def __init__(self, n_i, n_h, activation=tanh):
self.W_in = theano.shared(initialize_weights(n_i + n_h, n_h))
self.W_h = theano.shared(initialize_weights(n_h, n_h))
self.V_h = theano.shared(initialize_weights(n_h, n_h))
self.W_h_other = theano.shared(initialize_weights(n_h, n_h))
self.params = [self.W_in, self.W_h, self.V_h, self.W_h_other]
self.activation = activation
def build_model(args: Namespace, ddi:bool = False) -> nn.Module:
"""
Builds a MoleculeModel, which is a message passing neural network + feed-forward layers.
:param args: Arguments.
:return: A MoleculeModel containing the MPN encoder along with final linear layers with parameters initialized.
"""
output_size = args.num_tasks
args.output_size = output_size
if args.dataset_type == 'multiclass':
args.output_size *= args.multiclass_num_classes
if args.dataset_type == 'multilabel':
args.output_size = args.num_labels
if not ddi:
model = MoleculeModel(classification=args.dataset_type == 'classification', multiclass=args.dataset_type == 'multiclass')
else:
model = DDIModel(classification=args.dataset_type == 'classification',
multiclass=args.dataset_type == 'multiclass',
multilabel=args.dataset_type == 'multilabel')
if args.jt and args.jt_vocab_file is not None:
vocab = [x.strip("\r\n ") for x in open(args.jt_vocab_file, 'r')]
vocab = Vocab(vocab)
else:
vocab = None
model.create_encoder(args, vocab=vocab)
model.create_ffn(args)
initialize_weights(model)
return model
def build_model(args: Namespace) -> nn.Module:
"""
Builds a MoleculeModel, which is a message passing neural network + feed-forward layers.
:param args: Arguments.
:return: A MoleculeModel containing the MPN encoder along with final linear layers with parameters initialized.
"""
output_size = args.num_tasks
print(f'output_size is {args.num_tasks}')
model = QSARmodel(args,
classification=args.dataset_type == 'classification')
model.create_encoder(args)
model.create_ffn(args)
initialize_weights(model)
print(f'has initialize_weights')
return model
def __init__(self, n_i, n_h, activation=tanh):
self.WA_in = theano.shared(initialize_weights(n_i + n_h, n_h))
self.WA_h = theano.shared(initialize_weights(n_h, n_h))
self.VA_h = theano.shared(initialize_weights(n_h, n_h))
self.WB_in = theano.shared(initialize_weights(n_i + n_h, n_h))
self.WB_h = theano.shared(initialize_weights(n_h, n_h))
self.VB_h = theano.shared(initialize_weights(n_h, n_h))
self.Wother_in = theano.shared(initialize_weights(n_i + n_h, n_h))
self.Wother_h = theano.shared(initialize_weights(n_h, n_h))
self.params = [
self.WA_in, self.WA_h, self.VA_h, self.WB_in, self.WB_h, self.VB_h,
self.Wother_in, self.Wother_h
]
self.activation = activation
def __init__(self, n_i, n_h, activation=tanh):
self.W_xr = theano.shared(initialize_weights(n_i, n_h))
self.W_hr = theano.shared(initialize_weights(n_h, n_h))
self.W_xz = theano.shared(initialize_weights(n_i, n_h))
self.W_hz = theano.shared(initialize_weights(n_h, n_h))
self.W_xh = theano.shared(initialize_weights(n_i, n_h))
self.W_hh = theano.shared(initialize_weights(n_h, n_h))
self.params = [
self.W_xr, self.W_hr, self.W_xz, self.W_hz, self.W_xh, self.W_hh
]
self.activation = activation
def __init__(self, n_h, attention=1):
if attention == 1:
self.W1_c = theano.shared(initialize_weights(n_h, n_h))
self.W2_r = theano.shared(initialize_weights(n_h, n_h))
self.W1_h = theano.shared(initialize_weights(n_h, n_h))
self.w = theano.shared(initialize_weights(n_h, ))
self.f = self.forward
self.params = [self.W1_c, self.W1_h, self.w, self.W2_r]
elif attention == 2:
self.W1_c = theano.shared(initialize_weights(n_h, n_h))
self.W2_r = theano.shared(initialize_weights(n_h, n_h))
self.f = self.forward_bi
self.params = [self.W1_c, self.W2_r]
elif attention == 3:
self.W1_c = theano.shared(initialize_weights(n_h, n_h))
self.W2_r = theano.shared(initialize_weights(2 * n_h, n_h))
self.W1_h = theano.shared(initialize_weights(n_h, n_h))
self.w = theano.shared(initialize_weights(n_h, ))
self.w_b = theano.shared(initialize_weights(n_h, ))
self.f = self.forward_second_order
self.params = [self.W1_c, self.W1_h, self.w, self.w_b, self.W2_r]
else:
self.W1_c = theano.shared(initialize_weights(n_h, n_h))
self.W2_r = theano.shared(initialize_weights(n_h, n_h))
self.W1_h = theano.shared(initialize_weights(n_h, n_h))
self.W_m = theano.shared(initialize_weights(n_h, n_h))
self.w = theano.shared(initialize_weights(n_h, ))
self.f = self.forward_double
self.params = [self.W1_c, self.W1_h, self.w, self.W2_r, self.W_m]
def __init__(self, n_i, n_h):
self.W = theano.shared(initialize_weights(n_i, n_h))
self.params = [self.W]
def __init__(self, n_i, n_h, activation=tanh):
# Parameters for speaker
self.WA_xr = theano.shared(initialize_weights(n_i + n_h, n_h))
self.WA_xp = theano.shared(initialize_weights(n_i + n_h, n_h))
self.WA_xz = theano.shared(initialize_weights(n_i + n_h, n_h))
self.WA_xh = theano.shared(initialize_weights(n_i + n_h, n_h))
self.WA_hr = theano.shared(initialize_weights(n_h, n_h))
self.WA_hp = theano.shared(initialize_weights(n_h, n_h))
self.WA_hz = theano.shared(initialize_weights(n_h, n_h))
self.WA_hh = theano.shared(initialize_weights(n_h, n_h))
self.VA_hr = theano.shared(initialize_weights(n_h, n_h))
self.VA_hp = theano.shared(initialize_weights(n_h, n_h))
self.VA_hz = theano.shared(initialize_weights(n_h, n_h))
self.VA_hh = theano.shared(initialize_weights(n_h, n_h))
# Parameters for addressee
self.WB_xr = theano.shared(initialize_weights(n_i + n_h, n_h))
self.WB_xp = theano.shared(initialize_weights(n_i + n_h, n_h))
self.WB_xz = theano.shared(initialize_weights(n_i + n_h, n_h))
self.WB_xh = theano.shared(initialize_weights(n_i + n_h, n_h))
self.WB_hr = theano.shared(initialize_weights(n_h, n_h))
self.WB_hp = theano.shared(initialize_weights(n_h, n_h))
self.WB_hz = theano.shared(initialize_weights(n_h, n_h))
self.WB_hh = theano.shared(initialize_weights(n_h, n_h))
self.VB_hr = theano.shared(initialize_weights(n_h, n_h))
self.VB_hp = theano.shared(initialize_weights(n_h, n_h))
self.VB_hz = theano.shared(initialize_weights(n_h, n_h))
self.VB_hh = theano.shared(initialize_weights(n_h, n_h))
# Parameters for other
self.Wother_xr = theano.shared(initialize_weights(n_i + n_h, n_h))
self.Wother_xz = theano.shared(initialize_weights(n_i + n_h, n_h))
self.Wother_xh = theano.shared(initialize_weights(n_i + n_h, n_h))
self.Wother_hr = theano.shared(initialize_weights(n_h, n_h))
self.Wother_hz = theano.shared(initialize_weights(n_h, n_h))
self.Wother_hh = theano.shared(initialize_weights(n_h, n_h))
self.params = [
self.WA_xr, self.WA_xp, self.WA_xz, self.WA_xh, self.WA_hr,
self.WA_hp, self.WA_hz, self.WA_hh, self.VA_hr, self.VA_hp,
self.VA_hz, self.VA_hh, self.WB_xr, self.WB_xp, self.WB_xz,
self.WB_xh, self.WB_hr, self.WB_hp, self.WB_hz, self.WB_hh,
self.VB_hr, self.VB_hp, self.VB_hz, self.VB_hh, self.Wother_xr,
self.Wother_xz, self.Wother_xh, self.Wother_hr, self.Wother_hz,
self.Wother_hh
]
self.activation = activation
