AttentionFreeTransformer 源码解析（一）：AFTFull、AFTSimple、AFTLocal-编程知识

AttentionFreeTransformer 源码解析（一）：AFTFull、AFTSimple、AFTLocal

我觉得源码写的很好懂，我就不加注释了，直接上计算流程图。

`AFTFull`

在这里插入图片描述

class AFTFull(nn.Module):def __init__(self, max_seqlen, dim, hidden_dim=64):super().__init__()'''max_seqlen: the maximum number of timesteps (sequence length) to be fed indim: the embedding dimension of the tokenshidden_dim: the hidden dimension used inside AFT FullNumber of heads is 1 as done in the paper'''self.dim = dimself.hidden_dim = hidden_dimself.to_q = nn.Linear(dim, hidden_dim)self.to_k = nn.Linear(dim, hidden_dim)self.to_v = nn.Linear(dim, hidden_dim)self.project = nn.Linear(hidden_dim, dim)self.wbias = nn.Parameter(torch.Tensor(max_seqlen, max_seqlen))nn.init.xavier_uniform_(self.wbias)def forward(self, x):B, T, _ = x.shapeQ = self.to_q(x).view(B, T, self.hidden_dim)K = self.to_k(x).view(B, T, self.hidden_dim)V = self.to_v(x).view(B, T, self.hidden_dim)temp_wbias = self.wbias[:T, :T].unsqueeze(0) # sequences can still be variable length'''From the paper'''Q_sig = torch.sigmoid(Q)temp = torch.exp(temp_wbias) @ torch.mul(torch.exp(K), V)weighted = temp / (torch.exp(temp_wbias) @ torch.exp(K))Yt = torch.mul(Q_sig, weighted)Yt = Yt.view(B, T, self.hidden_dim)Yt = self.project(Yt)return Yt

`AFTSimple`

在这里插入图片描述

class AFTSimple(nn.Module):def __init__(self, max_seqlen, dim, hidden_dim=64):super().__init__()'''max_seqlen: the maximum number of timesteps (sequence length) to be fed indim: the embedding dimension of the tokenshidden_dim: the hidden dimension used inside AFT FullNumber of Heads is 1 as done in the paper.'''self.dim = dimself.hidden_dim = hidden_dimself.to_q = nn.Linear(dim, hidden_dim)self.to_k = nn.Linear(dim, hidden_dim)self.to_v = nn.Linear(dim, hidden_dim)self.project = nn.Linear(hidden_dim, dim)def forward(self, x):B, T, _ = x.shapeQ = self.to_q(x).view(B, T, self.hidden_dim)K = self.to_k(x).view(B, T, self.hidden_dim)V = self.to_v(x).view(B, T, self.hidden_dim)'''From the paper'''weights = torch.mul(torch.softmax(K, 1), V).sum(dim=1, keepdim=True)Q_sig = torch.sigmoid(Q)Yt = torch.mul(Q_sig, weights)Yt = Yt.view(B, T, self.hidden_dim)Yt = self.project(Yt)return Yt

`AFTLocal`

在这里插入图片描述

class AFTLocal(nn.Module):def __init__(self, max_seqlen, dim, hidden_dim=64, s=256):super().__init__()'''max_seqlen: the maximum number of timesteps (sequence length) to be fed indim: the embedding dimension of the tokenshidden_dim: the hidden dimension used inside AFT Fulls: the window size used for AFT-Local in the paperNumber of heads is 1 as done in the paper'''self.dim = dimself.hidden_dim = hidden_dimself.to_q = nn.Linear(dim, hidden_dim)self.to_k = nn.Linear(dim, hidden_dim)self.to_v = nn.Linear(dim, hidden_dim)self.project = nn.Linear(hidden_dim, dim)self.wbias = nn.Parameter(torch.Tensor(max_seqlen, max_seqlen))self.max_seqlen = max_seqlenself.s = snn.init.xavier_uniform_(self.wbias)def forward(self, x):B, T, _ = x.shapeQ = self.to_q(x).view(B, T, self.hidden_dim)K = self.to_k(x).view(B, T, self.hidden_dim)V = self.to_v(x).view(B, T, self.hidden_dim)self.wbias = nn.Parameter(torch.Tensor([[self.wbias[i][j] if math.fabs(i-j) < self.s else 0 for j in range(self.max_seqlen)] for i in range(self.max_seqlen)]))temp_wbias = self.wbias[:T, :T].unsqueeze(0) # sequences can still be variable length'''From the paper'''Q_sig = torch.sigmoid(Q)temp = torch.exp(temp_wbias) @ torch.mul(torch.exp(K), V)weighted = temp / (torch.exp(temp_wbias) @ torch.exp(K))Yt = torch.mul(Q_sig, weighted)Yt = Yt.view(B, T, self.hidden_dim)Yt = self.project(Yt)return Yt