| 39 | |
| 40 | # Nature CNN shared trunk + policy and value heads. |
| 41 | class ActorCritic(nn.Module): |
| 42 | def __init__(self, n_actions): |
| 43 | super().__init__() |
| 44 | self.conv = nn.Sequential( |
| 45 | _ortho(nn.Conv2d(4, 32, kernel_size=8, stride=4), 2 ** 0.5), nn.ReLU(), |
| 46 | _ortho(nn.Conv2d(32, 64, kernel_size=4, stride=2), 2 ** 0.5), nn.ReLU(), |
| 47 | _ortho(nn.Conv2d(64, 64, kernel_size=3, stride=1), 2 ** 0.5), nn.ReLU(), |
| 48 | nn.Flatten(), |
| 49 | _ortho(nn.Linear(64 * 7 * 7, 512), 2 ** 0.5), nn.ReLU(), |
| 50 | ) |
| 51 | # gain=0.01 keeps the initial action distribution close to uniform. |
| 52 | self.policy = _ortho(nn.Linear(512, n_actions), 0.01) |
| 53 | self.value = _ortho(nn.Linear(512, 1), 1.0) |
| 54 | |
| 55 | def forward(self, x): |
| 56 | h = self.conv(x.float() / 255.0) |
| 57 | return self.policy(h), self.value(h).squeeze(-1) |
| 58 | |
| 59 | |
| 60 | def compute_gae(rewards, values, dones, last_value): |