param - og_param).flatten()
loss += torch.linalg.norm(delta, ord=2)
loss += torch.linalg.norm(delta, ord=float("inf"))
loss.backward()
losses.append(loss)
optimizer.step()
self.maybe_print("Average Loss:", torch.mean(torch.tensor(losses)))
if self.holdout_set is not None:
new_holdout_n_correct = self.holdout_n_correct(network)
self.maybe_print("New holdout n correct:", new_holdout_n_correct, "/", len(self.holdout_set))
if new_holdout_n_correct < holdout_n_correct:
self.maybe_print("Holdout accuracy dropped, ending!")
break
holdout_n_correct = new_holdout_n_correct
else:
self.epoched_out = True
for param in parameters:
param.requires_grad = False
self.timing = dict({
"total": timer() - start,
})
return network
