| 280 | return out |
| 281 | |
| 282 | class MSAColGlobalAttention(nn.Module): |
| 283 | def __init__(self, d_msa=64, n_head=8, d_hidden=8): |
| 284 | super(MSAColGlobalAttention, self).__init__() |
| 285 | self.norm_msa = nn.LayerNorm(d_msa) |
| 286 | # |
| 287 | self.to_q = nn.Linear(d_msa, n_head*d_hidden, bias=False) |
| 288 | self.to_k = nn.Linear(d_msa, d_hidden, bias=False) |
| 289 | self.to_v = nn.Linear(d_msa, d_hidden, bias=False) |
| 290 | self.to_g = nn.Linear(d_msa, n_head*d_hidden) |
| 291 | self.to_out = nn.Linear(n_head*d_hidden, d_msa) |
| 292 | |
| 293 | self.scaling = 1/math.sqrt(d_hidden) |
| 294 | self.h = n_head |
| 295 | self.dim = d_hidden |
| 296 | |
| 297 | self.reset_parameter() |
| 298 | |
| 299 | def reset_parameter(self): |
| 300 | # query/key/value projection: Glorot uniform / Xavier uniform |
| 301 | nn.init.xavier_uniform_(self.to_q.weight) |
| 302 | nn.init.xavier_uniform_(self.to_k.weight) |
| 303 | nn.init.xavier_uniform_(self.to_v.weight) |
| 304 | |
| 305 | # gating: zero weights, one biases (mostly open gate at the begining) |
| 306 | nn.init.zeros_(self.to_g.weight) |
| 307 | nn.init.ones_(self.to_g.bias) |
| 308 | |
| 309 | # to_out: right before residual connection: zero initialize -- to make it sure residual operation is same to the Identity at the begining |
| 310 | nn.init.zeros_(self.to_out.weight) |
| 311 | nn.init.zeros_(self.to_out.bias) |
| 312 | |
| 313 | def forward(self, msa): |
| 314 | B, N, L = msa.shape[:3] |
| 315 | # |
| 316 | msa = self.norm_msa(msa) |
| 317 | # |
| 318 | query = self.to_q(msa).reshape(B, N, L, self.h, self.dim) |
| 319 | query = query.mean(dim=1) # (B, L, h, dim) |
| 320 | key = self.to_k(msa) # (B, N, L, dim) |
| 321 | value = self.to_v(msa) # (B, N, L, dim) |
| 322 | gate = torch.sigmoid(self.to_g(msa)) # (B, N, L, h*dim) |
| 323 | # |
| 324 | query = query * self.scaling |
| 325 | attn = einsum('bihd,bkid->bihk', query, key) # (B, L, h, N) |
| 326 | attn = F.softmax(attn, dim=-1) |
| 327 | # |
| 328 | out = einsum('bihk,bkid->bihd', attn, value).reshape(B, 1, L, -1) # (B, 1, L, h*dim) |
| 329 | out = gate * out # (B, N, L, h*dim) |
| 330 | # |
| 331 | out = self.to_out(out) |
| 332 | return out |
| 333 | |
| 334 | # Instead of triangle attention, use Tied axail attention with bias from coordinates..? |
| 335 | class BiasedAxialAttention(nn.Module): |