def __init__(self, num_channels, output_dim, initializer='xavier'):
super(DiscreteConvQNetwork, self).__init__()
self.base = create_dqn_base(num_channels, initializer=initializer)
self.V_stream = create_linear_network(7 * 7 * 64,
1,
hidden_units=[512],
initializer=initializer)
self.A_stream = create_linear_network(7 * 7 * 64,
output_dim,
hidden_units=[512],
initializer=initializer)
def __init__(self, num_inputs, num_actions, hidden_units=[256, 256],
initializer='xavier'):
super(GaussianPolicy, self).__init__()
# https://github.com/ku2482/rltorch/blob/master/rltorch/network/builder.py
self.policy = create_linear_network(
num_inputs, num_actions*2, hidden_units=hidden_units,
initializer=initializer)
def __init__(self, num_channels, output_dim, initializer='kaiming'):
super(ConvCategoricalPolicy, self).__init__()
self.policy = nn.Sequential(
*create_dqn_base(num_channels),
*create_linear_network(7 * 7 * 64,
output_dim,
hidden_units=[512],
output_activation='softmax',
initializer=initializer))
def __init__(self,
input_dim,
output_dim,
hidden_units=[],
initializer='xavier'):
super(LinearGaussianPolicy, self).__init__()
self.policy = create_linear_network(input_dim,
output_dim * 2,
hidden_units=hidden_units,
initializer=initializer)
def __init__(self,
input_dim,
output_dim,
hidden_units=[],
initializer='xavier'):
super(ContinuousLinearQNetwork, self).__init__()
self.Q = create_linear_network(input_dim + output_dim,
1,
hidden_units=hidden_units,
initializer=initializer)