def __init__(self, args, dim_in, dim_out, cond_dim):
super(OdeLayer, self).__init__()
self.layer_cond = args.layer_cond
self.set_param_shapes(args, dim_in, dim_out, cond_dim)
if args.layer_cond == 'hyper':
total_shape = sum([np.prod(s) for s in self.list_param_shapes])
self.hyper_net = HyperNet(cond_dim, args.hyper_dims + '-' + str(total_shape), args.nonlinearity)
else:
self.create_param_from_shapes(self.list_param_shapes)
weights_init(self)
def __init__(self,
x_dim,
y_dim,
hidden_dims,
nonlinearity,
trainable=False):
input_dim = max(x_dim, y_dim) * 2
super(SizeBalancedEmbedNet, self).__init__(input_dim, hidden_dims,
nonlinearity, trainable)
self.x_dim = x_dim
self.y_dim = y_dim
if x_dim > y_dim:
self.mm = nn.Linear(y_dim, x_dim)
else:
self.mm = nn.Linear(x_dim, y_dim)
weights_init(self)
def __init__(self,
prior_embed_dim,
ob_dim,
hidden_dims,
nonlinearity,
time_length=1.0):
super(TimePredNet, self).__init__()
self.time_length = time_length
self.ob2embed = nn.Linear(ob_dim, prior_embed_dim)
self.st_pred = MLP(input_dim=prior_embed_dim * 2,
hidden_dims=hidden_dims,
nonlinearity=nonlinearity,
act_last='softplus')
self.dur_pred = MLP(input_dim=prior_embed_dim * 2,
hidden_dims=hidden_dims,
nonlinearity=nonlinearity,
act_last='sigmoid')
weights_init(self)
def __init__(self, input_dim, hidden_dims, nonlinearity, act_last=None):
super(MLP, self).__init__()
self.act_last = act_last
hidden_dims = tuple(map(int, hidden_dims.split("-")))
prev_size = input_dim
layers = []
activation_fns = []
for h in hidden_dims:
layers.append(nn.Linear(prev_size, h))
prev_size = h
activation_fns.append(NONLINEARITIES[nonlinearity])
if act_last is not None:
activation_fns[-1] = NONLINEARITIES[self.act_last]
self.output_size = prev_size
self.layers = nn.ModuleList(layers)
self.activation_fns = nn.ModuleList(activation_fns)
weights_init(self)
def __init__(self, input_dim, hidden_dims, nonlinearity):
super(HyperNet, self).__init__()
self.mlp = MLP(input_dim, hidden_dims, nonlinearity, act_last=nonlinearity)
weights_init(self)
def __init__(self, dim, num):
super(LogisticReg, self).__init__()
self.dim = dim
self.w = Parameter(torch.Tensor(num, dim))
weights_init(self)