def test_forward_cpu_graph_invariant(model, data):
atom_data, adj_data = data[0], data[1]
y_actual = cuda.to_cpu(model(atom_data, adj_data).data)
permutation_index = numpy.random.permutation(atom_size)
permute_atom_data = permute_node(atom_data, permutation_index)
permute_adj_data = permute_adj(adj_data, permutation_index)
permute_y_actual = cuda.to_cpu(model(
permute_atom_data, permute_adj_data).data)
assert numpy.allclose(y_actual, permute_y_actual, rtol=1e-5, atol=1e-6)
def test_forward_cpu_graph_invariant(model, data):
atom_data, adj_data = data[0], data[1]
y_actual = cuda.to_cpu(model(atom_data, adj_data).data)
permutation_index = numpy.random.permutation(atom_size)
permute_atom_data = permute_node(atom_data, permutation_index)
permute_adj_data = permute_adj(adj_data, permutation_index)
permute_y_actual = cuda.to_cpu(
model(permute_atom_data, permute_adj_data).data)
assert numpy.allclose(y_actual, permute_y_actual, rtol=1e-5, atol=1e-6)
def test_forward_cpu_graph_invariant(model, data):
# type: (GraphConvPredictor, Tuple[numpy.ndarray, numpy.ndarray, numpy.ndarray]) -> None # NOQA
atom_data, adj_data = data[:2]
y_actual = cuda.to_cpu(model(atom_data, adj_data).data)
permutation_index = numpy.random.permutation(atom_size)
permute_atom_data = permute_node(atom_data, permutation_index)
permute_adj_data = permute_adj(adj_data, permutation_index)
permute_y_actual = cuda.to_cpu(
model(permute_atom_data, permute_adj_data).data)
assert numpy.allclose(y_actual, permute_y_actual, rtol=1e-5, atol=1e-5)
def test_forward_cpu_graph_invariant(model, data):
# type: (GraphConvPredictor, Tuple[numpy.ndarray, numpy.ndarray, numpy.ndarray]) -> None # NOQA
atom_data, adj_data = data[:2]
y_actual = cuda.to_cpu(model(atom_data, adj_data).data)
permutation_index = numpy.random.permutation(atom_size)
permute_atom_data = permute_node(atom_data, permutation_index)
permute_adj_data = permute_adj(adj_data, permutation_index)
permute_y_actual = cuda.to_cpu(
model(permute_atom_data, permute_adj_data).data)
assert numpy.allclose(y_actual, permute_y_actual, rtol=1e-5, atol=1e-5)
def test_forward_cpu_graph_invariant(update, data):
atom_data, adj_data = data[:2]
y_actual = cuda.to_cpu(update(atom_data, adj_data).data)
permutation_index = numpy.random.permutation(atom_size)
permute_atom_data = permute_node(atom_data, permutation_index, axis=1)
permute_adj_data = permute_adj(adj_data, permutation_index)
permute_y_actual = cuda.to_cpu(
update(permute_atom_data, permute_adj_data).data)
numpy.testing.assert_allclose(
permute_node(y_actual, permutation_index, axis=1), permute_y_actual,
rtol=1e-5, atol=1e-5)
def test_forward_cpu_graph_invariant(update, data):
atom_data, adj_data = data[:2]
y_actual = cuda.to_cpu(update(atom_data, adj_data).data)
permutation_index = numpy.random.permutation(atom_size)
permute_atom_data = permute_node(atom_data, permutation_index, axis=1)
permute_adj_data = permute_adj(adj_data, permutation_index)
permute_y_actual = cuda.to_cpu(
update(permute_atom_data, permute_adj_data).data)
numpy.testing.assert_allclose(
permute_node(y_actual, permutation_index, axis=1), permute_y_actual,
rtol=1e-5, atol=1e-5)
def test_forward_cpu_graph_invariant(model, data):
# This RSGCN uses dropout, so we need to forward with test mode
# to remove stochastic calculation.
atom_data, adj_data = data[0], data[1]
with chainer.using_config('train', False):
y_actual = cuda.to_cpu(model(atom_data, adj_data).data)
permutation_index = numpy.random.permutation(atom_size)
permute_atom_data = permute_node(atom_data, permutation_index)
permute_adj_data = permute_adj(adj_data, permutation_index)
with chainer.using_config('train', False):
permute_y_actual = cuda.to_cpu(
model(permute_atom_data, permute_adj_data).data)
assert numpy.allclose(y_actual, permute_y_actual, rtol=1.e-4, atol=1.e-5)
def test_forward_cpu_graph_invariant(model, data):
# This RSGCN uses dropout, so we need to forward with test mode
# to remove stochastic calculation.
atom_data, adj_data = data[0], data[1]
with chainer.using_config('train', False):
y_actual = cuda.to_cpu(model(atom_data, adj_data).data)
permutation_index = numpy.random.permutation(atom_size)
permute_atom_data = permute_node(atom_data, permutation_index)
permute_adj_data = permute_adj(adj_data, permutation_index)
with chainer.using_config('train', False):
permute_y_actual = cuda.to_cpu(model(
permute_atom_data, permute_adj_data).data)
assert numpy.allclose(y_actual, permute_y_actual, rtol=1.e-4, atol=1.e-5)
def test_forward_cpu_graph_invariant(model, data):
# type: (MPNN, Tuple[numpy.ndarray, numpy.ndarray, numpy.ndarray]) -> None
if model.message_func == 'edgenet':
return
# Because EdgeNet uses NN for expanding edge vector dimension,
# graph invariant is not ensured.
atom_data, adj_data = data[0], data[1]
y_actual = cuda.to_cpu(model(atom_data, adj_data).data)
permutation_index = numpy.random.permutation(atom_size)
permute_atom_data = permute_node(atom_data, permutation_index)
permute_adj_data = permute_adj(adj_data, permutation_index)
permute_y_actual = cuda.to_cpu(
model(permute_atom_data, permute_adj_data).data)
assert numpy.allclose(y_actual, permute_y_actual, rtol=1e-3, atol=1e-3)
def test_forward_cpu_graph_invariant(model, data):
# type: (MPNN, Tuple[numpy.ndarray, numpy.ndarray, numpy.ndarray]) -> None
if model.message_func == 'edgenet':
return
# Because EdgeNet uses NN for expanding edge vector dimension,
# graph invariant is not ensured.
atom_data, adj_data = data[0], data[1]
y_actual = cuda.to_cpu(model(atom_data, adj_data).data)
permutation_index = numpy.random.permutation(atom_size)
permute_atom_data = permute_node(atom_data, permutation_index)
permute_adj_data = permute_adj(adj_data, permutation_index)
permute_y_actual = cuda.to_cpu(
model(permute_atom_data, permute_adj_data).data)
assert numpy.allclose(y_actual, permute_y_actual, rtol=1e-3, atol=1e-3)
def test_forward_cpu_graph_invariant(update, data):
# type: (GINUpdate, Tuple[numpy.ndarray, numpy.ndarray, numpy.ndarray]) -> None # NOQA
atom_data, adj_data = data[:2]
y_actual = cuda.to_cpu(update(atom_data, adj_data).data)
permutation_index = numpy.random.permutation(atom_size)
permute_atom_data = permute_node(atom_data, permutation_index, axis=1)
permute_adj_data = permute_adj(adj_data, permutation_index)
permute_y_actual = cuda.to_cpu(
update(permute_atom_data, permute_adj_data).data)
numpy.testing.assert_allclose(permute_node(y_actual,
permutation_index,
axis=1),
permute_y_actual,
rtol=1e-3,
atol=1e-3)
def _test_forward_cpu_graph_invariant(
model, atom_data, adj_data, node_permute_axis=-1):
y_actual = cuda.to_cpu(model(atom_data, adj_data).data)
permutation_index = numpy.random.permutation(atom_size)
permute_atom_data = permute_node(atom_data, permutation_index,
axis=node_permute_axis)
permute_adj_data = adj_data.reshape(
batch_size, atom_size, atom_size, pair_feature_dim
).astype(numpy.float32)
permute_adj_data = permute_adj(
permute_adj_data, permutation_index, axis=[1, 2])
permute_adj_data = permute_adj_data.reshape(
batch_size, atom_size * atom_size, pair_feature_dim
).astype(numpy.float32)
permute_y_actual = cuda.to_cpu(model(
permute_atom_data, permute_adj_data).data)
assert numpy.allclose(y_actual, permute_y_actual, rtol=1.e-4, atol=1.e-6)
def test_forward_cpu_graph_invariant(update, data):
atom_data, adj_data, deg_conds = data[:3]
y_actual = cuda.to_cpu(update(atom_data, adj_data, deg_conds).data)
permutation_index = numpy.random.permutation(atom_size)
# atom_data: (batch_size, atom_size, hidden_channels)
permute_atom_data = permute_node(atom_data, permutation_index, axis=1)
permute_adj_data = permute_adj(adj_data, permutation_index)
# deg_conds: (num_degree_type, batch_size, atom_size, hidden_channels)
permute_deg_conds = permute_node(deg_conds, permutation_index, axis=2)
permute_y_actual = cuda.to_cpu(
update(permute_atom_data, permute_adj_data, permute_deg_conds).data)
numpy.testing.assert_allclose(permute_node(y_actual,
permutation_index,
axis=1),
permute_y_actual,
rtol=1e-5,
atol=1e-5)
def _test_forward_cpu_graph_invariant(model,
atom_data,
adj_data,
node_permute_axis=-1):
y_actual = cuda.to_cpu(model(atom_data, adj_data).data)
permutation_index = numpy.random.permutation(atom_size)
permute_atom_data = permute_node(atom_data,
permutation_index,
axis=node_permute_axis)
permute_adj_data = adj_data.reshape(batch_size, atom_size, atom_size,
pair_feature_dim).astype(numpy.float32)
permute_adj_data = permute_adj(permute_adj_data,
permutation_index,
axis=[1, 2])
permute_adj_data = permute_adj_data.reshape(
batch_size, atom_size * atom_size,
pair_feature_dim).astype(numpy.float32)
permute_y_actual = cuda.to_cpu(
model(permute_atom_data, permute_adj_data).data)
assert numpy.allclose(y_actual, permute_y_actual, rtol=1.e-4, atol=1.e-6)
