Method __init__

config/atari/model.py:367–535

EfficientZero network Parameters ---------- observation_shape: tuple or list shape of observations: [C, W, H] action_space_size: int action space num_blocks: int number of res blocks num_channels: int cha

(
        self,
        observation_shape,
        action_space_size,
        num_blocks,
        num_channels,
        reduced_channels_reward,
        reduced_channels_value,
        reduced_channels_policy,
        fc_reward_layers,
        fc_value_layers,
        fc_policy_layers,
        reward_support_size,
        value_support_size,
        downsample,
        inverse_value_transform,
        inverse_reward_transform,
        lstm_hidden_size,
        bn_mt=0.1,
        proj_hid=256,
        proj_out=256,
        pred_hid=64,
        pred_out=256,
        init_zero=False,
        state_norm=False
    )

class EfficientZeroNet(BaseNet):
    def __init__(
        self,
        observation_shape,
        action_space_size,
        num_blocks,
        num_channels,
        reduced_channels_reward,
        reduced_channels_value,
        reduced_channels_policy,
        fc_reward_layers,
        fc_value_layers,
        fc_policy_layers,
        reward_support_size,
        value_support_size,
        downsample,
        inverse_value_transform,
        inverse_reward_transform,
        lstm_hidden_size,
        bn_mt=0.1,
        proj_hid=256,
        proj_out=256,
        pred_hid=64,
        pred_out=256,
        init_zero=False,
        state_norm=False
    ):
        """EfficientZero network
        Parameters
        ----------
        observation_shape: tuple or list
            shape of observations: [C, W, H]
        action_space_size: int
            action space
        num_blocks: int
            number of res blocks
        num_channels: int
            channels of hidden states
        reduced_channels_reward: int
            channels of reward head
        reduced_channels_value: int
            channels of value head
        reduced_channels_policy: int
            channels of policy head
        fc_reward_layers: list
            hidden layers of the reward prediction head (MLP head)
        fc_value_layers: list
            hidden layers of the value prediction head (MLP head)
        fc_policy_layers: list
            hidden layers of the policy prediction head (MLP head)
        reward_support_size: int
            dim of reward output
        value_support_size: int
            dim of value output
        downsample: bool
            True -> do downsampling for observations. (For board games, do not need)
        inverse_value_transform: Any
            A function that maps value supports into value scalars
        inverse_reward_transform: Any