MCPcopy Create free account
hub / github.com/pytorch/examples / Policy

Class Policy

reinforcement_learning/actor_critic.py:36–70  ·  view source on GitHub ↗

implements both actor and critic in one model

Source from the content-addressed store, hash-verified

34
35
36class Policy(nn.Module):
37 """
38 implements both actor and critic in one model
39 """
40 def __init__(self):
41 super(Policy, self).__init__()
42 self.affine1 = nn.Linear(4, 128)
43
44 # actor's layer
45 self.action_head = nn.Linear(128, 2)
46
47 # critic's layer
48 self.value_head = nn.Linear(128, 1)
49
50 # action & reward buffer
51 self.saved_actions = []
52 self.rewards = []
53
54 def forward(self, x):
55 """
56 forward of both actor and critic
57 """
58 x = F.relu(self.affine1(x))
59
60 # actor: choses action to take from state s_t
61 # by returning probability of each action
62 action_prob = F.softmax(self.action_head(x), dim=-1)
63
64 # critic: evaluates being in the state s_t
65 state_values = self.value_head(x)
66
67 # return values for both actor and critic as a tuple of 2 values:
68 # 1. a list with the probability of each action over the action space
69 # 2. the value from state s_t
70 return action_prob, state_values
71
72
73model = Policy()

Callers 1

actor_critic.pyFile · 0.70

Calls

no outgoing calls

Tested by

no test coverage detected