def __init__(self, num_inputs, num_outputs, use_state_dependent_std):
super(GaussianActorLayer, self).__init__()
self.actor_mean = nn.Linear(num_inputs, num_outputs)
init(self.actor_mean, nn.init.orthogonal_, lambda x: nn.init.constant_(x, 0))
self.use_state_dependent_std = use_state_dependent_std
if self.use_state_dependent_std:
self.actor_logstd = nn.Linear(num_inputs, num_outputs)
init(self.actor_logstd, nn.init.orthogonal_, lambda x: nn.init.constant_(x, 0))
else:
self.logstd = nn.Parameter(torch.zeros(num_outputs), requires_grad=True)
def __init__(self, num_inputs, num_outputs, use_state_dependent_std, init_w=1e-3):
super(TanhGaussainActorLayer, self).__init__()
self.mean_actor = nn.Linear(num_inputs, num_outputs)
init(self.mean_actor, lambda x: nn.init.uniform_(x, -init_w, init_w),
lambda x: nn.init.uniform_(x, -init_w, init_w))
self.state_dependent_std = use_state_dependent_std
if self.state_dependent_std:
self.logstd_actor = nn.Linear(num_inputs, num_outputs)
init(self.logstd_actor, lambda x: nn.init.uniform_(x, -init_w, init_w),
lambda x: nn.init.uniform_(x, -init_w, init_w))
else:
self.logstd = nn.Parameter(torch.zeros(num_outputs), requires_grad=True)
self.logstd_actor = lambda _: self.logstd
def init_(m):
init(m, truncated_norm_init, lambda x: nn.init.constant_(x, 0))
def init_last_(m):
init(m, lambda x: nn.init.uniform_(x, -init_w, init_w),
lambda x: nn.init.uniform_(x, -init_w, init_w))
def init_(m):
init(m, fanin_init, lambda x: nn.init.constant_(x, init_b))
def __init__(self, num_inputs, num_outputs):
super(BernoulliActorLayer, self).__init__()
self.logit_actor = nn.Linear(num_inputs, num_outputs)
init(self.logit_actor, nn.init.orthogonal_, lambda x: nn.init. constant_(x, 0))
def __init__(self, num_inputs, num_outputs):
super(CategoricalActorLayer, self).__init__()
self.logit_actor = nn.Linear(num_inputs, num_outputs)
init(self.logit_actor, lambda x: nn.init.orthogonal_(x, 0.01), lambda x: nn.init.constant_(x, 0))