File size: 13,157 Bytes
ca7299e | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 | """
Adapted from [Openfold](https://github.com/aqlaboratory/openfold) IPA implementation.
"""
import math
from typing import Optional, List, Sequence
import torch
import torch.nn as nn
from src.common.rigid_utils import Rigid
from src.models.net.layers import Linear, NodeTransition, EdgeTransition, TorsionAngleHead, BackboneUpdate
def permute_final_dims(tensor: torch.Tensor, inds: List[int]):
zero_index = -1 * len(inds)
first_inds = list(range(len(tensor.shape[:zero_index])))
return tensor.permute(first_inds + [zero_index + i for i in inds])
def flatten_final_dims(t: torch.Tensor, no_dims: int):
return t.reshape(t.shape[:-no_dims] + (-1,))
def ipa_point_weights_init_(weights):
with torch.no_grad():
softplus_inverse_1 = 0.541324854612918
weights.fill_(softplus_inverse_1)
class InvariantPointAttention(nn.Module):
"""
Implements Algorithm 22.
"""
def __init__(
self,
c_s: int,
c_z: int,
c_hidden: int,
no_heads: int,
no_qk_points: int,
no_v_points: int,
inf: float = 1e5,
eps: float = 1e-8,
):
"""
Args:
c_s:
Single representation channel dimension
c_z:
Pair representation channel dimension
c_hidden:
Hidden channel dimension
no_heads:
Number of attention heads
no_qk_points:
Number of query/key points to generate
no_v_points:
Number of value points to generate
"""
super(InvariantPointAttention, self).__init__()
self.c_s = c_s
self.c_z = c_z
self.c_hidden = c_hidden
self.no_heads = no_heads
self.no_qk_points = no_qk_points
self.no_v_points = no_v_points
self.inf = inf
self.eps = eps
# These linear layers differ from their specifications in the
# supplement. There, they lack bias and use Glorot initialization.
# Here as in the official source, they have bias and use the default
# Lecun initialization.
hc = self.c_hidden * self.no_heads
self.linear_q = Linear(self.c_s, hc)
self.linear_kv = Linear(self.c_s, 2 * hc)
hpq = self.no_heads * self.no_qk_points * 3
self.linear_q_points = Linear(self.c_s, hpq)
hpkv = self.no_heads * (self.no_qk_points + self.no_v_points) * 3
self.linear_kv_points = Linear(self.c_s, hpkv)
self.linear_b = Linear(self.c_z, self.no_heads)
self.down_z = Linear(self.c_z, self.c_z // 4)
self.head_weights = nn.Parameter(torch.zeros((self.no_heads)))
ipa_point_weights_init_(self.head_weights)
concat_out_dim = (
self.c_z // 4 + self.c_hidden + self.no_v_points * 4
)
self.linear_out = Linear(self.no_heads * concat_out_dim, self.c_s, init="final")
self.softmax = nn.Softmax(dim=-1)
self.softplus = nn.Softplus()
def forward(
self,
s: torch.Tensor,
z: Optional[torch.Tensor],
r: Rigid,
mask: torch.Tensor,
_offload_inference: bool = False,
_z_reference_list: Optional[Sequence[torch.Tensor]] = None,
) -> torch.Tensor:
"""
Args:
s:
[*, N_res, C_s] single representation
z:
[*, N_res, N_res, C_z] pair representation
r:
[*, N_res] transformation object
mask:
[*, N_res] mask
Returns:
[*, N_res, C_s] single representation update
"""
if _offload_inference:
z = _z_reference_list
else:
z = [z]
#######################################
# Generate scalar and point activations
#######################################
# [*, N_res, H * C_hidden]
q = self.linear_q(s)
kv = self.linear_kv(s)
# [*, N_res, H, C_hidden]
q = q.view(q.shape[:-1] + (self.no_heads, -1))
# [*, N_res, H, 2 * C_hidden]
kv = kv.view(kv.shape[:-1] + (self.no_heads, -1))
# [*, N_res, H, C_hidden]
k, v = torch.split(kv, self.c_hidden, dim=-1)
# [*, N_res, H * P_q * 3]
q_pts = self.linear_q_points(s)
# This is kind of clunky, but it's how the original does it
# [*, N_res, H * P_q, 3]
q_pts = torch.split(q_pts, q_pts.shape[-1] // 3, dim=-1)
q_pts = torch.stack(q_pts, dim=-1)
q_pts = r[..., None].apply(q_pts)
# [*, N_res, H, P_q, 3]
q_pts = q_pts.view(
q_pts.shape[:-2] + (self.no_heads, self.no_qk_points, 3)
)
# [*, N_res, H * (P_q + P_v) * 3]
kv_pts = self.linear_kv_points(s)
# [*, N_res, H * (P_q + P_v), 3]
kv_pts = torch.split(kv_pts, kv_pts.shape[-1] // 3, dim=-1)
kv_pts = torch.stack(kv_pts, dim=-1)
kv_pts = r[..., None].apply(kv_pts)
# [*, N_res, H, (P_q + P_v), 3]
kv_pts = kv_pts.view(kv_pts.shape[:-2] + (self.no_heads, -1, 3))
# [*, N_res, H, P_q/P_v, 3]
k_pts, v_pts = torch.split(
kv_pts, [self.no_qk_points, self.no_v_points], dim=-2
)
##########################
# Compute attention scores
##########################
# [*, N_res, N_res, H]
b = self.linear_b(z[0])
if(_offload_inference):
z[0] = z[0].cpu()
# [*, H, N_res, N_res]
a = torch.matmul(
permute_final_dims(q, (1, 0, 2)), # [*, H, N_res, C_hidden]
permute_final_dims(k, (1, 2, 0)), # [*, H, C_hidden, N_res]
)
a *= math.sqrt(1.0 / (3 * self.c_hidden))
a += (math.sqrt(1.0 / 3) * permute_final_dims(b, (2, 0, 1)))
# [*, N_res, N_res, H, P_q, 3]
pt_displacement = q_pts.unsqueeze(-4) - k_pts.unsqueeze(-5)
pt_att = pt_displacement ** 2
# [*, N_res, N_res, H, P_q]
pt_att = sum(torch.unbind(pt_att, dim=-1))
head_weights = self.softplus(self.head_weights).view(
*((1,) * len(pt_att.shape[:-2]) + (-1, 1))
)
head_weights = head_weights * math.sqrt(
1.0 / (3 * (self.no_qk_points * 9.0 / 2))
)
pt_att = pt_att * head_weights
# [*, N_res, N_res, H]
pt_att = torch.sum(pt_att, dim=-1) * (-0.5)
# [*, N_res, N_res]
square_mask = mask.unsqueeze(-1) * mask.unsqueeze(-2)
square_mask = self.inf * (square_mask - 1)
# [*, H, N_res, N_res]
pt_att = permute_final_dims(pt_att, (2, 0, 1))
a = a + pt_att
a = a + square_mask.unsqueeze(-3)
a = self.softmax(a)
################
# Compute output
################
# [*, N_res, H, C_hidden]
o = torch.matmul(
a, v.transpose(-2, -3).to(dtype=a.dtype)
).transpose(-2, -3)
# [*, N_res, H * C_hidden]
o = flatten_final_dims(o, 2)
# [*, H, 3, N_res, P_v]
o_pt = torch.sum(
(
a[..., None, :, :, None]
* permute_final_dims(v_pts, (1, 3, 0, 2))[..., None, :, :]
),
dim=-2,
)
# [*, N_res, H, P_v, 3]
o_pt = permute_final_dims(o_pt, (2, 0, 3, 1))
o_pt = r[..., None, None].invert_apply(o_pt)
# [*, N_res, H * P_v]
o_pt_dists = torch.sqrt(torch.sum(o_pt ** 2, dim=-1) + self.eps)
o_pt_norm_feats = flatten_final_dims(
o_pt_dists, 2)
# [*, N_res, H * P_v, 3]
o_pt = o_pt.reshape(*o_pt.shape[:-3], -1, 3)
if(_offload_inference):
z[0] = z[0].to(o_pt.device)
# [*, N_res, H, C_z // 4]
pair_z = self.down_z(z[0]).to(dtype=a.dtype)
o_pair = torch.matmul(a.transpose(-2, -3), pair_z)
# [*, N_res, H * C_z // 4]
o_pair = flatten_final_dims(o_pair, 2)
o_feats = [o, *torch.unbind(o_pt, dim=-1), o_pt_norm_feats, o_pair]
# [*, N_res, C_s]
s = self.linear_out(
torch.cat(
o_feats, dim=-1
).to(dtype=z[0].dtype)
)
return s
class TranslationIPA(nn.Module):
def __init__(self,
c_s: int,
c_z: int,
coordinate_scaling: float,
no_ipa_blocks: int,
skip_embed_size: int,
transformer_num_heads: int = 4,
transformer_num_layers: int = 2,
c_hidden: int = 256,
no_heads: int = 8,
no_qk_points: int = 8,
no_v_points: int = 12,
dropout: float = 0.0,
):
super(TranslationIPA, self).__init__()
self.scale_pos = lambda x: x * coordinate_scaling
self.scale_rigids = lambda x: x.apply_trans_fn(self.scale_pos)
self.unscale_pos = lambda x: x / coordinate_scaling
self.unscale_rigids = lambda x: x.apply_trans_fn(self.unscale_pos)
self.trunk = nn.ModuleDict()
self.num_blocks = no_ipa_blocks
for b in range(no_ipa_blocks):
self.trunk[f'ipa_{b}'] = InvariantPointAttention(
c_s=c_s,
c_z=c_z,
c_hidden=c_hidden,
no_heads=no_heads,
no_qk_points=no_qk_points,
no_v_points=no_v_points,
)
self.trunk[f'ipa_ln_{b}'] = nn.LayerNorm(c_s)
self.trunk[f'skip_embed_{b}'] = Linear(
c_s,
skip_embed_size,
init="final"
)
_in_dim = c_s + skip_embed_size
transformer_layer = nn.TransformerEncoderLayer(
d_model=_in_dim,
nhead=transformer_num_heads,
dim_feedforward=_in_dim,
)
self.trunk[f'transformer_{b}'] = nn.TransformerEncoder(transformer_layer, transformer_num_layers)
self.trunk[f'linear_{b}'] = Linear(_in_dim, c_s, init="final")
self.trunk[f'node_transition_{b}'] = NodeTransition(c_s)
self.trunk[f'bb_update_{b}'] = BackboneUpdate(c_s)
if b < self.num_blocks - 1: # No need edge update on the last block
self.trunk[f'edge_transition_{b}'] = EdgeTransition(
node_embed_size=c_s,
edge_embed_in=c_z,
edge_embed_out=c_z,
)
self.torsion_pred = TorsionAngleHead(c_s, 1)
# self.chi_pred = TorsionAngleHead(c_s, 4)
def forward(self, node_embed, edge_embed, batch):
node_mask = batch['residue_mask'].type(torch.float)
diffuse_mask = (1 - batch['fixed_mask'].type(torch.float)) * node_mask
edge_mask = node_mask[..., None] * node_mask[..., None, :]
init_frames = batch['rigids_t'].type(torch.float)
curr_rigids = Rigid.from_tensor_7(torch.clone(init_frames))
init_rigids = Rigid.from_tensor_7(init_frames)
curr_rigids = self.scale_rigids(curr_rigids)
# Main trunk
init_node_embed = node_embed
for b in range(self.num_blocks):
ipa_embed = self.trunk[f'ipa_{b}'](
node_embed,
edge_embed,
curr_rigids,
node_mask
)
ipa_embed *= node_mask[..., None]
node_embed = self.trunk[f'ipa_ln_{b}'](node_embed + ipa_embed)
# MHA + FF
concat_node_embed = torch.cat([
node_embed, self.trunk[f'skip_embed_{b}'](init_node_embed)
], dim=-1)
concat_node_embed = torch.transpose(concat_node_embed, 0, 1)
transformed_embed = self.trunk[f'transformer_{b}'](concat_node_embed, src_key_padding_mask=1.0 - node_mask)
transformed_embed = torch.transpose(transformed_embed, 0, 1)
# residual
node_embed = node_embed + self.trunk[f'linear_{b}'](transformed_embed)
# MLP
node_embed = self.trunk[f'node_transition_{b}'](node_embed)
# apply mask
node_embed = node_embed * node_mask[..., None]
# backbone update
rigid_update = self.trunk[f'bb_update_{b}'](node_embed * diffuse_mask[..., None])
# apply update
curr_rigids = curr_rigids.compose_q_update_vec(rigid_update, diffuse_mask[..., None])
if b < self.num_blocks - 1:
edge_embed = self.trunk[f'edge_transition_{b}'](node_embed, edge_embed) * edge_mask[..., None]
# torsion prediction
psi_pred = self.torsion_pred(node_embed)
# chi_pred = self.chi_pred(node_embed)
# scale back
curr_rigids = self.unscale_rigids(curr_rigids)
model_out = {
'in_rigids': init_rigids,
'out_rigids': curr_rigids,
'psi': psi_pred,
# 'chi': chi_pred,
}
return model_out
|