def __init__(self, out_dim=1, hidden_dim=32):
super(SchNetReadout, self).__init__()
with self.init_scope():
self.linear1 = GraphLinear(hidden_dim)
self.linear2 = GraphLinear(out_dim)
self.out_dim = out_dim
self.hidden_dim = hidden_dim
def __init__(self, out_dim=1, in_channels=None, hidden_channels=32):
super(SchNetReadout, self).__init__()
with self.init_scope():
self.linear1 = GraphLinear(in_channels, hidden_channels)
self.linear2 = GraphLinear(hidden_channels, out_dim)
self.out_dim = out_dim
self.hidden_dim = in_channels
def __init__(self, hidden_dim=16, dropout_ratio=0.5):
super(GINUpdate, self).__init__()
with self.init_scope():
# two Linear + RELU
self.linear_g1 = GraphLinear(hidden_dim, hidden_dim)
self.linear_g2 = GraphLinear(hidden_dim, hidden_dim)
# end with
self.dropout_ratio = dropout_ratio
def __init__(self, in_channels, out_channels, n_edge_types=4, **kwargs):
super(RelGCNUpdate, self).__init__()
with self.init_scope():
self.graph_linear_self = GraphLinear(in_channels, out_channels)
self.graph_linear_edge = GraphLinear(in_channels,
out_channels * n_edge_types)
self.n_edge_types = n_edge_types
self.in_channels = in_channels
self.out_channels = out_channels
def __init__(self, in_channels, out_channels, num_edge_type=4):
super(RelGCNUpdate, self).__init__()
with self.init_scope():
self.graph_linear_self = GraphLinear(in_channels, out_channels)
self.graph_linear_edge = GraphLinear(in_channels,
out_channels * num_edge_type)
self.num_edge_type = num_edge_type
self.in_channels = in_channels
self.out_channels = out_channels
def __init__(self, in_channels=None, hidden_channels=16, out_channels=None,
dropout_ratio=0.5, **kwargs):
if out_channels is None:
out_channels = hidden_channels
super(GINUpdate, self).__init__()
with self.init_scope():
# two Linear + RELU
self.linear_g1 = GraphLinear(in_channels, hidden_channels)
self.linear_g2 = GraphLinear(hidden_channels, out_channels)
self.dropout_ratio = dropout_ratio
def __init__(self, out_dim, in_channels=None, nobias=False,
activation=functions.identity,
activation_agg=functions.identity):
super(GGNNReadout, self).__init__()
with self.init_scope():
self.i_layer = GraphLinear(in_channels, out_dim, nobias=nobias)
self.j_layer = GraphLinear(in_channels, out_dim, nobias=nobias)
self.out_dim = out_dim
self.in_channels = in_channels
self.nobias = nobias
self.activation = activation
self.activation_agg = activation_agg
def __init__(self,
out_dim,
hidden_dim=16,
nobias=False,
activation=functions.identity,
activation_agg=functions.identity):
super(GINReadout, self).__init__()
with self.init_scope():
self.i_layer = GraphLinear(None, out_dim, nobias=nobias)
self.j_layer = GraphLinear(None, out_dim, nobias=nobias)
self.out_dim = out_dim
self.hidden_dim = hidden_dim
self.nobias = nobias
self.activation = activation
self.activation_agg = activation_agg
def __init__(self, hidden_dim=64):
super(SchNetUpdate, self).__init__()
with self.init_scope():
self.linear = chainer.ChainList(
*[GraphLinear(hidden_dim) for _ in range(3)])
self.cfconv = CFConv(hidden_dim=hidden_dim)
self.hidden_dim = hidden_dim
def __init__(self, hidden_dim=16, num_edge_type=4):
super(GGNNUpdate, self).__init__()
with self.init_scope():
self.graph_linear = GraphLinear(hidden_dim,
num_edge_type * hidden_dim)
self.update_layer = links.GRU(2 * hidden_dim, hidden_dim)
self.num_edge_type = num_edge_type
def __init__(self,
out_channels=64,
num_edge_type=4,
ch_list=None,
n_atom_types=MAX_ATOMIC_NUM,
input_type='int',
scale_adj=None):
super(RelGCN, self).__init__()
if ch_list is None:
ch_list = [16, 128, 64]
with self.init_scope():
if input_type == 'int':
self.embed = EmbedAtomID(out_size=ch_list[0],
in_size=n_atom_types)
elif input_type == 'float':
self.embed = GraphLinear(None, ch_list[0])
else:
raise ValueError(
"[ERROR] Unexpected value input type={}".format(
input_type))
self.rgcn_convs = chainer.ChainList(*[
RelGCNUpdate(ch_list[i], ch_list[i + 1], num_edge_type)
for i in range(len(ch_list) - 1)
])
self.rgcn_readout = GGNNReadout(out_dim=out_channels,
hidden_dim=ch_list[-1],
nobias=True,
activation=functions.tanh)
self.input_type = input_type
self.scale_adj = scale_adj
def __init__(self, in_channels, out_channels, n_heads=3, n_edge_types=4,
dropout_ratio=-1., negative_slope=0.2, softmax_mode='across',
concat_heads=False):
super(RelGATUpdate, self).__init__()
with self.init_scope():
self.message_layer = GraphLinear(
in_channels, out_channels * n_edge_types * n_heads)
self.attention_layer = GraphLinear(out_channels * 2, 1)
self.in_channels = in_channels
self.out_channels = out_channels
self.n_heads = n_heads
self.n_edge_types = n_edge_types
self.dropout_ratio = dropout_ratio
self.softmax_mode = softmax_mode
self.concat_heads = concat_heads
self.negative_slope = negative_slope
def __init__(self,
hidden_channels=16,
n_edge_types=5,
activation=functions.relu):
super(GNNFiLMUpdate, self).__init__()
self.n_edge_types = n_edge_types
self.activation = activation
with self.init_scope():
self.W_linear = GraphLinear(in_size=None,
out_size=self.n_edge_types *
hidden_channels,
nobias=True) # W_l in eq. (6)
self.W_g = GraphLinear(in_size=None,
out_size=self.n_edge_types *
hidden_channels * 2,
nobias=True) # g in eq. (6)
self.norm_layer = links.LayerNormalization() # l in eq. (6)
def __init__(self, in_channels, out_channels, **kwargs):
super(RSGCNUpdate, self).__init__()
with self.init_scope():
self.graph_linear = GraphLinear(in_channels,
out_channels,
nobias=True)
self.in_channels = in_channels
self.out_channels = out_channels
def __init__(self, hidden_dim=64, num_rbf=300, radius_resolution=0.1,
gamma=10.0):
super(SchNetUpdate, self).__init__()
with self.init_scope():
self.linear = chainer.ChainList(
*[GraphLinear(hidden_dim) for _ in range(3)])
self.cfconv = CFConv(
num_rbf=num_rbf, radius_resolution=radius_resolution,
gamma=gamma, hidden_dim=hidden_dim)
self.hidden_dim = hidden_dim
def __init__(self, in_channels, out_channels, max_degree=6):
super(NFPUpdate, self).__init__()
num_degree_type = max_degree + 1
with self.init_scope():
self.graph_linears = chainer.ChainList(*[
GraphLinear(in_channels, out_channels)
for _ in range(num_degree_type)
])
self.max_degree = max_degree
self.in_channels = in_channels
self.out_channels = out_channels
def __init__(self, channels, in_channels=None, activation=relu):
super(GraphMLP, self).__init__()
if not isinstance(channels, (list, numpy.ndarray)):
raise TypeError('channels {} is expected to be list, actual {}'
.format(channels, type(channels)))
channels_list = [in_channels] + list(channels)
layers = [GraphLinear(channels_list[i], channels_list[i+1])
for i in range(len(channels_list) - 1)]
with self.init_scope():
self.layers = chainer.ChainList(*layers)
self.activation = activation
def __init__(self,
hidden_dim=16,
n_layers=4,
n_atom_types=MAX_ATOMIC_NUM,
num_edge_type=4,
weight_tying=True):
super(GGNNUpdate, self).__init__()
n_layer = 1 if weight_tying else n_layers
with self.init_scope():
self.graph_linears = chainer.ChainList(*[
GraphLinear(hidden_dim, num_edge_type * hidden_dim)
for _ in range(n_layer)
])
self.update_layer = links.GRU(2 * hidden_dim, hidden_dim)
self.n_layers = n_layers
self.num_edge_type = num_edge_type
self.weight_tying = weight_tying
def __init__(self, in_channels=None, hidden_channels=16,
out_channels=None, n_edge_types=4, **kwargs):
if out_channels is None:
out_channels = hidden_channels
super(GGNNUpdate, self).__init__()
if in_channels is None:
gru_in_channels = None
else:
gru_in_channels = in_channels + hidden_channels
with self.init_scope():
self.graph_linear = GraphLinear(
in_channels, n_edge_types * hidden_channels)
self.update_layer = links.GRU(gru_in_channels, out_channels)
self.n_edge_types = n_edge_types
self.in_channels = in_channels
self.hidden_channels = hidden_channels
self.out_channels = out_channels
def __init__(self,
hidden_channels,
out_dim,
update_layer,
readout_layer,
n_update_layers=None,
out_channels=None,
wle_dim=None,
n_atom_types=MAX_ATOMIC_NUM,
n_edge_types=4,
dropout_ratio=-1.0,
with_wle=True,
concat_hidden=False,
sum_hidden=False,
weight_tying=False,
scale_adj=False,
activation=None,
use_batchnorm=False,
n_activation=None,
update_kwargs=None,
readout_kwargs=None,
wle_kwargs=None,
n_wle_types=MAX_WLE_NUM):
super(GWLEGraphConvModel, self).__init__()
# General: length of hidden_channels must be n_layers + 1
if isinstance(hidden_channels, int):
if n_update_layers is None:
raise ValueError('n_update_layers is None')
else:
hidden_channels = [
hidden_channels for _ in range(n_update_layers + 1)
]
elif isinstance(hidden_channels, list):
if out_channels is None:
n_update_layers = len(hidden_channels) - 1
else:
n_update_layers = len(hidden_channels)
else:
raise TypeError('Unexpected value for hidden_channels {}'.format(
hidden_channels))
if readout_layer == GeneralReadout and hidden_channels[-1] != out_dim:
# When use GWM, hidden channels must be same. But GeneralReadout
# cannot change the dimension. So when use General Readout and GWM,
# hidden channel and out_dim should be same.
if with_wle:
raise ValueError('Unsupported combination.')
else:
hidden_channels[-1] = out_dim
# When use with_gwm, concat_hidden, sum_hidden and weight_tying option,
# hidden_channels must be same
if with_wle or concat_hidden or sum_hidden or weight_tying:
if not all(
[in_dim == hidden_channels[0] for in_dim in hidden_channels]):
raise ValueError(
'hidden_channels must be same but different {}'.format(
hidden_channels))
if with_wle and wle_dim is None:
print('[WARNING] wle_dim is None, set to {}'.format(
hidden_channels[0]))
wle_dim = hidden_channels[0]
if out_channels is None:
in_channels_list = hidden_channels[:-1]
out_channels_list = hidden_channels[1:]
else:
# For RelGAT concat_heads option
in_channels_list = hidden_channels
out_channels_list = out_channels
assert len(in_channels_list) == n_update_layers
assert len(out_channels_list) == n_update_layers
n_use_update_layers = 1 if weight_tying else n_update_layers
n_readout_layers = n_use_update_layers if concat_hidden or sum_hidden else 1
n_activation = n_use_update_layers if n_activation is None else n_activation
if update_kwargs is None:
update_kwargs = {}
if readout_kwargs is None:
readout_kwargs = {}
if wle_kwargs is None:
wle_kwargs = {}
with self.init_scope():
self.embed = EmbedAtomID(out_size=hidden_channels[0],
in_size=n_atom_types) # +1 for label 0
self.update_layers = chainer.ChainList(*[
update_layer(in_channels=in_channels_list[i],
out_channels=out_channels_list[i],
n_edge_types=n_edge_types,
**update_kwargs)
for i in range(n_use_update_layers)
])
# when use weight_tying option, hidden_channels must be same. So we can use -1 index
self.readout_layers = chainer.ChainList(*[
readout_layer(
out_dim=out_dim,
# in_channels=hidden_channels[-1],
in_channels=None,
**readout_kwargs) for _ in range(n_readout_layers)
])
if with_wle:
self.embed_wle = links.EmbedID(out_size=wle_dim,
in_size=n_wle_types)
# Gates
self.gate_W1 = GraphLinear(in_size=hidden_channels[0],
out_size=hidden_channels[0])
self.gate_W2 = GraphLinear(in_size=wle_dim,
out_size=hidden_channels[0])
if use_batchnorm:
self.bnorms = chainer.ChainList(*[
GraphBatchNormalization(out_channels_list[i])
for i in range(n_use_update_layers)
])
self.readout_layer = readout_layer
self.update_layer = update_layer
self.weight_tying = weight_tying
self.with_wle = with_wle
self.concat_hidden = concat_hidden
self.sum_hidden = sum_hidden
self.scale_adj = scale_adj
self.activation = activation
self.dropout_ratio = dropout_ratio
self.use_batchnorm = use_batchnorm
self.n_activation = n_activation
self.n_update_layers = n_update_layers
self.n_edge_types = n_edge_types
def __init__(self, in_channels, out_size):
super(NFPReadout, self).__init__()
with self.init_scope():
self.output_weight = GraphLinear(in_channels, out_size)
self.in_channels = in_channels
self.out_size = out_size
def model():
l = GraphLinear(in_size=in_size, out_size=out_size)
l.cleargrads()
return l
