def __init__(self):
super(Policy, self).__init__()
self.num_envs = num_envs
# shared layer
self.affine1 = nn.Linear(4, 128)
self.affine_pi0 = nn.Linear(4, 128)
# not shared layers
self.mu_heads = nn.ModuleList(
[nn.Linear(128, 1) for i in range(self.num_envs + 1)])
self.sigma2_heads = nn.ModuleList(
[nn.Linear(128, 1) for i in range(self.num_envs + 1)])
self.value_heads = nn.ModuleList(
[nn.Linear(128, 1) for i in range(self.num_envs)])
self.apply(weights_init)
# +1 for the distilled policy
for i in range(self.num_envs + 1):
mu = self.mu_heads[i]
sigma = self.sigma2_heads[i]
mu.data = normalized_columns_initializer(mu.weight.data, 0.01)
mu.bias.data.fill_(0)
sigma.bias.data.fill_(0)
if i != self.num_envs:
value = self.value_heads[i]
value.weight.data = normalized_columns_initializer(
value.weight.data, 1.0)
value.bias.data.fill_(0)
# initialize lists for holding run information
self.div = [[] for i in range(self.num_envs)]
self.saved_actions = [[] for i in range(self.num_envs)]
#self.entropies = [[] for i in range(num_envs)]
self.entropies = [[] for i in range(self.num_envs)]
self.rewards = [[] for i in range(self.num_envs)]
self.log_prob = [[] for i in range(self.num_envs)]
def __init__(self):
super(Policy, self).__init__()
self.num_envs = num_envs
# shared layer
self.affine1 = nn.Linear(4, 128)
# not shared layers
self.mu_heads = nn.ModuleList(
[nn.Linear(128, 1) for i in range(self.num_envs)])
self.sigma2_heads = nn.ModuleList(
[nn.Linear(128, 1) for i in range(self.num_envs)])
self.apply(weights_init)
for i in range(self.num_envs):
mu = self.mu_heads[i]
sigma = self.sigma2_heads[i]
mu.data = normalized_columns_initializer(mu.weight.data, 0.01)
mu.bias.data.fill_(0)
sigma.bias.data.fill_(0)
# initialize lists for holding run information
self.div = [[] for i in range(num_envs)]
