in python/dgllife/model/pretrain/moleculenet/freesolv.py [0:0]
def create_freesolv_model(model_name):
"""Create a model.
Parameters
----------
model_name : str
Name for the model.
Returns
-------
Created model
"""
n_tasks = 1
if model_name == 'GCN_canonical_FreeSolv':
num_gnn_layers = 2
dropout = 0.05769700663189804
return GCNPredictor(in_feats=74,
hidden_feats=[32] * num_gnn_layers,
activation=[F.relu] * num_gnn_layers,
residual=[True] * num_gnn_layers,
batchnorm=[False] * num_gnn_layers,
dropout=[dropout] * num_gnn_layers,
predictor_hidden_feats=64,
predictor_dropout=dropout,
n_tasks=n_tasks)
elif model_name == 'GCN_attentivefp_FreeSolv':
num_gnn_layers = 4
dropout = 0.09905316493862346
return GCNPredictor(in_feats=39,
hidden_feats=[32] * num_gnn_layers,
activation=[F.relu] * num_gnn_layers,
residual=[True] * num_gnn_layers,
batchnorm=[False] * num_gnn_layers,
dropout=[dropout] * num_gnn_layers,
predictor_hidden_feats=32,
predictor_dropout=dropout,
n_tasks=n_tasks)
elif model_name == 'GAT_canonical_FreeSolv':
dropout = 0.02327359604429937
return GATPredictor(in_feats=74,
hidden_feats=[256],
num_heads=[4],
feat_drops=[dropout],
attn_drops=[dropout],
alphas=[0.6211392042947481],
residuals=[True],
biases=[False],
predictor_hidden_feats=256,
predictor_dropout=dropout,
n_tasks=n_tasks)
elif model_name == 'GAT_attentivefp_FreeSolv':
dropout = 0.06949846918000477
num_gnn_layers = 2
return GATPredictor(in_feats=39,
hidden_feats=[32] * num_gnn_layers,
num_heads=[8] * num_gnn_layers,
feat_drops=[dropout] * num_gnn_layers,
attn_drops=[dropout] * num_gnn_layers,
alphas=[0.6294479518124414] * num_gnn_layers,
residuals=[True] * num_gnn_layers,
biases=[False] * num_gnn_layers,
predictor_hidden_feats=64,
predictor_dropout=dropout,
n_tasks=n_tasks)
elif model_name == 'Weave_canonical_FreeSolv':
return WeavePredictor(node_in_feats=74,
edge_in_feats=13,
num_gnn_layers=1,
gnn_hidden_feats=64,
graph_feats=64,
gaussian_expand=False,
n_tasks=n_tasks)
elif model_name == 'Weave_attentivefp_FreeSolv':
return WeavePredictor(node_in_feats=39,
edge_in_feats=11,
num_gnn_layers=1,
gnn_hidden_feats=32,
graph_feats=16,
gaussian_expand=False,
n_tasks=n_tasks)
elif model_name == 'MPNN_canonical_FreeSolv':
return MPNNPredictor(node_in_feats=74,
edge_in_feats=13,
node_out_feats=32,
edge_hidden_feats=32,
num_step_message_passing=4,
num_step_set2set=2,
num_layer_set2set=3,
n_tasks=n_tasks)
elif model_name == 'MPNN_attentivefp_FreeSolv':
return MPNNPredictor(node_in_feats=39,
edge_in_feats=11,
node_out_feats=32,
edge_hidden_feats=64,
num_step_message_passing=2,
num_step_set2set=2,
num_layer_set2set=1,
n_tasks=n_tasks)
elif model_name == 'AttentiveFP_canonical_FreeSolv':
return AttentiveFPPredictor(node_feat_size=74,
edge_feat_size=13,
num_layers=4,
num_timesteps=1,
graph_feat_size=32,
dropout=0.07118127568309571,
n_tasks=n_tasks)
elif model_name == 'AttentiveFP_attentivefp_FreeSolv':
return AttentiveFPPredictor(node_feat_size=39,
edge_feat_size=11,
num_layers=1,
num_timesteps=1,
graph_feat_size=128,
dropout=0.1457037675069287,
n_tasks=n_tasks)
elif model_name == 'gin_supervised_contextpred_FreeSolv':
jk = 'concat'
model = GINPredictor(
num_node_emb_list=[120, 3],
num_edge_emb_list=[6, 3],
num_layers=5,
emb_dim=300,
JK=jk,
dropout=0.5,
readout='sum',
n_tasks=n_tasks
)
model.gnn.JK = jk
return model
elif model_name == 'gin_supervised_infomax_FreeSolv':
jk = 'last'
model = GINPredictor(
num_node_emb_list=[120, 3],
num_edge_emb_list=[6, 3],
num_layers=5,
emb_dim=300,
JK=jk,
dropout=0.5,
readout='mean',
n_tasks=n_tasks
)
model.gnn.JK = jk
return model
elif model_name == 'gin_supervised_edgepred_FreeSolv':
jk = 'sum'
model = GINPredictor(
num_node_emb_list=[120, 3],
num_edge_emb_list=[6, 3],
num_layers=5,
emb_dim=300,
JK=jk,
dropout=0.5,
readout='sum',
n_tasks=n_tasks
)
model.gnn.JK = jk
return model
elif model_name == 'gin_supervised_masking_FreeSolv':
jk = 'concat'
model = GINPredictor(
num_node_emb_list=[120, 3],
num_edge_emb_list=[6, 3],
num_layers=5,
emb_dim=300,
JK=jk,
dropout=0.5,
readout='sum',
n_tasks=n_tasks
)
model.gnn.JK = jk
return model
else:
return None