def test_to_networkx():
graph = Graph({
'!i': [0, 1, 2],
'open': [True, False, True]
}, {
'!i': [0, 0],
'!j': [1, 2],
'length': [0.5, 1.5],
'label': ['x', 'y']
},
title='test')
ngraph = graph.to_networkx()
assert (len(ngraph.nodes) == len(graph.nodes))
assert (len(ngraph.edges) == len(graph.edges))
for _, n in ngraph.nodes.items():
assert ('open' in n)
for _, e in ngraph.edges.items():
assert ('length' in e)
assert ('label' in e)
for k in range(len(graph.nodes)):
i = graph.nodes['!i'][k]
assert (graph.nodes['open'][k] == ngraph.nodes[i]['open'])
for k in range(len(graph.edges)):
i = graph.edges['!i'][k]
j = graph.edges['!j'][k]
assert (graph.edges['length'][k] == ngraph[i][j]['length'])
assert (graph.edges['label'][k] == ngraph[i][j]['label'])
def __call__(self, atoms1, atoms2):
args = dict(use_charge=self.use_charge, adjacency=self.adjacency)
g1 = Graph.from_ase(atoms1, **args)
g2 = Graph.from_ase(atoms2, **args)
R1 = self._mlgk(g1, g1).diagonal()**-0.5
R2 = self._mlgk(g2, g2).diagonal()**-0.5
R12 = self._mlgk(g1, g2)
K = R1[:, None] * R12 * R2[None, :]
D = np.sqrt(np.maximum(2 - 2 * K, 0))
return max(D.min(axis=1).max(), D.min(axis=0).max())
def test_graph_cache_invalidation(inplace):
G = [Graph.from_networkx(nx.cycle_graph(4)) for _ in range(2)]
backend = CUDABackend()
cached = [backend._register_graph(g) for g in G]
if inplace:
Graph.unify_datatype(G, inplace=inplace)
H = G
else:
H = Graph.unify_datatype(G, inplace=inplace)
for h in H:
_, h_cached = backend._register_graph(h)
for _, g_cached in cached:
assert (h_cached is not g_cached)
def test_dict_init():
G = Graph(nodes={
'order': {
0: 1,
1: -2
},
'conjugate': {
0: True,
1: False
}
},
edges={
'!ij': {
0: (0, 1)
},
'length': {
0: 3.2
},
'weight': {
0: 1
}
},
title='graph')
for g in [G, eval(repr(G).strip('><'))]:
assert (g.title == 'graph')
assert (len(g.nodes) == 2)
assert (len(g.nodes.columns) == 2)
assert (len(g.edges) == 1)
assert (len(g.edges.columns) == 3)
def test_from_rdkit_linear_hydrocarbon():
for i in range(2, 10):
smi = 'C' * i
mol = Chem.MolFromSmiles(smi)
g = Graph.from_rdkit(mol)
assert(len(g.nodes) == i)
assert(len(g.edges) == i - 1)
def test_permute(inplace):
nxg = nx.Graph(title='Simple')
f = [np.pi, np.e, np.sqrt(2), -1.0, 2.0]
nxg.add_node(0, f=f[0])
nxg.add_node(1, f=f[1])
nxg.add_node(2, f=f[2])
nxg.add_node(3, f=f[3])
nxg.add_node(4, f=f[4])
nxg.add_edge(0, 1, h=f[0] * f[1])
nxg.add_edge(0, 2, h=f[0] * f[2])
nxg.add_edge(0, 4, h=f[0] * f[4])
nxg.add_edge(1, 2, h=f[1] * f[2])
nxg.add_edge(1, 3, h=f[1] * f[3])
nxg.add_edge(2, 3, h=f[2] * f[3])
nxg.add_edge(3, 4, h=f[3] * f[4])
g = Graph.from_networkx(nxg)
for perm in it.permutations(range(len(g.nodes))):
g1 = g.copy(deep=True)
g2 = g1.permute(perm, inplace=inplace)
if inplace:
assert (g1 is g2)
m = {idx: i for i, idx in enumerate(g2.nodes['!i'])}
for i in range(len(g2.edges)):
i1 = m[g2.edges['!i'][i]]
i2 = m[g2.edges['!j'][i]]
assert (g2.edges.h[i] == pytest.approx(g2.nodes.f[i1] *
g2.nodes.f[i2]))
def test_copy(deep):
nxg = nx.Graph(title='Simple')
nxg.add_node(0, f=0, g='a')
nxg.add_node(1, f=1, g='b')
nxg.add_node(2, f=2, g='c')
nxg.add_edge(0, 1, h=1.0)
nxg.add_edge(0, 2, h=2.0)
g1 = Graph.from_networkx(nxg)
g1.custom_attributes = (1.5, 'hello', False)
g2 = g1.copy(deep=deep)
assert (hasattr(g2, 'custom_attributes'))
assert (g1.custom_attributes == g2.custom_attributes)
assert (g1.title == g2.title)
assert (len(g1.nodes) == len(g2.nodes))
assert (len(g1.edges) == len(g2.edges))
for n1, n2 in zip(g1.nodes.rows(), g2.nodes.rows()):
assert (n1 == n2)
for e1, e2 in zip(g1.edges.rows(), g2.edges.rows()):
assert (e1 == e2)
# Changes in one graph should reflect in its shallow copy, but should not
# affect its deep copy.
for i in range(len(g1.nodes)):
g1.nodes.f[i] += 10.0
for n1, n2 in zip(g1.nodes.rows(), g2.nodes.rows()):
if deep:
assert (n1 != n2)
else:
assert (n1 == n2)
def test_rcm_fancy_graphs(n, gen):
nxg = gen(n)
if nxg.number_of_edges() > 0:
g = Graph.from_networkx(nxg)
p = rcm(g)
assert (np.min(p) == 0)
assert (np.max(p) == n - 1)
assert (len(np.unique(p)) == n)
def test_attribute_consistency_from_networkx():
nxg1 = nx.Graph(title='H2O')
nxg1.add_node(0, label1=3.14)
nxg1.add_node(1, label2=True)
with pytest.raises(TypeError):
Graph.from_networkx(nxg1)
nxg2 = nx.Graph(title='H2O')
nxg2.add_node(0)
nxg2.add_node(1)
nxg2.add_node(2)
nxg2.add_edge(0, 1, weight=1)
nxg2.add_edge(0, 2, multi=2)
with pytest.raises(TypeError):
Graph.from_networkx(nxg2)
def test_empty_init():
G = Graph(nodes=[], edges=[])
for g in [G, eval(repr(G).strip('><'))]:
assert (g.title == '')
assert (len(g.nodes) == 0)
assert (len(g.nodes.columns) == 0)
assert (len(g.edges) == 0)
assert (len(g.edges.columns) == 0)
def test_graph_cache():
G = [Graph.from_networkx(nx.cycle_graph(4)) for _ in range(2)]
backend = CUDABackend()
i1, cached1 = backend._register_graph(G[0])
i2, cached2 = backend._register_graph(G[0])
i3, cached3 = backend._register_graph(G[1])
assert (i1 == i2)
assert (cached1 is cached2)
assert (i1 != i3)
assert (cached1 is not cached3)
def test_rcm_fancy_graphs(n, gen):
nxg = gen(n)
if nxg.number_of_edges() > 0:
g = Graph.from_networkx(nxg)
p = pbr(g)
assert(np.min(p) == 0)
assert(np.max(p) == n - 1)
assert(len(np.unique(p)) == n)
g_perm = g.permute(p)
assert(n_tiles(g.adjacency_matrix) >= n_tiles(g_perm.adjacency_matrix))
def test_empty_from_networkx():
nxg = nx.Graph(title='Null')
G = Graph.from_networkx(nxg)
for g in [G, eval(repr(G).strip('><'))]:
assert (g.title == 'Null')
assert (len(g.nodes) == 0)
assert (len(g.nodes.columns) == 0)
assert (len(g.edges) == 0)
assert (len(g.edges.columns) == 1) # +1 for the hidden edge index
def test_adjacency_matrix_full():
for n in [2, 3, 5, 7, 8, 11, 13, 16, 17, 23, 25]:
g = Graph.from_networkx(nx.complete_graph(n))
A = g.adjacency_matrix.todense()
assert (A.shape[0] == n)
assert (A.shape[1] == n)
for i in range(n):
for j in range(n):
if i == j:
assert (A[i, j] == 0)
else:
assert (A[i, j] == 1)
def test_adjacency_matrix_regular():
for d in [2, 3, 4]:
for n in [8, 12, 16]:
g = Graph.from_networkx(nx.random_regular_graph(d, n))
A = g.adjacency_matrix.todense()
assert (A.shape[0] == n)
assert (A.shape[1] == n)
D = A.sum(axis=0)
for deg in D.flat:
assert (deg == d)
for i in range(n):
for j in range(n):
assert (A[i, j] == A[j, i])
def test_weighted_from_networkx():
nxg = nx.Graph(title='Simple')
nxg.add_node(0)
nxg.add_node(1)
nxg.add_edge(0, 1, w=1.0)
G = Graph.from_networkx(nxg, weight='w')
for g in [G, eval(repr(G).strip('><'))]:
assert (g.title == 'Simple')
assert (len(g.nodes) == 2)
assert (len(g.nodes.columns) == 0)
assert (len(g.edges) == 1)
assert (len(g.edges.columns) == 2) # +2 for edge index and weight
assert ('!ij' in g.edges.columns)
assert ('!w' in g.edges.columns)
def test_molecular_kernel_on_organics(benchmark):
graphs = [
Graph.from_ase(molecule(name)) for name in g2.names
if len(molecule(name)) > 1
]
kernel = Tang2019MolecularKernel()
def fun(kernel, graphs):
return kernel(graphs, nodal=True)
benchmark.pedantic(fun,
args=(kernel, graphs),
iterations=3,
rounds=3,
warmup_rounds=1)
def test_molecule_from_networkx():
nxg = nx.Graph(title='H2O')
nxg.add_node('O1', charge=1, conjugate=False, mass=8.0)
nxg.add_node('H1', charge=-1, conjugate=True, mass=1.0)
nxg.add_node('H2', charge=2, conjugate=True, mass=1.0)
nxg.add_edge('O1', 'H1', order=1, length=0.5, weight=3.2)
nxg.add_edge('O1', 'H2', order=2, length=1.0, weight=0.5)
G = Graph.from_networkx(nxg)
for g in [G, eval(repr(G).strip('><'))]:
assert (g.title == 'H2O')
assert (len(g.nodes) == 3)
assert (len(g.nodes.columns) == 3)
assert (len(g.edges) == 2)
assert (len(g.edges.columns) == 3 + 1) # +1 for the hidden edge index
def test_simple_from_networkx():
nxg = nx.Graph(title='Simple')
nxg.add_node(0)
nxg.add_node('c')
nxg.add_node('X')
nxg.add_node(3.14)
nxg.add_edge(0, 'c')
nxg.add_edge(3.14, 'X')
G = Graph.from_networkx(nxg)
for g in [G, eval(repr(G).strip('><'))]:
assert (g.title == 'Simple')
assert (len(g.nodes) == 4)
assert (len(g.nodes.columns) == 0)
assert (len(g.edges) == 2)
assert (len(g.edges.columns) == 1) # +1 for the hidden edge index
def test_adjacency_matrix_cycle():
for n in [2, 3, 5, 7, 8, 11, 13, 16, 17, 23, 25]:
g = Graph.from_networkx(nx.cycle_graph(n))
A = g.adjacency_matrix.todense()
assert (A.shape[0] == n)
assert (A.shape[1] == n)
for i in range(n):
for j in range(n):
d = i - j
if d > n / 2:
d -= n
elif d < -n / 2:
d += n
if abs(d) == 1:
assert (A[i, j] == 1)
else:
assert (A[i, j] == 0)
def test_large_from_networkx():
nxg = nx.Graph(title='Large')
nn = 1000
ne = 100000
np.random.seed(0)
for i in range(nn):
nxg.add_node(i, label=np.random.randn())
for _ in range(ne):
i, j = np.random.randint(nn, size=2)
nxg.add_edge(i, j, weight=np.random.rand())
G = Graph.from_networkx(nxg)
for g in [G, eval(repr(G).strip('><'))]:
assert (g.title == 'Large')
assert (len(g.nodes) == nn)
assert (len(g.edges) <= ne)
def test_from_rdkit_options():
m = Chem.MolFromSmiles('CCOCC')
g = Graph.from_rdkit(m, bond_type='order')
assert('order' in g.edges.columns)
g = Graph.from_rdkit(m, bond_type='type')
assert('type' in g.edges.columns)
g = Graph.from_rdkit(m, set_ring_list=True)
assert('ring_list' in g.nodes.columns)
g = Graph.from_rdkit(m, set_ring_list=False)
assert('ring_list' not in g.nodes.columns)
g = Graph.from_rdkit(m, set_ring_stereo=True)
assert('ring_stereo' in g.edges.columns)
g = Graph.from_rdkit(m, set_ring_stereo=False)
assert('ring_stereo' not in g.edges.columns)
def test_dict_init():
G = Graph(nodes={
'!i': [0, 1],
'order': [1, -2],
'conjugate': [True, False]
},
edges={
'!i': [0],
'!j': [1],
'length': [3.2],
'weight': [1]
},
title='graph')
for g in [G, eval(repr(G).strip('><'))]:
assert (g.title == 'graph')
assert (len(g.nodes) == 2)
assert (len(g.nodes.columns) == 3) # +1 for the hidden !i index
assert (len(g.edges) == 1)
assert (len(g.edges.columns) == 4) # +2 for the hidden !i, !j index
def __call__(self, X, ij, lmin=0, timing=False):
"""Compute a list of pairwise similarities between graphs.
Parameters
----------
X: list of N graphs
The graphs must all have same node and edge attributes.
ij: list of pairs of ints
Pair indices for which the graph kernel is to be evaluated.
lmin: 0 or 1
Number of steps to skip in each random walk path before similarity
is computed.
(lmin + 1) corresponds to the starting value of l in the summation
of Eq. 1 in Tang & de Jong, 2019 https://doi.org/10.1063/1.5078640
(or the first unnumbered equation in Section 3.3 of Kashima, Tsuda,
and Inokuchi, 2003).
Returns
-------
gramian: ndarray
A vector with the same length as ij
"""
timer = Timer()
backend = self.backend
traits = self.traits(lmin=lmin, )
''' assert graph attributes are compatible with each other '''
pred_or_tuple = Graph.has_unified_types(X)
if pred_or_tuple is not True:
group, first, second = pred_or_tuple
raise TypeError(
f'The two graphs have mismatching {group} attributes or '
'attribute types. If the attributes match in name but differ '
'in type, try `Graph.unify_datatype` as an automatic fix.\n'
f'First graph: {first}\n'
f'Second graph: {second}\n')
''' generate jobs '''
timer.tic('generating jobs')
jobs = backend.array(
np.array(ij, dtype=np.uint32).ravel().view(
dtype=np.dtype([('i', np.uint32), ('j', np.uint32)])))
timer.toc('generating jobs')
''' create output buffer '''
timer.tic('creating output buffer')
gramian = backend.empty(len(jobs), np.float32)
timer.toc('creating output buffer')
''' call GPU kernel '''
timer.tic('calling GPU kernel (overall)')
backend(
X,
self.node_kernel,
self.edge_kernel,
self.p,
self.q,
self.eps,
self.ftol,
self.gtol,
jobs,
gramian,
traits,
timer,
)
timer.toc('calling GPU kernel (overall)')
''' collect result '''
timer.tic('collecting result')
gramian = gramian.astype(self.element_dtype)
timer.toc('collecting result')
if timing:
timer.report(unit='ms')
timer.reset()
return gramian
def test_from_rdkit_feature_atom_aromatic(testset):
smi, aromatic = testset
g = Graph.from_rdkit(Chem.MolFromSmiles(smi))
for n in g.nodes.rows():
assert(n.aromatic == aromatic)
def test_from_rdkit_feature_atom_hybridization(testset):
smi, hybridization = testset
g = Graph.from_rdkit(Chem.MolFromSmiles(smi))
for n in g.nodes.rows():
assert(n.hybridization == hybridization)
def test_from_rdkit_feature_atom_hcount(testset):
smi, hcount = testset
g = Graph.from_rdkit(Chem.MolFromSmiles(smi))
for n in g.nodes.rows():
assert(n.hcount == hcount)
def test_from_rdkit_feature_atom_charge(testset):
smi, charge = testset
g = Graph.from_rdkit(Chem.MolFromSmiles(smi))
assert(g.nodes.charge[0] == charge)
def test_from_rdkit_feature_atomic_number(testset):
smi, atomic_number = testset
g = Graph.from_rdkit(Chem.MolFromSmiles(smi))
for n in g.nodes.rows():
assert(n.atomic_number == atomic_number)
def test_from_rdkit_feature_bond_stereo(testset):
smi, stereo = testset
g = Graph.from_rdkit(Chem.MolFromSmiles(smi))
assert(g.edges.stereo[1] == stereo)