def __init__(self, num_inputs, num_outputs):
super(DiagGaussian, self).__init__()
init_ = lambda m: init(m, init_normc_, lambda x: nn.init.constant_(
x, 0))
self.fc_mean = init_(nn.Linear(num_inputs, num_outputs, bias=True))
# self.fc_mean.weight.data = torch.from_numpy(MIRROR_FACT).float()
self.logstd = AddBias(torch.zeros(num_outputs))
self.synlogstd = AddBias(torch.ones(num_inputs) * 0)
self.opto_probs = AddBias(torch.ones(1) * 0.01)
self.syn_probs = 1.0
def __init__(self, num_inputs, num_outputs):
super(GaussianPolicy, self).__init__()
self.fc_mean = orthongonal_init_(nn.Linear(num_inputs, num_outputs))
self.logstd = AddBias(torch.zeros(num_outputs))
self.saved_log_probs = []
self.rewards = []
def __init__(self, num_inputs, num_outputs):
super(DiagGaussian, self).__init__()
init_ = lambda m: init(m, init_normc_, lambda x: nn.init.constant_(
x, 0))
self.fc_mean = init_(nn.Linear(num_inputs, num_outputs))
self.logstd = AddBias(torch.zeros(num_outputs))
def __init__(self, num_inputs, action_space):
super(MLPPolicy, self).__init__()
self.obs_filter = ObsNorm((1, num_inputs), clip=5)
self.action_space = action_space
self.a_fc1 = nn.Linear(num_inputs, 64, bias=False)
self.a_ab1 = AddBias(64)
self.a_fc2 = nn.Linear(64, 64, bias=False)
self.a_ab2 = AddBias(64)
self.a_fc_mean = nn.Linear(64, action_space.shape[0], bias=False)
self.a_ab_mean = AddBias(action_space.shape[0])
self.a_ab_logstd = AddBias(action_space.shape[0])
self.v_fc1 = nn.Linear(num_inputs, 64, bias=False)
self.v_ab1 = AddBias(64)
self.v_fc2 = nn.Linear(64, 64, bias=False)
self.v_ab2 = AddBias(64)
self.v_fc3 = nn.Linear(64, 1, bias=False)
self.v_ab3 = AddBias(1)
self.apply(weights_init_mlp)
tanh_gain = nn.init.calculate_gain('tanh')
#self.a_fc1.weight.data.mul_(tanh_gain)
#self.a_fc2.weight.data.mul_(tanh_gain)
self.a_fc_mean.weight.data.mul_(0.01)
#self.v_fc1.weight.data.mul_(tanh_gain)
#self.v_fc2.weight.data.mul_(tanh_gain)
self.train()
def __init__(self,
state_shape,
action_dim,
lstm_size,
use_cuda,
use_lstm,
name=''):
super(MLPBase, self).__init__()
init_ = lambda m: init(m, init_normc_, lambda x: nn.init.constant_(
x, 0))
self.state_shape = state_shape
self.action_dim = action_dim
self.lstm_size = lstm_size
self.use_cuda = use_cuda
self.use_lstm = use_lstm
self.stddev = 1
# self.conv1 = nn.Conv2d(self.state_shape[0], 64, kernel_size=9, stride=1, padding = 4)
# # self.bn1 = nn.BatchNorm2d(64)
# self.conv2 = nn.Conv2d(64, 64, kernel_size=7, stride=1, padding = 3)
# # self.bn2 = nn.BatchNorm2d(64)
# self.conv3 = nn.Conv2d(64, 128, kernel_size=5, stride=1, padding = 2)
# # self.bn3 = nn.BatchNorm2d(128)
# self.conv4 = nn.Conv2d(128, 128, kernel_size=5, stride=1, padding = 2)
# # self.bn4 = nn.BatchNorm2d(128)
# self.conv5 = nn.Conv2d(128, 128, kernel_size=3, stride=1, padding = 1)
# # self.bn5 = nn.BatchNorm2d(128)
# self.conv6 = nn.Conv2d(128, 128, kernel_size=3, stride=1, padding = 1)
# self.max4 = nn.MaxPool2d(2)
# # self.bn6 = nn.BatchNorm2d(128)
self.hidden11 = init_(nn.Linear(self.state_shape[3], 512))
self.hidden12 = init_(nn.Linear(self.state_shape[4], 512))
# self.hidden13 = init_(nn.Linear(self.state_shape[0], 256))
# self.hidden14 = init_(nn.Linear(self.state_shape[3], 256))
self.hidden2 = init_(nn.Linear(1024, 512))
if self.use_lstm:
self.lstm = nn.LSTMCell(512, self.lstm_size)
self.action_head = init_(nn.Linear(self.lstm_size,
self.action_dim))
self.action_sigma = init_(
nn.Linear(self.lstm_size, self.action_dim))
self.value_head = init_(nn.Linear(self.lstm_size, 1))
self.lstm.bias_ih.data.fill_(0)
self.lstm.bias_hh.data.fill_(0)
else:
self.action_head = init_(nn.Linear(512, self.action_dim))
self.action_sigma = init_(nn.Linear(512, self.action_dim))
self.value_head = init_(nn.Linear(512, 1))
self.logstd = AddBias(torch.zeros(action_dim))
def __init__(self, num_inputs, action_space):
super(CNNPolicy, self).__init__()
self.conv1 = nn.Conv2d(num_inputs, 32, 8, stride=4, bias=False)
self.ab1 = AddBias(32)
self.conv2 = nn.Conv2d(32, 64, 4, stride=2, bias=False)
self.ab2 = AddBias(64)
self.conv3 = nn.Conv2d(64, 32, 3, stride=1, bias=False)
self.ab3 = AddBias(32)
self.linear1 = nn.Linear(32 * 7 * 7, 512, bias=False)
self.ab_fc1 = AddBias(512)
self.critic_linear = nn.Linear(512, 1, bias=False)
self.ab_fc2 = AddBias(1)
num_outputs = action_space.n
self.actor_linear = nn.Linear(512, num_outputs, bias=False)
self.ab_fc3 = AddBias(num_outputs)
self.apply(weights_init)
relu_gain = nn.init.calculate_gain('relu')
self.conv1.weight.data.mul_(relu_gain)
self.conv2.weight.data.mul_(relu_gain)
self.conv3.weight.data.mul_(relu_gain)
self.linear1.weight.data.mul_(relu_gain)
self.train()
def __init__(self, num_inputs, action_space):
super(MLPPolicy, self).__init__()
self.obs_filter = ObsNorm((1, num_inputs), clip=5)
self.action_space = action_space
self.a_fc1 = nn.Linear(num_inputs, 64, bias=False)
self.a_ab1 = AddBias(64)
self.a_fc2 = nn.Linear(64, 64, bias=False)
self.a_ab2 = AddBias(64)
self.v_fc1 = nn.Linear(num_inputs, 64, bias=False)
self.v_ab1 = AddBias(64)
self.v_fc2 = nn.Linear(64, 64, bias=False)
self.v_ab2 = AddBias(64)
self.v_fc3 = nn.Linear(64, 1, bias=False)
self.v_ab3 = AddBias(1)
if action_space.__class__.__name__ == "Discrete":
num_outputs = action_space.n
self.dist = Categorical(64, num_outputs)
elif action_space.__class__.__name__ == "Box":
num_outputs = action_space.shape[0]
self.dist = DiagGaussian(64, num_outputs)
else:
raise NotImplementedError
self.apply(weights_init_mlp)
tanh_gain = nn.init.calculate_gain('tanh')
#self.a_fc1.weight.data.mul_(tanh_gain)
#self.a_fc2.weight.data.mul_(tanh_gain)
#self.v_fc1.weight.data.mul_(tanh_gain)
#self.v_fc2.weight.data.mul_(tanh_gain)
if action_space.__class__.__name__ == "Box":
self.dist.fc_mean.weight.data.mul_(0.01)
self.train()
def __init__(self, num_inputs, num_outputs, zero_init=False):
super(DiagGaussian, self).__init__()
init_ = lambda m: init(m, init_normc_, lambda x: nn.init.constant_(
x, 0))
self.fc_mean = nn.Linear(num_inputs, num_outputs)
if not zero_init:
self.fc_mean = init_(self.fc_mean)
else:
self.fc_mean.weight.data.fill_(0.0)
self.fc_mean.bias.data.fill_(0.0)
self.logstd = AddBias(torch.zeros(num_outputs))
def __init__(self, num_inputs, num_outputs, num_agents, sigmoid):
super(DiagGaussian, self).__init__()
# An instance (i.e., with specific weight and bias initializing function) of the function "init"
init_ = lambda m: init(m, init_normc_, lambda x: nn.init.constant_(
x, 0))
# Number of inputs and outputs for each agent
self.num_agents = num_agents
self.atom_num_inputs = num_inputs // num_agents
self.atom_num_outputs = num_outputs // num_agents
self.fc_mean = init_(
nn.Linear(self.atom_num_inputs, self.atom_num_outputs))
self.logstd = AddBias(torch.zeros(self.atom_num_outputs))
self.sigmoid = sigmoid
def __init__(self, num_inputs, action_space):
super(CNNPolicy, self).__init__()
self.conv1 = nn.Conv2d(num_inputs, 32, 8, stride=4, bias=False)
self.ab1 = AddBias(32)
self.conv2 = nn.Conv2d(32, 64, 4, stride=2, bias=False)
self.ab2 = AddBias(64)
self.conv3 = nn.Conv2d(64, 32, 3, stride=1, bias=False)
self.ab3 = AddBias(32)
self.linear1 = nn.Linear(32 * 7 * 7, 512, bias=False)
self.ab_fc1 = AddBias(512)
self.critic_linear = nn.Linear(512, 1, bias=False)
self.ab_fc2 = AddBias(1)
if action_space.__class__.__name__ == "Discrete":
num_outputs = action_space.n
self.dist = Categorical(512, num_outputs)
elif action_space.__class__.__name__ == "Box":
num_outputs = action_space.shape[0]
self.dist = DiagGaussian(512, num_outputs)
else:
raise NotImplementedError
self.apply(weights_init)
relu_gain = nn.init.calculate_gain('relu')
self.conv1.weight.data.mul_(relu_gain)
self.conv2.weight.data.mul_(relu_gain)
self.conv3.weight.data.mul_(relu_gain)
self.linear1.weight.data.mul_(relu_gain)
if action_space.__class__.__name__ == "Box":
self.dist.fc_mean.weight.data.mul_(0.01)
self.train()
def __init__(self,
state_shape,
action_dim,
lstm_size,
use_cuda,
use_lstm,
name=''):
super(MLPBase, self).__init__()
init_ = lambda m: init(m, init_normc_, lambda x: nn.init.constant_(
x, 0))
self.state_shape = state_shape
self.action_dim = action_dim
self.lstm_size = lstm_size
self.use_cuda = use_cuda
self.use_lstm = use_lstm
self.stddev = 1
self.hidden1 = init_(
nn.Linear(self.state_shape[1] + self.state_shape[2], 512))
# self.hidden12 = init_(nn.Linear(self.state_shape[4], 512))
# self.hidden13 = init_(nn.Linear(self.state_shape[0], 256))
# self.hidden14 = init_(nn.Linear(self.state_shape[3], 256))
self.hidden2 = init_(nn.Linear(512, 512))
if self.use_lstm:
self.lstm = nn.LSTMCell(512, self.lstm_size)
self.action_head = init_(nn.Linear(self.lstm_size,
self.action_dim))
self.action_sigma = init_(
nn.Linear(self.lstm_size, self.action_dim))
self.value_head = init_(nn.Linear(self.lstm_size, 1))
self.lstm.bias_ih.data.fill_(0)
self.lstm.bias_hh.data.fill_(0)
else:
self.action_head = init_(nn.Linear(512, self.action_dim))
self.action_sigma = init_(nn.Linear(512, self.action_dim))
self.value_head = init_(nn.Linear(512, 1))
self.logstd = AddBias(torch.zeros(action_dim))
def __init__(self, num_inputs, num_outputs):
super(DiagGaussian, self).__init__()
self.fc_mean = nn.Linear(num_inputs, num_outputs)
self.logstd = AddBias(torch.zeros(num_outputs))
def __init__(self, num_inputs, num_outputs): # input: 64, out: 2
super(DiagGaussian, self).__init__()
self.fc_mean = nn.Linear(num_inputs,
num_outputs) # actor의 마지막 layer가 여기 있음
self.logstd = AddBias(torch.zeros(num_outputs))
def __init__(self, module):
super(SplitBias, self).__init__()
self.module = module
self.add_bias = AddBias(module.bias.data)
self.module.bias = None
def __init__(self, num_outputs):
super(DiagGaussianNoFC, self).__init__()
self.logstd = AddBias(torch.zeros(num_outputs))
def __init__(self, num_inputs, num_outputs):
super(DiagGaussian, self).__init__()
self.fc_mean = nn.Linear(64, num_outputs, bias=False)
self.ab_mean = AddBias(num_outputs)
self.ab_logstd = AddBias(num_outputs)
def __init__(self, num_inputs, num_outputs):
super(Categorical, self).__init__()
self.linear = nn.Linear(num_inputs, num_outputs, bias=False)
self.ab = AddBias(num_outputs)
