def build(self):
"""Build."""
self.fc1 = nn.Linear(self.observation_space.shape[0], 256)
self.fc2 = nn.Linear(256, 256)
self.dist_torque = DiagGaussian(256,
self.action_space['torque'].shape[0],
constant_log_std=False)
self.dist_position = DiagGaussian(256,
self.action_space['position'].shape[0],
constant_log_std=False)
for p in self.dist_torque.fc_mean.parameters():
nn.init.constant_(p, 0.)
for p in self.dist_position.fc_mean.parameters():
nn.init.constant_(p, 0.)
def __init__(self, ob_shape, action_shape, nunits):
"""Init."""
super().__init__()
self.nunits = nunits
self.fc1 = nn.Linear(ob_shape, self.nunits)
self.fc2 = nn.Linear(self.nunits, self.nunits)
self.dist = DiagGaussian(self.nunits, action_shape)
self.ob_shape = ob_shape
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."""
inshape = self.observation_space.shape[0]
self.fc1 = nn.Linear(inshape, self.nunits)
self.fc2 = nn.Linear(self.nunits, self.nunits)
self.dist = DiagGaussian(self.nunits, self.action_space.shape[0])
for p in self.dist.fc_mean.parameters():
nn.init.constant_(p, 0.)
self.vf_fc1 = nn.Linear(inshape, self.nunits)
self.vf_fc2 = nn.Linear(self.nunits, self.nunits)
self.vf_out = nn.Linear(self.nunits, 1)
def build(self):
"""Build."""
inshape = (self.observation_space.spaces[0].shape[0] +
self.observation_space.spaces[1].shape[0])
self.fc1 = nn.Linear(inshape, 32)
self.fc2 = nn.Linear(32, 32)
self.fc3 = nn.Linear(32, 32)
self.dist = DiagGaussian(32, self.action_space.shape[0])
for p in self.dist.fc_mean.parameters():
nn.init.constant_(p, 0.)
self.vf_fc1 = nn.Linear(inshape, 32)
self.vf_fc2 = nn.Linear(32, 32)
self.vf_fc3 = nn.Linear(32, 32)
self.vf_out = nn.Linear(32, 1)
def build(self):
"""Build."""
self.fc1 = nn.Linear(self.observation_space.shape[0], 128)
self.fc2 = nn.Linear(128, 128)
self.dist = DiagGaussian(128, self.action_space.shape[0])