def test_vector_edge_loss(self):
nodes = [Node(), Node(), Node()]
edges = [
Edge(nodes[0], nodes[1], Attribute(Tensor([-50., -10., -5.]))),
Edge(nodes[1], nodes[0], Attribute(Tensor([-40., 100., 120.]))),
Edge(nodes[0], nodes[2], Attribute(Tensor([1., 3., 4.]))),
Edge(nodes[0], nodes[0], Attribute(Tensor([2., 2., 2.])))
]
g1 = Graph(nodes, edges)
g2 = deepcopy(g1)
g2.ordered_edges[0].attr.val = Tensor([-45., -11., -5.])
g2.ordered_edges[1].attr.val = Tensor([-40., 200., 121.])
g2.ordered_edges[2].attr.val = Tensor([1.1, 3., 3.9])
g2.ordered_edges[3].attr.val = Tensor([2., 2., 2.1])
loss = GraphLoss(e_fn=MSELoss())
loss_val = loss(g1, g2).detach().numpy()
# division by 3 because there are three entries per vector
# division by 4 because there are four edges
target_loss_val = ((-50. + 45.)**2 + (-10. + 11.)**2 +
(-40. + 40.)**2 + (100. - 200.)**2 +
(120. - 121.)**2 + (1 - 1.1)**2 + (4. - 3.9)**2 +
(2 - 2.1)**2) / 3 / 4
self.assertTrue(np.isclose(loss_val, target_loss_val))
def _create_tuple(x_min: int, x_max: int) -> Tuple[Graph, Graph]:
"""
Creates a sorting task training sample, i.e. a tuple consisting of (input, target).
:param x_min: Number min value
:param x_max: Number max value (exclusive)
:return: Two graphs, both fully connected, one with randomly initialized attributes, the other one with targets.
"""
assert x_min < x_max
vec = torch.arange(x_min, x_max)
nodes = [Node(Attribute(x)) for x in vec]
g_in = Graph(nodes, [])
g_out = deepcopy(g_in)
def sorted_edge_attribute_generator(n1: Node, n2: Node) -> Attribute:
val = [0, 1] if n1.attr.val < n2.attr.val else [1, 0]
return Attribute(torch.Tensor(val).float())
g_in.add_all_edges(
reflexive=False,
attribute_generator=lambda n1, n2: Attribute(torch.randn(2)))
g_out.add_all_edges(
reflexive=False,
attribute_generator=sorted_edge_attribute_generator)
return g_in, g_out
def test_combined_loss(self):
nodes = [
Node(Attribute(Tensor([-4., -8.]))),
Node(Attribute(Tensor([1., 5.]))),
Node(Attribute(Tensor([4., 4.]))),
Node(Attribute(Tensor([0., 1., 5.])))
]
edges = [
Edge(nodes[0], nodes[1], Attribute(Tensor([1., 2., 3.]))),
Edge(nodes[1], nodes[2], Attribute(Tensor([1., 2.]))),
Edge(nodes[2], nodes[1], Attribute(Tensor([5.]))),
Edge(nodes[1], nodes[3], Attribute(Tensor([1., 2., 3., 4.])))
]
u = Attribute(
Tensor([[1., 2., 4., 3.], [8., 3., 0., 3.], [1., 7., 5., 3.]]))
g1 = Graph(nodes, edges, attr=u)
g2 = deepcopy(g1)
g2.ordered_nodes[0].attr.val = Tensor([-4., -8.1])
g2.ordered_nodes[1].attr.val = Tensor([2., 6.])
g2.ordered_nodes[3].attr.val = Tensor([1., 1.5, 5.])
g2.ordered_edges[0].attr.val = Tensor([2., 3., 4.])
g2.ordered_edges[1].attr.val = Tensor([5., 10.])
g2.attr.val = Tensor([[2., 2., 4., 3.], [100, 3., 1., 3.],
[1., 14., 5., 3.]])
loss = GraphLoss(e_fn=MSELoss(), v_fn=L1Loss(), u_fn=MSELoss())
loss_val = loss(g1, g2).detach().numpy()
e_loss = (1. + (4**2 + 8**2) / 2) / 4
v_loss = (.1 / 2 + 2. / 2 + (1 + .5) / 3) / 4
u_loss = (1 + (8 - 100)**2 + 1 + 7**2) / 12 / 1
target_loss_val = v_loss + e_loss + u_loss
self.assertTrue(np.isclose(loss_val, target_loss_val))
def forward(self, aggr_e: Attribute, aggr_v: Attribute,
u: Attribute) -> Attribute:
x = torch.cat((aggr_e.val, aggr_v.val, u.val))
x = F.relu(self.dense(x))
x = F.dropout(x, p=self.drop_p, training=self.training)
return Attribute(x)
def forward(self, e: Attribute, v_r: Attribute, v_s: Attribute,
u: Attribute) -> Attribute:
x = torch.cat((e.val, v_r.val, v_s.val, u.val))
x = F.relu(self.dense(x))
x = F.dropout(x, p=self.drop_p, training=self.training)
return Attribute(x)
def add_reflexive_edges(self, attribute_generator: Optional[Callable[[Node], Attribute]] = None):
if attribute_generator is None:
attribute_generator = lambda _: Attribute()
for n in self.nodes:
e = Edge(n, n, attribute_generator(n))
self.add_edge(e)
def _feed(aggr: Aggregation, in_val: List[List[float]],
out_target: List[float]) -> bool:
in_attrs = [Attribute(torch.Tensor(np.array(x))) for x in in_val]
out_target = np.array(out_target)
out_tensor: torch.Tensor = aggr(in_attrs).val
out_val = out_tensor.detach().numpy()
return np.allclose(out_val, out_target)
def test_scalar_edge_loss(self):
nodes = [Node(), Node(), Node()]
edges = [
Edge(nodes[0], nodes[1], Attribute(Tensor([-50.]))),
Edge(nodes[1], nodes[0], Attribute(Tensor([-40.]))),
Edge(nodes[0], nodes[2], Attribute(Tensor([1.])))
]
g1 = Graph(nodes, edges)
g2 = deepcopy(g1)
g2.ordered_edges[0].attr.val = Tensor([-45.])
g2.ordered_edges[1].attr.val = Tensor([-40.])
g2.ordered_edges[2].attr.val = Tensor([1.1])
loss = GraphLoss(e_fn=MSELoss())
loss_val = loss(g1, g2).detach().numpy()
target_loss_val = ((-50. + 45.)**2 + (-40. + 40.)**2 +
(1 - 1.1)**2) / 3
self.assertTrue(np.isclose(loss_val, target_loss_val))
def test_identity_property(self):
nodes = [
Node(Attribute([1, 23, 4])),
Node(Attribute("stringattr")),
Node(Attribute([1])),
Node(Attribute(5))
]
edges = [
Edge(nodes[0], nodes[1], Attribute({'dict': 1234})),
Edge(nodes[0], nodes[1], Attribute([1, 2, 3])),
Edge(nodes[0], nodes[2], Attribute(5)),
Edge(nodes[1], nodes[2], Attribute([3, 4, 5])),
Edge(nodes[1], nodes[1], Attribute())
]
global_state = Attribute([[1, 2, 3], [5, 6, 7]])
g1 = Graph(nodes, edges, global_state)
g1_prime = Graph.from_dict(g1.asdict())
self.assertTrue(g1 == g1_prime)
def forward(self, attrs: List[Attribute]) -> Attribute:
n = len(attrs)
assert n > 0, "Maximum aggregation cannot be applied to empty lists."
attr_vals = [a.val for a in attrs]
attr_vals = torch.stack(attr_vals)
attr_vals_max = torch.max(
attr_vals, dim=0, keepdim=False)[0] # returns (Tensor, LongTensor)
return Attribute(attr_vals_max)
def forward(self, attrs: List[Attribute]) -> Attribute:
n = len(attrs)
assert n > 0, "Average aggregation cannot be applied to empty lists."
attr_vals = [a.val for a in attrs]
attr_vals = torch.stack(attr_vals)
attr_vals_sum = torch.sum(attr_vals, dim=0, keepdim=False)
attr_vals_avg = attr_vals_sum / n
return Attribute(attr_vals_avg)
def forward(self, aggr_e: Attribute, v: Attribute,
u: Attribute) -> Attribute:
img_tensor = v.val
assert aggr_e.val.is_cuda, "Following reshaping does only work on GPU"
aggr_channels = aggr_e.val.unsqueeze(-1).unsqueeze(
-1) # add height and width dimension for broadcasting
img_plus_aggr = torch.add(img_tensor, aggr_channels) / 2
conv_out = self.dense_layers(img_plus_aggr)
return Attribute(conv_out)
def test_forward_pass(self):
linear_block = LinearIndependentGNBlock(e_config=(4, 8, True),
v_config=(1, 1, False),
u_config=(16, 16, True))
nodes = Node.from_vals(
[torch.randn(1), torch.randn(1),
torch.randn(1)])
edges = [
Edge(nodes[0], nodes[1], Attribute(torch.randn(4))),
Edge(nodes[1], nodes[2], Attribute(torch.randn(4))),
Edge(nodes[2], nodes[1], Attribute(torch.randn(4)))
]
g_in = Graph(nodes, edges, Attribute(torch.randn(16)))
# noinspection PyUnusedLocal
g_out = linear_block(g_in)
self.assertTrue(
True
) # the assertion is that the forward pass works without errors
def forward(self, attrs: List[Attribute]) -> Attribute:
assert len(
attrs) > 0, "Sum aggregation cannot be applied to empty lists"
in_shape = attrs[0].val.shape
attr_vals = [a.val.view(-1) for a in attrs]
attr_vals_concat = torch.stack(attr_vals)
attr_vals_sum = torch.sum(attr_vals_concat, dim=0, keepdim=False)
assert attr_vals_sum.shape == in_shape
return Attribute(attr_vals_sum)
def load_graph(self, path: str) -> Graph:
# read JSON from page description file
json_file_path = glob(os.path.join(path, '*.json'))
assert len(json_file_path) == 1, "Number of json files in '{}' must be exactly one.".format(path)
json_file_path = json_file_path[0]
with open(json_file_path, encoding='utf-8') as json_file:
pages_json: List = json.load(json_file)
# read screenshot paths
imgs = self.load_images(os.path.join(path, 'image'))
assert len(pages_json) == len(imgs), "Number of pages and number of screenshots mismatch in '{}'.".format(path)
# extract nodes
nodes = {}
for page_json in pages_json:
desktop_img, mobile_img = imgs[page_json['id']-1]
node_attribute = PageAttribute.from_json(page_json, desktop_img, mobile_img)
url = page_json['base_url']
if url in nodes:
logging.debug("Found two nodes with the same URL.")
continue
node = Node(node_attribute)
nodes[url] = node
# extract edges
edges = set()
for page_json in pages_json:
url = page_json['base_url']
source_node = nodes[url]
for edge_json in page_json.get('urls', []):
target_url = edge_json['url']
if target_url not in nodes:
logging.debug("Invalid link target URL. Could not find a node that corresponds to it.")
continue
target_node = nodes[target_url]
edge_attribute = LinkAttribute(target_url)
edge = Edge(sender=source_node, receiver=target_node, attr=edge_attribute)
edges.add(edge)
return Graph(nodes=list(nodes.values()), edges=list(edges), attr=Attribute(None))
def test_learn_identity(self):
# construct input graph with random values
nodes = Node.from_vals(
[torch.randn(1), torch.randn(1),
torch.randn(1)])
edges = [
Edge(nodes[0], nodes[1], Attribute(torch.randn(4))),
Edge(nodes[1], nodes[2], Attribute(torch.randn(4))),
Edge(nodes[2], nodes[1], Attribute(torch.randn(4)))
]
g_in = Graph(nodes, edges, Attribute(torch.randn(16)))
g_target = deepcopy(g_in)
block_1 = LinearIndependentGNBlock(e_config=(4, 8, True),
v_config=(1, 12, True),
u_config=(16, 16, True))
block_2 = LinearIndependentGNBlock(e_config=(8, 4, False),
v_config=(12, 1, False),
u_config=(16, 16, False))
model = torch.nn.Sequential(block_1, block_2)
opt = optim.SGD(model.parameters(), lr=.1, momentum=0)
loss_fn = GraphLoss(e_fn=MSELoss(), v_fn=MSELoss(), u_fn=MSELoss())
loss = torch.Tensor([1.])
for step in range(100):
model.train()
opt.zero_grad()
g_out = model(g_in)
loss = loss_fn(g_out, g_target)
loss.backward()
opt.step()
final_loss = loss.detach().numpy()
print(final_loss)
self.assertTrue(final_loss < 1e-3)
def add_all_edges(self, reflexive: bool = True,
attribute_generator: Optional[Callable[[Node, Node], Attribute]] = None) -> None:
"""
Modifies the graph in-place, such that it is fully connected, adding n^n edges, where n is the number of nodes.
:param reflexive: Whether to connect nodes to themselves
:param attribute_generator: New edges will be given an attribute generated by the attribute generator
"""
if attribute_generator is None:
attribute_generator = lambda sn, rn: Attribute()
for n1 in self.nodes:
for n2 in self.nodes:
if not reflexive and n1 == n2:
continue
e = Edge(n1, n2, attribute_generator(n1, n2))
self.add_edge(e)
def __init__(self, nodes: List[Node], edges: List[Edge], attr: Optional[Attribute] = None):
if attr is None:
attr = Attribute(val=None)
assert isinstance(nodes, list) and isinstance(edges, list), "Nodes and edges must be lists"
self.nodes = set(nodes) # V
self.edges = set(edges) # E
self.attr = attr # e
# store ordered representations for easy alignment of two graphs with equal structure
self.ordered_nodes = nodes
self.ordered_edges = edges
self.device: torch.device = None
self._check_integrity()
def test_graph_equality(self):
v_0, v_1, v_2 = Node(Attribute(0.)), Node(Attribute(1.)), Node(
Attribute(2.))
vs = [v_0, v_1, v_2] # nodes
es = [Edge(v_0, v_1), Edge(v_0, v_2)] # edges
g_0 = Graph(nodes=vs, edges=es)
v_0, v_1, v_2 = Node(Attribute(0.)), Node(Attribute(1.)), Node(
Attribute(2.))
vs = [v_0, v_1, v_2] # nodes
es = [Edge(v_0, v_1), Edge(v_0, v_2)] # edges
g_1 = Graph(nodes=vs, edges=es)
self.assertTrue(g_0 == g_1)
def forward(self, aggr_e: Attribute, aggr_v: Attribute,
u: Attribute) -> Attribute:
return Attribute(self.dense(u.val))
def from_dict(d: Dict) -> 'Graph':
nodes_dict = {k: Node.from_dict(node_dict) for k, node_dict in d['nodes'].items()}
nodes = list(nodes_dict.values())
edges = [Edge.from_dict(e, nodes_dict) for e in d['edges']]
attr = Attribute.from_dict(d['attr'])
return Graph(nodes, edges, attr)
def test_lists_equal_objects_comparator(self):
l1 = [Node(Attribute(1)), Node(Attribute(2)), Node(Attribute("test"))]
l2 = [Node(Attribute(1)), Node(Attribute(2)), Node(Attribute("test"))]
self.assertTrue(lists_equal(l1, l2, comparator=Node.eq_attr))
def sorted_edge_attribute_generator(n1: Node, n2: Node) -> Attribute:
val = [0, 1] if n1.attr.val < n2.attr.val else [1, 0]
return Attribute(torch.Tensor(val).float())
def forward(self, aggr_e: Attribute, aggr_v: Attribute,
u: Attribute) -> Attribute:
return Attribute(aggr_e.val + aggr_v.val - u.val)
def test_sets_equal_objects_comparator(self):
s1 = {Node(Attribute(1)), Node(Attribute(2)), Node(Attribute("test"))}
s2 = {Node(Attribute(2)), Node(Attribute(1)), Node(Attribute("test"))}
self.assertTrue(sets_equal(s1, s2, comparator=Node.eq_attr))
def forward(self, e: Attribute, v_r: Attribute, v_s: Attribute,
u: Attribute) -> Attribute:
return Attribute(global_avg_pool(v_s.val))
def test_basic(self):
"""
Basic test w/o PyTorch, all attributes are scalars, edges do not have attributes.
Feeds a graph through a basic graph block twice and compares to the target values after both passes.
"""
# create data structure
v_0, v_1, v_2 = Node(Attribute(1)), Node(Attribute(10)), Node(
Attribute(20))
vs = [v_0, v_1, v_2] # nodes
es = [Edge(v_0, v_1), Edge(v_0, v_2), Edge(v_1, v_2)]
g_0 = Graph(nodes=vs, edges=es, attr=Attribute(0))
# create block w/ functions
block = GNBlock(phi_e=SenderIdentityEdgeUpdate(),
phi_v=EdgeNodeSumNodeUpdate(),
phi_u=MixedGlobalStateUpdate(),
rho_ev=ScalarSumAggregation(),
rho_vu=ScalarSumAggregation(),
rho_eu=ScalarSumAggregation())
g_1 = block(g_0)
v_0, v_1, v_2 = Node(Attribute(1)), Node(Attribute(10 + 1)), Node(
Attribute(20 + 11))
vs = [v_0, v_1, v_2] # nodes
es = [
Edge(v_0, v_1, Attribute(1)),
Edge(v_0, v_2, Attribute(1)),
Edge(v_1, v_2, Attribute(10))
]
g_1_target = Graph(nodes=vs, edges=es, attr=Attribute(35))
self.assertTrue(g_1 == g_1_target)
g_2 = block(g_1)
v_0, v_1, v_2 = Node(Attribute(1)), Node(Attribute(10 + 2)), Node(
Attribute(20 + 11 + 12))
vs = [v_0, v_1, v_2] # nodes
es = [
Edge(v_0, v_1, Attribute(1)),
Edge(v_0, v_2, Attribute(1)),
Edge(v_1, v_2, Attribute(11))
]
g_2_target = Graph(nodes=vs,
edges=es,
attr=Attribute(1 + 12 + 43 - 35))
self.assertTrue(g_2 == g_2_target)
g: Graph = batch[0]['graph'].to(device)
g_hash = int(batch[0]['rank'])
assert g_hash not in cache
for n in g.nodes:
desktop_img: Tensor = n.attr.val['desktop_img']
mobile_img: Tensor = n.attr.val['mobile_img']
# desktop and mobile feature vector
x1, x2 = net(desktop_img.unsqueeze(0), mobile_img.unsqueeze(0))
x: torch.Tensor = torch.cat((x1, x2), dim=1).view(-1)
x = x.detach().cpu()
x = torch.Tensor(x.data)
del n.attr
n.attr = Attribute(x)
g.to(torch.device('cpu'))
cache[g_hash] = g
if len(cache) == 4000:
save_cached_graphs(cache, file_ctr)
file_ctr += 1
cache = {}
save_cached_graphs(cache, file_ctr)
def forward(self, attrs: List[Attribute]) -> Attribute:
return Attribute(self.constant_val)
def forward(self, attrs: List[Attribute]) -> Attribute:
return Attribute(sum([a.val for a in attrs]))