def __init__(self):
"""Build network."""
super().__init__(None, None)
self.fc1 = nn.Linear(9, 32)
self.fc2 = nn.Linear(32, 32)
self.vf = nn.Linear(32, 1)
self.dist = Categorical(32, 9)
def build(self):
"""Build network."""
inshape = self.observation_space.shape[0]
self.net = FeedForwardNet(inshape, [32, 32], activate_last=True)
if hasattr(self.action_space, 'n'):
self.dist = Categorical(32, self.action_space.n)
else:
self.dist = DiagGaussian(32, self.action_space.shape[0])
self.vf = torch.nn.Linear(32, 1)
def build(self):
"""Build."""
self.fc1 = nn.Linear(self.observation_space.shape[0], 32)
self.fc2 = nn.Linear(32, 32)
self.fc3 = nn.Linear(32, 32)
if isinstance(self.action_space, gym.spaces.Box):
self.dist = TanhDiagGaussian(32, self.action_space.shape[0])
else:
self.dist = Categorical(32, self.action_space.n)
def build(self):
"""Build network."""
self.conv1 = nn.Conv2d(4, 32, 8, 4)
self.conv2 = nn.Conv2d(32, 64, 4, 2)
self.conv3 = nn.Conv2d(64, 64, 3, 1)
shape = self.observation_space.shape[1:]
for c in [self.conv1, self.conv2, self.conv3]:
shape = conv_out_shape(shape, c)
self.nunits = 64 * np.prod(shape)
self.fc = nn.Linear(self.nunits, 512)
self.dist = Categorical(512, self.action_space.n)
def build(self):
"""Build."""
self.conv1 = nn.Conv2d(4, 16, 8, 4)
self.conv2 = nn.Conv2d(16, 32, 4, 2)
shape = self.observation_space.shape[1:]
for c in [self.conv1, self.conv2]:
shape = conv_out_shape(shape, c)
self.nunits = 32 * np.prod(shape)
self.fc = nn.Linear(self.nunits, 256)
self.vf = nn.Linear(256, 1)
self.dist = Categorical(256, self.action_space.n)
nn.init.orthogonal_(self.vf.weight.data, gain=1.0)
nn.init.constant_(self.vf.bias.data, 0)
def build(self):
"""Build."""
self.fc1 = nn.Linear(self.observation_space.shape[0], 128)
self.fc2 = nn.Linear(128, 128)
self.dist = Categorical(128, self.action_space.n)
def __init__(self, action_space):
super().__init__()
self.fc1 = nn.Linear(64, 32)
self.dist = Categorical(32, action_space.n)
def build(self):
"""Build Network."""
self.fc1 = nn.Linear(self.observation_space.shape[0], 32)
self.fc2 = nn.Linear(32, 32)
self.fc3 = nn.Linear(32, 32)
self.dist = Categorical(32, self.action_space.n)
