def __init__(self, gnn_dim, n_layers, nn_dim):
super(pair_matrix_model, self).__init__()
with self.init_scope():
self.ggnn_gwm = ggnn_gwm(hidden_channels=gnn_dim,
n_update_layers=n_layers)
self.nn1 = GraphMLP(channels=[nn_dim, 5])
self.nn2 = GraphMLP(channels=[nn_dim, 2])
def __init__(self, in_channels=None, hidden_channels=16, out_channels=None,
dropout_ratio=0.5, n_layers=2, **kwargs):
if out_channels is None:
out_channels = hidden_channels
super(GINUpdate, self).__init__()
channels = [hidden_channels] * (n_layers - 1) + [out_channels]
with self.init_scope():
# two Linear + RELU
self.graph_mlp = GraphMLP(
channels=channels, in_channels=in_channels,
activation=functions.relu)
self.dropout_ratio = dropout_ratio
def __init__(self, out_dim, hidden_dim, n_layers, concat_hidden,
weight_tying, nn_hidden_dim, gwm):
super(ggnngwm_stop_step, self).__init__()
with self.init_scope():
self.ggnngwm = GWMGraphConvModel(out_dim=out_dim,
hidden_channels=hidden_dim,
n_update_layers=n_layers,
update_layer=GGNNUpdate,
concat_hidden=concat_hidden,
weight_tying=weight_tying,
with_gwm=gwm,
n_edge_types=5)
# output: binary classification after softmax
self.mlp = GraphMLP(channels=[nn_hidden_dim, 2])
def __init__(self, out_dim, hidden_dim, n_layers, concat_hidden,
weight_tying, nn_hidden_dim, gwm, topK):
super(ggnngwm_atom, self).__init__()
with self.init_scope():
self.ggnngwm = GWMGraphConvModel(out_dim=out_dim,
hidden_channels=hidden_dim,
n_update_layers=n_layers,
update_layer=GGNNUpdate,
concat_hidden=concat_hidden,
weight_tying=weight_tying,
with_gwm=gwm,
n_edge_types=5)
# output: one sigmoid score of each atom
self.mlp = GraphMLP(channels=[nn_hidden_dim, 1])
self.topK = topK
def __init__(self, out_dim, hidden_dim, n_layers, concat_hidden,
weight_tying, nn_hidden_dim, gwm):
super(ggnngwn_action_step, self).__init__()
with self.init_scope():
self.ggnngwm = GWMGraphConvModel(out_dim=out_dim,
hidden_channels=hidden_dim,
n_update_layers=n_layers,
update_layer=GGNNUpdate,
concat_hidden=concat_hidden,
weight_tying=weight_tying,
with_gwm=gwm,
n_edge_types=5)
# output: [remove, single, double, triple] softmax score of each action for one pair
self.mlp = GraphMLP(channels=[nn_hidden_dim, 5 - 1]) # no aromatic
self.out_dim = out_dim