def __init__(self,
hidden_dims,
input_dim,
layer_type="concat",
nonlinearity="softplus"):
super(AmortizedBiasODEnet, self).__init__()
base_layer = {
"ignore": diffeq_layers.IgnoreLinear,
"hyper": diffeq_layers.HyperLinear,
"squash": diffeq_layers.SquashLinear,
"concat": diffeq_layers.ConcatLinear,
"concat_v2": diffeq_layers.ConcatLinear_v2,
"concatsquash": diffeq_layers.ConcatSquashLinear,
"blend": diffeq_layers.BlendLinear,
"concatcoord": diffeq_layers.ConcatLinear,
}[layer_type]
self.input_dim = input_dim
# build layers and add them
layers = []
activation_fns = []
hidden_shape = input_dim
for dim_out in hidden_dims:
layer = base_layer(hidden_shape, dim_out)
layers.append(layer)
activation_fns.append(NONLINEARITIES[nonlinearity])
hidden_shape = dim_out
self.layers = nn.ModuleList(layers)
self.activation_fns = nn.ModuleList(activation_fns[:-1])
def _make_layer(self, block, hidden, planes, num_blocks, stride):
strides = [stride] + [1] * (num_blocks - 1)
layers = []
for stride in strides:
layers.append(block(self.in_planes, hidden, planes, stride))
self.in_planes = planes * block.expansion
return nn.Sequential(*layers)
def create_net_with_time(n_inputs,
n_outputs,
n_layers=1,
n_units=100,
layer_type="concat",
nonlinear="tanh"):
base_layer = LAYER[layer_type]
base_activate = NONLINEARITIES[nonlinear]
layers = [base_layer(n_inputs, n_units)]
for i in range(n_layers):
layers.append(base_activate())
layers.append(base_layer(n_units, n_units))
layers.append(base_activate())
layers.append(nn.Linear(n_units, n_outputs))
return nn.ModuleList(layers)