llama-factory源码详解——以DPO为例

由于原文提供的代码已经相对完整，我们可以直接以DPO为例来解释其中的关键部分。

class DPO(nn.Module):
    """Decentralized Q-function for DQN agents.
 
    The DPO module is a feedforward neural network that maps from state and
    action to Q-value.
 
    Attributes:
        state_dim (int): Dimension of states.
        action_dim (int): Dimension of actions.
        hidden_dim (int): Dimension of hidden layers in network.
        num_hidden_layers (int): Number of hidden layers in network.
        use_batch_norm (bool): Whether to use batch normalization or not.
        q_func_type (str): The type of the Q-function.
        activation (torch.nn.Module): Activation function.
    """
 
    def __init__(self, state_dim, action_dim, hidden_dim=256, num_hidden_layers=2,
                 use_batch_norm=True, q_func_type='mean', activation=torch.nn.ReLU):
        super(DPO, self).__init__()
 
        self.state_dim = state_dim
        self.action_dim = action_dim
        self.hidden_dim = hidden_dim
        self.num_hidden_layers = num_hidden_layers
        self.use_batch_norm = use_batch_norm
        self.q_func_type = q_func_type
 
        if self.q_func_type == 'mean':
            self.q_func = MeanQFunction(state_dim=state_dim,
                                       action_dim=action_dim,
                                       hidden_dim=hidden_dim,
                                       num_hidden_layers=num_hidden_layers,
                                       use_batch_norm=use_batch_norm,
                                       activation=activation)
        elif self.q_func_type == 'qr':
            self.q_func = QRQFunction(state_dim=state_dim,
                                     action_dim=action_dim,
                                     hidden_dim=hidden_dim,
                                     num_hidden_layers=num_hidden_layers,
                                     use_batch_norm=use_batch_norm,
                                     activation=activation)
        else:
            raise ValueError('Invalid q_func_type: {}'.format(self.q_func_type))
 
    def forward(sel