def __init__(self,
n_in,
n_out,
n_hidden=None,
n_layers=2,
activation=shifted_softplus):
super(MLP, self).__init__()
# If no neurons are given, initialize
if n_hidden is None:
c_neurons = n_in
self.n_neurons = []
for i in range(n_layers):
self.n_neurons.append(c_neurons)
c_neurons = c_neurons // 2
self.n_neurons.append(n_out)
else:
if type(n_hidden) is int:
n_hidden = [n_hidden] * (n_layers - 1)
self.n_neurons = [n_in] + n_hidden + [n_out]
layers = [
Dense(self.n_neurons[i],
self.n_neurons[i + 1],
activation=activation) for i in range(n_layers - 1)
]
layers.append(
Dense(self.n_neurons[-2], self.n_neurons[-1], activation=None))
self.out_net = nn.Sequential(*layers)
def __init__(self,
n_in,
n_out,
n_hidden=None,
n_layers=2,
activation=shifted_softplus):
super(MLP, self).__init__()
# get list of number of nodes in input, hidden & output layers
if n_hidden is None:
c_neurons = n_in
self.n_neurons = []
for i in range(n_layers):
self.n_neurons.append(c_neurons)
c_neurons = c_neurons // 2
self.n_neurons.append(n_out)
else:
# get list of number of nodes hidden layers
if type(n_hidden) is int:
n_hidden = [n_hidden] * (n_layers - 1)
self.n_neurons = [n_in] + n_hidden + [n_out]
# assign a Dense layer (with activation function) to each hidden layer
layers = [
Dense(self.n_neurons[i],
self.n_neurons[i + 1],
activation=activation) for i in range(n_layers - 1)
]
# assign a Dense layer (without activation function) to the output layer
layers.append(
Dense(self.n_neurons[-2], self.n_neurons[-1], activation=None))
# put all layers together to make the network
self.out_net = nn.Sequential(*layers)
def __init__(
self,
n_atom_basis,
n_hidden,
n_heads=8,
activation=None,
):
super(EgoAttention, self).__init__()
# dense layer as mh_attention
assert (n_atom_basis % n_heads == 0), "Mismatch Head Numbers."
n_per_head = n_atom_basis // n_heads
self.n_heads = n_heads
self.n_per_head = n_per_head
self.mh_q = Dense(n_atom_basis,
n_atom_basis,
bias=False,
activation=None)
self.mh_k = Dense(n_atom_basis,
n_atom_basis,
bias=False,
activation=None)
self.mh_v = Dense(n_atom_basis,
n_atom_basis,
bias=False,
activation=None)
self.mh_o = Dense(n_atom_basis,
n_atom_basis,
bias=False,
activation=None)
self.layer_norm_in = nn.LayerNorm([n_atom_basis
]) ###(input.size()[-1])
def __init__(
self,
nin,
nout,
elements,
n_acsf,
n_apf,
n_acsf_nodes=100,
n_apf_nodes=50,
n_hidden=50,
n_layers=3,
trainable=False,
onehot=True,
activation=shifted_softplus,
):
super(PairGatedNetwork, self).__init__()
self.nelem = len(elements)
self.gate = ElementalGate(elements, trainable=trainable, onehot=onehot)
acsf_input = (n_acsf * len(elements)) + len(elements) + 1
self.dense_radial = Dense(acsf_input,
n_acsf_nodes,
activation=activation)
apf_input = (n_apf * len(elements)) + len(elements) + 1
self.dense_apf = Dense(apf_input, n_apf_nodes, activation=activation)
dense_input = (n_acsf_nodes + n_apf_nodes + len(elements) + 1) * 2 + 2
self.dense_layers = MLP(
dense_input,
nout,
n_hidden=n_hidden,
n_layers=n_layers,
activation=activation,
)
def __init__(self, n_in, n_filters, n_out, filter_network,
cutoff_network=None,
activation=None, normalize_filter=False, axis=2):
super(CFConv, self).__init__()
self.in2f = Dense(n_in, n_filters, bias=False)
self.f2out = Dense(n_filters, n_out, activation=activation)
self.filter_network = filter_network
self.cutoff_network = cutoff_network
self.agg = Aggregate(axis=axis, mean=normalize_filter)
def __init__(self, n_atom_basis=128, max_z=100, kmax=150, n_interactions=1, activation=shifted_softplus):
super(SchnetWithEdgeUpdate, self).__init__()
self.n_interactions = n_interactions
self.embedding = nn.Embedding(max_z, n_atom_basis - 1)
self.edge_update_net = nn.Sequential(
Dense(3 * n_atom_basis, 2 * n_atom_basis, activation=activation),
Dense(2 * n_atom_basis, n_atom_basis),
)
self.msg_edge_net = nn.Sequential(
Dense(n_atom_basis, n_atom_basis, activation=activation),
Dense(n_atom_basis, n_atom_basis, activation=activation),
)
self.msg_atom_fc = Dense(n_atom_basis, n_atom_basis)
self.state_trans_net = nn.Sequential(
Dense(n_atom_basis, n_atom_basis, activation=activation),
Dense(n_atom_basis, n_atom_basis),
)
# self.n_dihedral_edge_attrs = 2
# n_dihedral_edge_feats = 16
# self.dihedral_net = gnn.NNConv(n_atom_basis, n_dihedral_edge_feats, nn.Sequential(
# Dense(self.n_dihedral_edge_attrs, n_atom_basis, activation=F.relu),
# Dense(n_atom_basis, n_atom_basis * n_dihedral_edge_feats),
# ))
# n_angle_edge_attrs = 1
# n_angle_edge_feats = 8
# self.angle_net = gnn.NNConv(n_atom_basis, n_angle_edge_feats, nn.Sequential(
# Dense(n_angle_edge_attrs, n_atom_basis, activation=F.relu),
# Dense(n_atom_basis, n_atom_basis * n_angle_edge_feats),
# ))
# self.init_atom_fc = Dense(n_atom_basis + n_dihedral_edge_feats, n_atom_basis, activation=activation)
self.init_edge_fc = Dense(kmax, n_atom_basis, activation=activation)
def __init__(self,
n_in,
n_filters,
n_out,
filter_network,
cutoff_network=None,
activation=None,
normalize_filter=False,
axis=2,
n_heads_weights=0,
n_heads_conv=0,
device=torch.device("cpu"),
hyperparams=[0, 0],
dropout=0,
exp=False):
super(CFConv, self).__init__()
self.device = device
self.n_heads_weights = n_heads_weights
self.n_heads_conv = n_heads_conv
self.atomic_embedding_dim = n_out
self.in2f = Dense(n_in, n_filters, bias=False, activation=None)
self.f2out = Dense(n_filters, n_out, bias=True, activation=activation)
self.filter_network = filter_network
self.cutoff_network = cutoff_network
#sum over indices
self.agg = Aggregate(axis=axis, mean=normalize_filter)
#added multiheaded attention to weights
self.attention_dim = int(n_out / 4) #arbitrary -> could modify at will
if n_heads_weights > 0:
self.Attention = AttentionHeads(n_in, self.attention_dim,n_heads=self.n_heads_weights,EXP = exp,\
atomic_embedding_dim=n_out ,device=self.device,SM=False,hyperparams = hyperparams,dropout = dropout)
#added multiheaded attention to convolution
if n_heads_conv > 0:
self.AttentionConv = AttentionHeads(n_in,self.attention_dim,n_heads=self.n_heads_conv,EXP=exp,\
atomic_embedding_dim=n_out,device=self.device,SM=False,hyperparams = hyperparams,dropout = dropout)#for now should be single head
#NOTE: the EXP determines if the scalar attention value should be exp(A) or just (A).
#NOTE: exp(A) can be unstable, as can softmax below
#add possibility to use softmax over weights
self.softmax = nn.Softmax(dim=3)
#not currently used, but could add if deemed beneficial
self.dropout = nn.Dropout(dropout)
def __init__(
self,
n_atom_basis,
n_hidden,
activation=None,
dropout_rate=0,
epsilon=0.01,
):
super(AdaptiveComputationTime, self).__init__()
### # Regularization of Atomic Embedding:
self.ponder_net = nn.Sequential(
nn.Dropout(dropout_rate),
Dense(n_atom_basis, n_hidden, activation=activation),
Dense(n_hidden, 1, activation=None),
)
###self.affine_net = Dense(n_atom_basis, n_atom_basis, bias=True, activation=None)
self.epsilon = epsilon
self.sharpen_power = 5. ### to make the linear-interpolation sharper
def __init__(
self,
n_atom_basis,
n_scales,
n_heads=8,
use_time_embedding=True,
):
super(MultiScaleAttention, self).__init__()
# dense layer as mh_attention
assert (n_atom_basis % n_heads == 0), "Mismatch Head Numbers."
assert (n_atom_basis % 2 == 0), "Must Be Even number of Atom Features."
n_per_head = n_atom_basis // n_heads
self.n_per_head = n_per_head
self.n_heads = n_heads
self.n_scales = n_scales
self.n_atom_basis = n_atom_basis
self.use_time_embedding = use_time_embedding
self.mh_q = Dense(n_atom_basis,
n_atom_basis,
bias=False,
activation=None)
self.mh_k = Dense(n_atom_basis,
n_atom_basis,
bias=False,
activation=None)
self.mh_v = Dense(n_atom_basis,
n_atom_basis,
bias=False,
activation=None)
self.mh_o = Dense(n_atom_basis,
n_atom_basis,
bias=False,
activation=None)
def __init__(self,
n_in,
n_filters,
n_out,
filter_network,
cutoff_network=None,
activation=None,
normalize_filter=False,
axis=2,
weight_init=xavier_uniform_):
super(CFConv, self).__init__()
self.in2f = Dense(n_in,
n_filters,
bias=False,
activation=None,
weight_init=weight_init)
self.f2out = Dense(n_filters,
n_out,
bias=True,
activation=activation,
weight_init=weight_init)
self.filter_network = filter_network
self.cutoff_network = cutoff_network
self.agg = Aggregate(axis=axis, mean=normalize_filter)