def get_fbvs_max_size(self, g: MultiGraph, k: int) -> set:
if len(g) <= k + 2:
return set(g.nodes()[:k])
# Construct a trivial FVS of size k + 1 on the first k + 3 vertices of G.
nodes = g.nodes()
# The set of nodes currently under consideration.
node_set = set(nodes[:(k + 2)])
# The current best solution, of size (k + 1) before each compression step,
# and size <= k at the end.
soln = set(nodes[:k])
for i in range(k + 2, len(nodes)):
soln.add(nodes[i])
node_set.add(nodes[i])
if len(soln) < k + 1:
continue
assert (len(soln) == (k + 1))
assert (len(node_set) == (i + 1))
new_soln = self.ic_compression(g.subgraph(node_set), soln, k)
if new_soln is None:
return None
soln = new_soln
assert (len(soln) <= k)
return soln
def simplify(streckennetz: nx.MultiGraph) -> nx.MultiGraph:
print('simplifying...')
# Remove the shorter edge of two parallel edges
edges_to_remove = []
for edges_index in streckennetz.edges():
edges = streckennetz[edges_index[0]][edges_index[1]]
if len(edges) > 1:
min_length = edges[0]['length']
min_length_index = 0
for i in edges:
if edges[i]['length'] < min_length:
min_length = edges[i]['length']
min_length_index = i
edges_to_remove.append(list(edges_index) + [min_length_index])
print('found', len(edges_to_remove), 'parallel edges')
streckennetz.remove_edges_from(edges_to_remove)
# Remove nodes, that only have a single edge
while True:
nodes_to_remove = []
for node in streckennetz.nodes():
if streckennetz.degree(
node) < 2 and 'type' not in streckennetz.nodes[node]:
nodes_to_remove.append(node)
if nodes_to_remove:
print('found', len(nodes_to_remove), 'deadend nodes')
streckennetz.remove_nodes_from(nodes_to_remove)
nodes_to_remove = []
else:
break
return streckennetz
def get_graph_extents(
networkX_multigraph: nx.MultiGraph
) -> Tuple[float, float, float, float]:
"""
Parameters
----------
networkX_multigraph
A `NetworkX` `MultiGraph` with `x` and `y` node parameters.
Returns
-------
Tuple
A tuple of `min_x`, `min_y`, `max_x`, `max_y` values.
"""
# get min and maxes for x and y
min_x = np.inf
max_x = -np.inf
min_y = np.inf
max_y = -np.inf
for n, d in networkX_multigraph.nodes(data=True):
if d['x'] < min_x:
min_x = d['x']
if d['x'] > max_x:
max_x = d['x']
if d['y'] < min_y:
min_y = d['y']
if d['y'] > max_y:
max_y = d['y']
return min_x, min_y, max_x, max_y
def reduction2(g: MultiGraph, w: set, h: MultiGraph, k: int) -> (int, int, bool):
for v in h.nodes():
# Check if G[W ∪ {v}] contains a cycle.
if not is_forest(g.subgraph(w.union({v}))):
g.remove_node(v)
h.remove_nodes_from([v])
return (k - 1, v, True)
return (k, None, False)
def reduction1(g: MultiGraph, w: set, h: MultiGraph, k: int) -> (int, int, bool): changed = False for v in g.nodes(): if g.degree(v) <= 1: g.remove_node(v) h.remove_nodes_from([v]) changed = True return (k, None, changed)
def test_relabel_nodes_multigraph():
"""failed after switching to dg.relabel_nodes"""
G = MultiGraph([('a', 'b'), ('a', 'b')])
mapping = {'a': 'aardvark', 'b': 'bear'}
G = relabel_nodes(G, mapping, copy=False)
assert sorted(G.nodes()) == ['aardvark', 'bear']
assert_edges_equal(sorted(G.edges()), [('aardvark', 'bear'),
('aardvark', 'bear')])
def graph_minus(g: MultiGraph, w: set) -> MultiGraph: gx = MultiGraph() for (n1, n2) in g.edges(): if n1 not in w and n2 not in w: gx.add_edge(n1, n2) for n in g.nodes(): if n not in w: gx.add_node(n) return gx
def reduction3(self, g: MultiGraph, k: int) -> (int, List[int], bool):
"""
If there is a vertex v of degree at most 1, delete v.
"""
changed = False
for v in g.nodes():
if g.degree(v) <= 1:
g.remove_node(v)
changed = True
return k, None, changed
def reduction3(self, g: MultiGraph, k: int) -> (int, List[int], bool):
"""
If there is a vertex v of degree at most 1, delete v.
"""
changed = False
for v in g.nodes():
if g.degree(v) <= 1:
g.remove_node(v)
changed = True
return k, None, changed
def reduction1(self, g: MultiGraph, w: set, h: MultiGraph, k: int) -> (int, int, bool):
"""
Delete all nodes of degree 0 or 1 as they can't be part of any cycles.
"""
changed = False
for v in g.nodes():
if g.degree(v) <= 1:
g.remove_node(v)
h.remove_nodes_from([v])
changed = True
return k, None, changed
def get_node_feature_matrix(self, G:nx.MultiGraph, doc, embd_dim=96,
embedding_type=None, as_torch=True, is_padding_pos=True, **kwargs):
"""
Returns X with shape [num_nodes, node_feat_dim]
"""
assert G is not None
NDV = G.nodes(data=True); NV = G.nodes(data=False)
_is_head_node = lambda x: 'obj' in x
head_nodes = list(filter(_is_head_node, NV))
objs = self.clevr_parser.filter_clevr_objs(doc.ents)
N, M = len(NDV), (embd_dim+3) if is_padding_pos else embd_dim
feat_mats = []
for i, entity in enumerate(objs):
if entity.label_ not in ('CLEVR_OBJS', 'CLEVR_OBJ'):
continue
ent_mat = self.clevr_parser.get_clevr_entity_matrix_embedding(entity, dim=embd_dim,
include_obj_node_emd=True,
is_padding_pos=is_padding_pos)
ent_mat = ent_mat.reshape((-1, embd_dim))
if is_padding_pos:
head_node = G.nodes.get(head_nodes[i])
pos = head_node.get('pos') # pos = (x,y,z): Tuple[float]
if not pos:
pos = (0.0, 0.0, 0.0)
pos = np.tile(np.asarray(pos, dtype=float), ent_mat.shape[0]).reshape(-1, 3)
# np.pad(ent_mat, ((0,0), (0, 3)), 'constant', constant_values=(0, 0.0) )
ent_mat = np.concatenate((ent_mat, pos), axis=1) # [n, embd_dim+3]
feat_mats.append(ent_mat)
if len(feat_mats) > 1:
feat_mats = reduce(lambda a, b: np.vstack((a, b)), feat_mats)
else:
feat_mats = feat_mats[0]
assert feat_mats.shape == (N, M)
if as_torch:
feat_mats = torch.from_numpy(feat_mats).float()
if kwargs.get('is_cuda'):
device = 'cuda' if torch.cuda.is_available() else 'cpu'
feat_mats = feat_mats.to(device)
return feat_mats
def reduction3(g: MultiGraph, w: set, h: MultiGraph, k: int) -> (int, int, bool): for v in h.nodes(): if g.degree(v) == 2: # If v has a neighbour in H, short-curcuit it. if len(h[v]) >= 1: # Delete v and make its neighbors adjacent. [n1, n2] = g.neighbors(v) g.remove_node(v) g.add_edge(n1, n2) # Update H accordingly. h.remove_nodes_from([v]) if n1 not in w and n2 not in w: h.add_edge(n1, n2) return (k, None, True) return (k, None, False)
def subgraph_by_timestamp(mg: nx.MultiGraph, start: int, end: int) -> nx.Graph:
edges = filter(
lambda edge: start <= edge[2]["timestamp"] and edge[2]["timestamp"] <
end,
mg.edges(data=True),
)
g = nx.Graph()
for node in mg.nodes():
g.add_node(node)
for u, v, data in edges:
if g.has_edge(u, v):
g[u][v]["weight"] += data["weight"]
else:
g.add_edge(u, v, weight=data["weight"])
return g
def reduction2(self, g: MultiGraph, w: set, h: MultiGraph, k: int) -> (int, int, bool):
"""
If there exists a node v in H such that G[W ∪ {v}]
contains a cycle, then include v in the solution, delete v and decrease the
parameter by 1. That is, the new instance is (G - {v}, W, k - 1).
If v introduces a cycle, it must be part of X as none of the vertices in W
will be available to neutralise this cycle.
"""
for v in h.nodes():
# Check if G[W ∪ {v}] contains a cycle.
if not is_acyclic(g.subgraph(w.union({v}))):
g.remove_node(v)
h.remove_nodes_from([v])
return k - 1, v, True
return k, None, False
def update_graph_with_spatial_re(cls, G: nx.MultiGraph, doc,
**kwargs) -> nx.MultiGraph:
spatial_res = cls.filter_spatial_re(doc.ents)
if spatial_res is None:
return G
NV = G.nodes(data=False)
_is_head_node = lambda x: 'obj' in x
head_nodes = list(filter(_is_head_node, NV))
objs = cls.filter_clevr_objs(doc.ents)
relations = cls.extract_spatial_relations(doc)
assert len(objs) == len(head_nodes)
assert len(relations) == len(spatial_res)
o2n_map = dict(zip(objs, head_nodes))
sr2r_map = dict(zip(spatial_res, relations))
def add_nodes(n0, n1, r):
if not G.has_edge(*(n0, n1, 'spatial_re')):
#G.add_edge(n0, n1, **{'label':'spatial_re', 'val': r})
G.add_edge(n0, n1, spatial_re=r)
# G.add_edge(n0, n1,key=r)
has_and = 'and' in doc.text
# has_or = 'or' in doc.text
# is_logical_re = has_and or has_or
for i, ent in enumerate(doc.ents):
if ent in objs:
continue
if ent in spatial_res:
_r = sr2r_map[ent]
try:
if i > 1:
if has_and:
n1 = o2n_map[doc.ents[i + 1]]
n0 = o2n_map[doc.ents[0]]
else:
n0 = o2n_map[doc.ents[i - 1]]
n1 = o2n_map[doc.ents[i + 1]]
else:
n0 = o2n_map[doc.ents[i - 1]]
n1 = o2n_map[doc.ents[i + 1]]
add_nodes(n0, n1, _r)
except IndexError as ie:
print(ie)
except UnboundLocalError as lr:
print(lr)
return G
def connect_matching_pair_edges(self, Gu: nx.MultiGraph,
ls, rs, obj_node_id='obj',
connect_obj_rel_edges=False) -> nx.Graph:
"""
Ground truth (_gt_)generator function for a combined graph. Used for training S_0
:param Gu: A (unconnected) composed graph of Gs, Gt. No relational links between head nodes,
thus, the number_connected_components(Gs|Gt) = number of object nodes in each graph.
:param Gs: The source Graph, i.e., the text graph representation
:param Gt: The target Graph, i.e., the grounding (image features) graph representation
:param obj_node_id: the identifier determining a obj (or head) node
:return: matching_pairs: List of matching pair tuples between Gs, Gt
"""
matching_pairs, unmatched_pairs = self.get_matching_pairs_in_bipartite_graph(Gu, ls, rs, obj_node_id)
NDV = Gu.nodes(data=True)
# Connect the matching pairs if not connected #
for pair in matching_pairs:
s_node, t_node = pair
if not Gu.has_edge(s_node, t_node):
Gu.add_edge(s_node, t_node, '<gt>')
# Connect Attr Nodes #
is_head_node = lambda x: obj_node_id in x
Ns = nx.neighbors(Gu, s_node)
Nt = list(nx.neighbors(Gu, t_node))
for ns in Ns:
if is_head_node(ns):
continue
# Check label equality only, 'val' equality already verified
ns_label = NDV[ns]['label']
#logger.debug(f'Source attr node label = {ns_label}')
# TODO: potential issue here, Nt should always have a matching attr node
for nt in filter(lambda x: NDV[x]['label'] == ns_label, Nt):
if not Gu.has_edge(ns, nt):
Gu.add_edge(ns, nt, key='<gt>')
if connect_obj_rel_edges:
# TODO: Add a obj relation edge among all Gs, Gt obj nodes. Actually
# should be done earlier in the parse cycle.
pass
return Gu
def reduction3(self, g: MultiGraph, w: set, h: MultiGraph, k: int) -> (int, int, bool):
"""
If there is a node v ∈ V(H) of degree 2 in G such
that at least one neighbor of v in G is from V (H), then delete this node
and make its neighbors adjacent (even if they were adjacent before; the graph
could become a multigraph now).
"""
for v in h.nodes():
if g.degree(v) == 2:
# If v has a neighbour in H, short-curcuit it.
if len(h[v]) >= 1:
# Delete v and make its neighbors adjacent.
[n1, n2] = g.neighbors(v)
g.remove_node(v)
g.add_edge(n1, n2)
# Update H accordingly.
h.remove_nodes_from([v])
if n1 not in w and n2 not in w:
h.add_edge(n1, n2)
return k, None, True
return k, None, False
def build_node(g: nx.MultiGraph, new_node: int, n: int, const_amt: int, fee_type: str) -> List:
selected = []
# 2 initial edges so routes can pass through new node
while len(selected) < n:
max_reward = 0
best_channel = None
# try all channels
for node in g.nodes():
if node == new_node or node in g[new_node]:
continue
# create new channel
g.add_edge_with_init(new_node, node, default=True)
# calculate max reward
reward, fee = maximise_fee(
g, (new_node, node), const_amt, fee_type)
# select channel if highest reward
if max_reward <= reward:
max_reward = reward
best_channel = (node, fee)
# reset graph for next channel
g.remove_edge(new_node, node)
if best_channel is None:
raise Exception('Profit not possible.')
# add selected channel and redo for next channel
selected.append((new_node, best_channel[0], best_channel[1]))
g.add_edge_with_init(new_node, best_channel[0], default=True)
g.update_fee(new_node, best_channel[0], best_channel[1], fee_type)
# return node with selected channels
return selected
def reduction4(self, g: MultiGraph, k: int) -> (int, List[int], bool):
"""
If there is a vertex v of degree 2, delete v and connect its two neighbors by a new edge.
"""
for v in g.nodes():
if g.degree(v) == 2:
# Delete v and make its neighbors adjacent.
ne = g.neighbors(v)
# We must check whether v has 2 neighbors, or just one but connected to v by multiple edges
if len(ne) == 2:
[n1, n2] = ne
else:
[n1] = ne
n2 = n1
g.remove_node(v)
# Only add the edge if there are currently less than 2 edges between these two nodes
es = g[n1].get(n2, {})
if len(es) < 2:
g.add_edge(n1, n2)
return k, None, True
return k, None, False
def preprocess_2(self, g: MultiGraph, f: set, active_v) -> set:
mif_set = set()
while not is_independent_set(g, f):
mif_set = mif_set.union(f)
for component in connected_components(g.subgraph(f)):
if len(component) > 1:
if active_v in component:
active_v = component.pop()
compressed_node = active_v
else:
compressed_node = component.pop()
g = self.compress(g, component, compressed_node, False)
f = f.intersection(g.nodes())
# Maybe faster with
# f = f.difference(component)
# f.add(compressed_node)
mif_set = mif_set.union(component)
break
mif_set2 = self.mif_main(g, f, active_v)
if mif_set2:
mif_set = mif_set2.union(mif_set)
return mif_set
def reduction4(self, g: MultiGraph, k: int) -> (int, List[int], bool):
"""
If there is a vertex v of degree 2, delete v and connect its two neighbors by a new edge.
"""
for v in g.nodes():
if g.degree(v) == 2:
# Delete v and make its neighbors adjacent.
ne = g.neighbors(v)
# We must check whether v has 2 neighbors, or just one but connected to v by multiple edges
if len(ne) == 2:
[n1, n2] = ne
else:
[n1] = ne
n2 = n1
g.remove_node(v)
# Only add the edge if there are currently less than 2 edges between these two nodes
es = g[n1].get(n2, {})
if len(es) < 2:
g.add_edge(n1, n2)
return k, None, True
return k, None, False
def mif_preprocess_2(g: MultiGraph, f: set, active_v, k: int) -> set: mif_set = set() while not is_independent_set(g, f): mif_set = mif_set.union(f) for component in nxc.connected_components(g.subgraph(f)): if len(component) > 1: if active_v in component: active_v = component.pop() compressed_node = active_v else: compressed_node = component.pop() g = compress(g, component, compressed_node, True) f = f.intersection(g.nodes()) # Maybe faster with # f = f.difference(component) # f.add(compressed_node) mif_set = mif_set.union(component) break mif_set2 = mif_main(g, f, active_v, k) if mif_set2: mif_set = mif_set2.union(mif_set) if k == None or len(mif_set) >= k: return mif_set return None
class ShuttlingGraph(list):
def __init__(self, shuttlingEdges=list() ):
super(ShuttlingGraph, self).__init__(shuttlingEdges)
self.currentPosition = None
self.currentPositionName = None
self.nodeLookup = dict()
self.currentPositionObservable = Observable()
self.graphChangedObservable = Observable()
self.initGraph()
self._hasChanged = True
def initGraph(self):
self.shuttlingGraph = MultiGraph()
for edge in self:
self.shuttlingGraph.add_node(edge.startName)
self.shuttlingGraph.add_node(edge.stopName)
self.shuttlingGraph.add_edge(edge.startName, edge.stopName, key=hash(edge), edge=edge,
weight=abs(edge.stopLine-edge.startLine))
self.nodeLookup[edge.startLine] = edge.startName
self.nodeLookup[edge.stopLine] = edge.stopName
def rgenerateNodeLookup(self):
self.nodeLookup.clear()
for edge in self:
self.nodeLookup[edge.startLine] = edge.startName
self.nodeLookup[edge.stopLine] = edge.stopName
@property
def hasChanged(self):
return self._hasChanged
@hasChanged.setter
def hasChanged(self, value):
self._hasChanged = value
def position(self, line):
return self.nodeLookup.get(line)
def setPosition(self, line):
if self.currentPosition!=line:
self.currentPosition = line
self.currentPositionName = self.position(line)
self.currentPositionObservable.fire( line=line, text=firstNotNone(self.currentPositionName, "") )
def getMatchingPosition(self,graph):
"""Try to match node name/position to the current settings in the provided ShuttlingGraph."""
if not graph:
return self.currentPosition # no change
# Matching node name. Need to set the corresponding position
for edge in self:
if edge.startName == graph.currentPositionName:
return edge.startLine
if edge.stopName == graph.currentPositionName:
return edge.stopLine
#if graph.currentPosition:
# return graph.currentPosition #just use the graph's position
return self.currentPosition
def addEdge(self, edge):
self._hasChanged = True
self.append(edge)
self.shuttlingGraph.add_edge(edge.startName, edge.stopName, key=hash(edge), edge=edge, weight=abs(edge.stopLine-edge.startLine))
self.nodeLookup[edge.startLine] = edge.startName
self.nodeLookup[edge.stopLine] = edge.stopName
self.graphChangedObservable.firebare()
self.setPosition(self.currentPosition)
def isValidEdge(self, edge):
return ((edge.startLine not in self.nodeLookup or self.nodeLookup[edge.startLine] == edge.startName)
and (edge.stopLine not in self.nodeLookup or self.nodeLookup[edge.stopLine] == edge.stopName))
def getValidEdge(self):
index = 0
while self.shuttlingGraph.has_node("Start_{0}".format(index)):
index += 1
startName = "Start_{0}".format(index)
index = 0
while self.shuttlingGraph.has_node("Stop_{0}".format(index)):
index += 1
stopName = "Stop_{0}".format(index)
index = 0
startLine = (max( self.nodeLookup.keys() )+1) if self.nodeLookup else 1
stopLine = startLine + 1
return ShuttleEdge(startName, stopName, startLine, stopLine, 0, 0, 0, 0)
def removeEdge(self, edgeno):
self._hasChanged = True
edge = self.pop(edgeno)
self.shuttlingGraph.remove_edge(edge.startName, edge.stopName, hash(edge))
if self.shuttlingGraph.degree(edge.startName) == 0:
self.shuttlingGraph.remove_node(edge.startName)
if self.shuttlingGraph.degree(edge.stopName) == 0:
self.shuttlingGraph.remove_node(edge.stopName)
self.graphChangedObservable.firebare()
self.rgenerateNodeLookup()
self.setPosition(self.currentPosition)
def setStartName(self, edgeno, startName):
self._hasChanged = True
startName = str(startName)
edge = self[edgeno]
if edge.startName != startName:
self.shuttlingGraph.remove_edge(edge.startName, edge.stopName, key=hash(edge))
if self.shuttlingGraph.degree(edge.startName) == 0:
self.shuttlingGraph.remove_node(edge.startName)
edge.startName = startName
self.shuttlingGraph.add_edge(edge.startName, edge.stopName, key=hash(edge), edge=edge,
weight=abs(edge.stopLine-edge.startLine) )
self.graphChangedObservable.firebare()
self.setPosition(self.currentPosition)
self.rgenerateNodeLookup()
return True
def setStopName(self, edgeno, stopName):
self._hasChanged = True
stopName = str(stopName)
edge = self[edgeno]
if edge.stopName != stopName:
self.shuttlingGraph.remove_edge(edge.startName, edge.stopName, key=hash(edge))
if self.shuttlingGraph.degree(edge.stopName) == 0:
self.shuttlingGraph.remove_node(edge.stopName)
edge.stopName = stopName
self.shuttlingGraph.add_edge(edge.startName, edge.stopName, key=hash(edge), edge=edge,
weight=abs(edge.stopLine-edge.startLine) )
self.graphChangedObservable.firebare()
self.rgenerateNodeLookup()
self.setPosition(self.currentPosition)
return True
def setStartLine(self, edgeno, startLine):
self._hasChanged = True
edge = self[edgeno]
if startLine != edge.startLine and (startLine not in self.nodeLookup or self.nodeLookup[startLine] == edge.startName):
self.nodeLookup.pop(edge.startLine)
edge.startLine = startLine
self.shuttlingGraph.edge[edge.startName][edge.stopName][hash(edge)]['weight'] = abs(edge.stopLine-edge.startLine)
self.rgenerateNodeLookup()
self.graphChangedObservable.firebare()
self.setPosition(self.currentPosition)
return True
return False
def setStopLine(self, edgeno, stopLine):
self._hasChanged = True
edge = self[edgeno]
if stopLine != edge.stopLine and (stopLine not in self.nodeLookup or self.nodeLookup[stopLine] == edge.stopName):
self.nodeLookup.pop(edge.stopLine)
edge.stopLine = stopLine
self.shuttlingGraph.edge[edge.startName][edge.stopName][hash(edge)]['weight'] = abs(edge.stopLine-edge.startLine)
self.rgenerateNodeLookup()
self.graphChangedObservable.firebare()
self.setPosition(self.currentPosition)
return True
return False
def setIdleCount(self, edgeno, idleCount):
self._hasChanged = True
self[edgeno].idleCount = idleCount
return True
def setSteps(self, edgeno, steps):
self._hasChanged = True
self[edgeno].steps = steps
return True
def shuttlePath(self, fromName, toName ):
fromName = firstNotNone(fromName, self.currentPositionName)
fromName = fromName if fromName else self.position(float(self.currentPosition))
if fromName not in self.shuttlingGraph:
raise ShuttlingGraphException("Shuttling failed, origin '{0}' is not a valid shuttling node".format(fromName))
if toName not in self.shuttlingGraph:
raise ShuttlingGraphException("Shuttling failed, target '{0}' is not a valid shuttling node".format(toName))
sp = shortest_path(self.shuttlingGraph, fromName, toName)
path = list()
for a, b in pairs_iter(sp):
edge = sorted(self.shuttlingGraph.edge[a][b].values(), key=itemgetter('weight'))[0]['edge']
path.append((a, b, edge, self.index(edge)))
return path
def nodes(self):
return self.shuttlingGraph.nodes()
def toXmlElement(self, root):
mydict = dict( ( (key, str(getattr(self, key))) for key in ('currentPosition', 'currentPositionName') if getattr(self, key) is not None ) )
myElement = ElementTree.SubElement(root, "ShuttlingGraph", attrib=mydict )
for edge in self:
edge.toXmlElement( myElement )
return myElement
def setStartType(self, edgeno, Type):
self._hasChanged = True
self[edgeno].startType = str(Type)
return True
def setStopType(self, edgeno, Type):
self._hasChanged = True
self[edgeno].stopType = str(Type)
return True
def setStartLength(self, edgeno, length):
edge = self[edgeno]
if length!=edge.startLength:
if length+edge.stopLength<edge.sampleCount:
self._hasChanged = True
edge.startLength = int(length)
else:
return False
return True
def setStopLength(self, edgeno, length):
edge = self[edgeno]
if length!=edge.stopLength:
if edge.startLength+length<edge.sampleCount:
self._hasChanged = True
edge.stopLength = int(length)
else:
return False
return True
@staticmethod
def fromXmlElement( element ):
edgeElementList = element.findall("ShuttleEdge")
edgeList = [ ShuttleEdge.fromXmlElement(e) for e in edgeElementList ]
return ShuttlingGraph(edgeList)
def mif_main(g: MultiGraph, f: set, t, k: int) -> set: k_set = k != None new_k1 = new_k2 = None if k_set and k > g.order(): return None if f == g.nodes() or (k_set and k <= 0): return f if (not f): g_degree = g.degree() g_max_degree_node = max(g_degree, key=lambda n: g_degree[n]) if (g_degree[g_max_degree_node] <= 1): return set(g.nodes()) else: fx = f.copy() fx.add(g_max_degree_node) gx = g.copy() gx.remove_node(g_max_degree_node) if k_set: new_k1 = k-1 new_k2 = k mif_set1 = mif_preprocess_1(g, fx, t, new_k1) mif_set2 = mif_preprocess_1(gx, f, t, new_k2) if not mif_set1: return mif_set2 elif not mif_set2: return mif_set1 else: return max(mif_set1, mif_set2, key=len) # Set t as active vertex if t == None or not t in f: t = next(iter(f)) gd_over_3 = None gd_2 = None for v in g.neighbors_iter(t): (gd_v, gn_v) = generalized_degree(g, f, t, v) if gd_v <= 1: f.add(v) if k_set: new_k1 = k-1 return mif_preprocess_1(g, f, t, new_k1) elif gd_v >= 3: gd_over_3 = v else: gd_2 = (v, gn_v) if gd_over_3 != None: # Cannot simply use "if gd_over_3" because v might be 0 fx = f.copy() fx.add(gd_over_3) gx = g.copy() gx.remove_node(gd_over_3) if k_set: new_k1 = k-1 new_k2 = k mif_set1 = mif_preprocess_1(g, fx, t, new_k1) mif_set2 = mif_preprocess_1(gx, f, t, new_k2) if not mif_set1: return mif_set2 elif not mif_set2: return mif_set1 else: return max(mif_set1, mif_set2, key=len) elif gd_2 != None: (v, gn) = gd_2 fx1 = f.copy() fx2 = f.copy() fx1.add(v) for n in gn: fx2.add(n) gx = g.copy() gx.remove_node(v) if k_set: new_k1 = k-2 new_k2 = k-1 try: cyc.find_cycle(gx.subgraph(fx2)) mif_set1 = None except: mif_set1 = mif_preprocess_1(gx, fx2, t, new_k1) mif_set2 = mif_preprocess_1(g, fx1, t, new_k2) if not mif_set1: return mif_set2 elif not mif_set2: return mif_set1 else: return max(mif_set1, mif_set2, key=len) return None
def fvs_via_mif(g: MultiGraph, k: int) -> set: mif_set = mif(g, g.order()-k) if mif_set: nodes = set(g.nodes()) mif_set = nodes.difference(mif_set) return mif_set
def mif_main(self, g: MultiGraph, f: set, t) -> set:
if f == g.nodes():
return f
if not f:
g_degree = g.degree()
g_max_degree_node = max(g_degree, key=lambda n: n[1])[0]
if g_degree[g_max_degree_node] <= 1:
return set(g.nodes())
else:
fx = f.copy()
fx.add(g_max_degree_node)
gx = g.copy()
gx.remove_node(g_max_degree_node)
mif_set1 = self.preprocess_1(g, fx, t)
mif_set2 = self.preprocess_1(gx, f, t)
if not mif_set1:
return mif_set2
elif not mif_set2:
return mif_set1
else:
return max(mif_set1, mif_set2, key=len)
# Set t as active vertex
if t is None or t not in f:
t = next(iter(f))
gd_over_3 = None
gd_2 = None
for v in g.neighbors(t):
(gd_v, gn_v) = self.generalized_degree(g, f, t, v)
if gd_v <= 1:
f.add(v)
return self.preprocess_1(g, f, t)
elif gd_v >= 3:
gd_over_3 = v
else:
gd_2 = (v, gn_v)
if gd_over_3 is not None:
# Cannot simply use "if gd_over_3" because v might be 0
fx = f.copy()
fx.add(gd_over_3)
gx = g.copy()
gx.remove_node(gd_over_3)
mif_set1 = self.preprocess_1(g, fx, t)
mif_set2 = self.preprocess_1(gx, f, t)
if not mif_set1:
return mif_set2
elif not mif_set2:
return mif_set1
else:
return max(mif_set1, mif_set2, key=len)
elif gd_2 is not None:
(v, gn) = gd_2
fx1 = f.copy()
fx2 = f.copy()
fx1.add(v)
for n in gn:
fx2.add(n)
gx = g.copy()
gx.remove_node(v)
try:
find_cycle(gx.subgraph(fx2))
mif_set1 = None
except:
mif_set1 = self.preprocess_1(gx, fx2, t)
mif_set2 = self.preprocess_1(g, fx1, t)
if not mif_set1:
return mif_set2
elif not mif_set2:
return mif_set1
else:
return max(mif_set1, mif_set2, key=len)
return None
def get_matching_pairs_in_bipartite_graph(self, Gu: nx.MultiGraph, ls, rs, obj_node_id='obj') -> (
List[Tuple], List):
"""
Compares source and target nodes by labels and values and returns the matching pairs, or unmatched pairs
if there are source nodes that can't be matched.
N.b. the presence of a single unmatched Gs node suffices to deem the Gu as unmatched
:param Gu:
:param ls:
:param rs:
:param obj_node_id: identifier for the head node
:return:
"""
NDV = Gu.nodes(data=True)
# Compare and connect nodes in left partition and right partition
# N.b. there could be zero connections in case of mismatch (False Caption for e.g.)
is_head_node = lambda x: obj_node_id in x
Gs_head_nodes = sorted(list(filter(is_head_node, ls)))
Gt_head_nodes = sorted(list(filter(is_head_node, rs)))
logger.debug(f'Number of head nodes in Gs (text graph) = {len(Gs_head_nodes)}')
logger.debug(f'Number of head nodes in Gt (image graph) = {len(Gt_head_nodes)}')
matching_pairs = [] # Holds the (s,t) matching pairs
unmatched_pairs = [] # Holds the (s) source nodes that did not match
for gs_head_node in Gs_head_nodes:
print(f'Matching source head node: {NDV[gs_head_node]}')
gs_hn_val = NDV[gs_head_node]['val']
gs_hn_graph, gs_hn_doc = self.clevr_parser.parse(gs_hn_val)
gs_hn_span = gs_hn_doc.ents[0]
matching_grounding_node = None
# Try matching the src head node to one of the available grounding objects
for gt_head_node in Gt_head_nodes:
# gt = nx.subgraph(Gt, gt_head_node)
# is_equal = __is_equal_head_nodes(gs, gt)
gt_hn_val = NDV[gt_head_node]['val']
if gs_hn_val == gt_hn_val:
matching_grounding_node = gt_head_node
break
gt_hn_graph, gt_hn_doc = self.clevr_parser.parse(gt_hn_val)
gt_hn_span = gt_hn_doc.ents[0]
# Compare <Z>
attr1 = gs_hn_span._.size if gs_hn_span._.has_size else None
attr2 = gt_hn_span._.size
if attr1 is not None and len(attr1) > 0:
_is_equal_attr = self.clevr_parser.entity_recognizer.is_equal_size(attr1, attr2)
if not _is_equal_attr:
continue
# Compare <C>
attr1 = gs_hn_span._.color if gs_hn_span._.has_color else None
attr2 = gt_hn_span._.color
if attr1 and (attr1.text != attr2.text):
# Color is stipulated for source, but doesn't match target
continue
# Compare <M>
attr1 = gs_hn_span._.material if gs_hn_span._.has_material else None
attr2 = gt_hn_span._.material
if attr1 is not None and len(attr1) > 0:
_is_equal_attr = self.clevr_parser.entity_recognizer.is_equal_material(attr1, attr2)
if not _is_equal_attr:
continue
# Compare <S>
attr1 = gs_hn_span._.shape if gs_hn_span._.has_shape else None
attr2 = gt_hn_span._.shape
if attr1 is not None and len(attr1) > 0:
_is_equal_attr = self.clevr_parser.entity_recognizer.is_equal_shape(attr1, attr2)
if not _is_equal_attr:
continue
# Found Grounding Node Match
matching_grounding_node = gt_head_node
break
if matching_grounding_node:
matching_pairs.append((gs_head_node, matching_grounding_node))
else:
unmatched_pairs.append(gs_head_node)
logger.info(f'\tNumber of matching pairs (gs,gt) found = {len(matching_pairs)}'
f'\n\tNumber of unmatched pairs (gs) found = {len(unmatched_pairs)}')
logger.info(f'matching_pairs = {matching_pairs}')
return matching_pairs, unmatched_pairs
def get_embeddings(self, G: nx.MultiGraph, doc, embd_dim=96,
embedding_type=None, **kwargs) -> np.ndarray:
"""
Example:
Text: "There is a green metal block; the tiny metal thing is to the left of it"
Gs -> ['obj', '<C>', '<M>', '<S>', 'obj2', '<Z2>', '<M2>', '<S2>']
doc.ents -> (green metal block, tiny metal thing)
NodeDataView({'obj': {'label': 'CLEVR_OBJ', 'val': 'green metal block'},
'<C>': {'label': 'color', 'val': 'green'},
'<M>': {'label': 'material', 'val': 'metal'},
'<S>': {'label': 'shape', 'val': 'block'},
'obj2': {'label': 'CLEVR_OBJ', 'val': 'tiny metal thing'},
'<Z2>': {'label': 'size', 'val': 'tiny'},
'<M2>': {'label': 'material', 'val': 'metal'},
'<S2>': {'label': 'shape', 'val': 'thing'}})
:param G: MultiGraph containing all CLEVR nodes
:param doc: Spacy Doc
:param embd_dim:
:param embedding_type:
:return: A feature vector matrix corresponding to the value of the Graph of size (N * M) where N is the number
of nodes in the graph and M corresopnds to the embd_sz (default = 96)
Note: the head_node ('obj' node) will be a mean of all attrs vecs (of embd_sz)
"""
import warnings
warnings.warn("Deprecated: use `get_node_feature_matrix` instead",
DeprecationWarning, stacklevel=2)
assert G is not None
NDV = G.nodes(data=True)
NV = G.nodes(data=False)
_is_head_node = lambda x: 'obj' in x
head_nodes = list(filter(_is_head_node, NV))
objs = self.clevr_parser.filter_clevr_objs(doc.ents)
N = len(NDV)
M = embd_dim
feat_mats = []
for i, entity in enumerate(objs):
if entity.label_ not in ('CLEVR_OBJS', 'CLEVR_OBJ'):
continue
ent_mat = self.clevr_parser.get_clevr_entity_matrix_embedding(entity, dim=96, include_obj_node_emd=True)
feat_mats.append(ent_mat)
head_node = G.nodes.get(head_nodes[i])
pos = head_node.get('pos') # pos = (x,y,z): Tuple[float]
if len(feat_mats) > 1:
feat_mats = reduce(lambda a, b: np.vstack((a, b)), feat_mats)
else:
feat_mats = feat_mats[0]
assert feat_mats.shape == (N, M)
spatial_ents = self.clevr_parser.filter_spatial_re(doc.ents)
for i, entity in enumerate(spatial_ents):
ent_vec = entity.vector.reshape(1, -1) #(1, 96)
feat_mats = np.vstack((feat_mats, ent_vec))
matching_ents = self.clevr_parser.filter_matching_re(doc.ents)
for i, entity in enumerate(matching_ents):
ent_vec = entity.vector.reshape(1, -1) # (1, 96)
feat_mats = np.vstack((feat_mats, ent_vec))
return feat_mats
