DeBERTa部分源码理解+注释

文件在我的github里：https://github.com/VICKY-ZZ/DeBERTa

# 参考博客：https://yam.gift/2020/06/27/Paper/2020-06-27-DeBERTa/
class DientangledSelfAttention(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.num_attention_heads = config.num_attention_heads
        self.attention_head_size = int(config.hidden_size/config.num_attention_heads)
        # 为啥不直接等于hidden_size呢？？？
        self.all_head_size = self.num_attention_heads*self.attention_head_size
        # Wq,c,生成Qc，变成三份--QKV
        self.in_proj = torch.nn.Linear(config.hidden_size,self.all_head_size*3, bias = False)
        self.q_bias = torch.nn.Parameter(
            torch.zeros((self.all_head_size),dtype=torch.float)
        )
        self.v_bias = torch.nn.Parameter(
            torch.zeros((self.all_head_size), dtype = torch.float)
        )

        # ----------pos篇
        self.pos_att_type = ['p2c','c2p']
        self.max_relative_positions = config.max_relative_positions
        # pos的dropout for what???
        self.pos_dropout = StableDropout(config.hidden_dropout_prob)
        self.pos_proj = torch.nn.Linear(config.hidden_size,self.all_head_size)
        self.dropout = StableDropout(config.attention_probs_dropout_prob)
    def transpose_for_scores(self,x):
        new_x_shape = x.size()[:-1]+(self.num_attention_heads,-1)
        # 相当于不要x的最后一维（应该是hidden_size），然后换成num_att_head,每个head大小(head_size)
        x = x.view(*new_x_shape)