def __init__(self, state_size, action_size, fc1_units=256, fc2_units=128):
super(Actor, self).__init__()
self.net = FCNet(state_size,
action_size, [fc1_units, fc2_units],
random_seed=RANDOM_SEED,
output_func=torch.tanh)
self.to(DEVICE)
def __init__(self, state_size, action_size, shared_feature_size, fc_units):
super(CategoricalActorCritic, self).__init__()
self.shared_net = FCNet(input_size=state_size,
output_size=shared_feature_size,
hidden_sizes=fc_units,
random_seed=RANDOM_SEED)
self.actor = Softmax(feature_size=shared_feature_size,
action_size=action_size)
self.critic = nn.Linear(shared_feature_size, 1)
self.to(DEVICE)
class Critic(nn.Module):
def __init__(self, state_size, fc1_units=256, fc2_units=128):
super(Critic, self).__init__()
self.net = FCNet(state_size,
1, [fc1_units, fc2_units],
random_seed=RANDOM_SEED)
self.to(DEVICE)
def reset_parameters(self):
self.net.reset_parameters()
def forward(self, state_action):
return self.net.forward(state_action)
class Actor(nn.Module):
def __init__(self, state_size, action_size, fc1_units=256, fc2_units=128):
super(Actor, self).__init__()
self.net = FCNet(state_size,
action_size, [fc1_units, fc2_units],
random_seed=RANDOM_SEED,
output_func=torch.tanh)
self.to(DEVICE)
def reset_parameters(self):
self.net.reset_parameters()
def forward(self, state):
return self.net.forward(state)
class Actor(nn.Module):
def __init__(self, state_size, action_size, fc1_units=64, fc2_units=32):
super(Actor, self).__init__()
self.net = FCNet(state_size,
action_size, [fc1_units, fc2_units],
random_seed=RANDOM_SEED)
self.to(DEVICE)
def reset_parameters(self):
self.net.reset_parameters()
def forward(self, state):
logits = self.net.forward(state)
dist = torch.distributions.Categorical(logits=logits)
action = dist.sample() # action.shape = (N)
log_prob = dist.log_prob(action) # log_prob.shape = (N)
entropy = dist.entropy().sum(-1).unsqueeze(-1)
return entropy, action.unsqueeze(-1), log_prob.unsqueeze(-1)
class CategoricalActorCritic(nn.Module):
def __init__(self, state_size, action_size, shared_feature_size, fc_units):
super(CategoricalActorCritic, self).__init__()
self.shared_net = FCNet(input_size=state_size,
output_size=shared_feature_size,
hidden_sizes=fc_units,
random_seed=RANDOM_SEED)
self.actor = Softmax(feature_size=shared_feature_size,
action_size=action_size)
self.critic = nn.Linear(shared_feature_size, 1)
self.to(DEVICE)
def forward(self, states):
features = self.shared_net.forward(states)
entropy, actions, log_prob = self.actor.forward(features=features)
critic_v = self.critic.forward(features)
return entropy, actions, log_prob, critic_v
def __init__(self, state_size, fc1_units=256, fc2_units=128):
super(Critic, self).__init__()
self.net = FCNet(state_size,
1, [fc1_units, fc2_units],
random_seed=RANDOM_SEED)
self.to(DEVICE)
def __init__(self, state_size, action_size, fc1_units=64, fc2_units=32):
super(Actor, self).__init__()
self.net = FCNet(state_size,
action_size, [fc1_units, fc2_units],
random_seed=RANDOM_SEED)
self.to(DEVICE)
def __init__(self, state_size, action_size, fc1_units=256, fc2_units=128):
super(NavModel, self).__init__()
self.net = FCNet(state_size,
action_size, [fc1_units, fc2_units],
random_seed=RANDOM_SEED)
self.to(DEVICE)