File size: 6,203 Bytes
3f0529e |
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 |
# Copyright (c) Meta Platforms, Inc. and affiliates.
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.
import torch
from openfold.model.triangular_attention import (
TriangleAttentionEndingNode,
TriangleAttentionStartingNode,
)
from openfold.model.triangular_multiplicative_update import (
TriangleMultiplicationIncoming,
TriangleMultiplicationOutgoing,
)
from torch import nn
from esm.esmfold.v1.misc import (
Attention,
Dropout,
PairToSequence,
ResidueMLP,
SequenceToPair,
)
class TriangularSelfAttentionBlock(nn.Module):
def __init__(
self,
sequence_state_dim,
pairwise_state_dim,
sequence_head_width,
pairwise_head_width,
dropout=0,
**__kwargs,
):
super().__init__()
assert sequence_state_dim % sequence_head_width == 0
assert pairwise_state_dim % pairwise_head_width == 0
sequence_num_heads = sequence_state_dim // sequence_head_width
pairwise_num_heads = pairwise_state_dim // pairwise_head_width
assert sequence_state_dim == sequence_num_heads * sequence_head_width
assert pairwise_state_dim == pairwise_num_heads * pairwise_head_width
assert pairwise_state_dim % 2 == 0
self.sequence_state_dim = sequence_state_dim
self.pairwise_state_dim = pairwise_state_dim
self.layernorm_1 = nn.LayerNorm(sequence_state_dim)
self.sequence_to_pair = SequenceToPair(
sequence_state_dim, pairwise_state_dim // 2, pairwise_state_dim
)
self.pair_to_sequence = PairToSequence(pairwise_state_dim, sequence_num_heads)
self.seq_attention = Attention(
sequence_state_dim, sequence_num_heads, sequence_head_width, gated=True
)
self.tri_mul_out = TriangleMultiplicationOutgoing(
pairwise_state_dim,
pairwise_state_dim,
)
self.tri_mul_in = TriangleMultiplicationIncoming(
pairwise_state_dim,
pairwise_state_dim,
)
self.tri_att_start = TriangleAttentionStartingNode(
pairwise_state_dim,
pairwise_head_width,
pairwise_num_heads,
inf=1e9,
) # type: ignore
self.tri_att_end = TriangleAttentionEndingNode(
pairwise_state_dim,
pairwise_head_width,
pairwise_num_heads,
inf=1e9,
) # type: ignore
self.mlp_seq = ResidueMLP(sequence_state_dim, 4 * sequence_state_dim, dropout=dropout)
self.mlp_pair = ResidueMLP(pairwise_state_dim, 4 * pairwise_state_dim, dropout=dropout)
assert dropout < 0.4
self.drop = nn.Dropout(dropout)
self.row_drop = Dropout(dropout * 2, 2)
self.col_drop = Dropout(dropout * 2, 1)
torch.nn.init.zeros_(self.tri_mul_in.linear_z.weight)
torch.nn.init.zeros_(self.tri_mul_in.linear_z.bias)
torch.nn.init.zeros_(self.tri_mul_out.linear_z.weight)
torch.nn.init.zeros_(self.tri_mul_out.linear_z.bias)
torch.nn.init.zeros_(self.tri_att_start.mha.linear_o.weight)
torch.nn.init.zeros_(self.tri_att_start.mha.linear_o.bias)
torch.nn.init.zeros_(self.tri_att_end.mha.linear_o.weight)
torch.nn.init.zeros_(self.tri_att_end.mha.linear_o.bias)
torch.nn.init.zeros_(self.sequence_to_pair.o_proj.weight)
torch.nn.init.zeros_(self.sequence_to_pair.o_proj.bias)
torch.nn.init.zeros_(self.pair_to_sequence.linear.weight)
torch.nn.init.zeros_(self.seq_attention.o_proj.weight)
torch.nn.init.zeros_(self.seq_attention.o_proj.bias)
torch.nn.init.zeros_(self.mlp_seq.mlp[-2].weight)
torch.nn.init.zeros_(self.mlp_seq.mlp[-2].bias)
torch.nn.init.zeros_(self.mlp_pair.mlp[-2].weight)
torch.nn.init.zeros_(self.mlp_pair.mlp[-2].bias)
def forward(self, sequence_state, pairwise_state, mask=None, chunk_size=None, **__kwargs):
"""
Inputs:
sequence_state: B x L x sequence_state_dim
pairwise_state: B x L x L x pairwise_state_dim
mask: B x L boolean tensor of valid positions
Output:
sequence_state: B x L x sequence_state_dim
pairwise_state: B x L x L x pairwise_state_dim
"""
assert len(sequence_state.shape) == 3
assert len(pairwise_state.shape) == 4
if mask is not None:
assert len(mask.shape) == 2
batch_dim, seq_dim, sequence_state_dim = sequence_state.shape
pairwise_state_dim = pairwise_state.shape[3]
assert sequence_state_dim == self.sequence_state_dim
assert pairwise_state_dim == self.pairwise_state_dim
assert batch_dim == pairwise_state.shape[0]
assert seq_dim == pairwise_state.shape[1]
assert seq_dim == pairwise_state.shape[2]
# Update sequence state
bias = self.pair_to_sequence(pairwise_state)
# Self attention with bias + mlp.
y = self.layernorm_1(sequence_state)
y, _ = self.seq_attention(y, mask=mask, bias=bias)
sequence_state = sequence_state + self.drop(y)
sequence_state = self.mlp_seq(sequence_state)
# Update pairwise state
pairwise_state = pairwise_state + self.sequence_to_pair(sequence_state)
# Axial attention with triangular bias.
tri_mask = mask.unsqueeze(2) * mask.unsqueeze(1) if mask is not None else None
pairwise_state = pairwise_state + self.row_drop(
self.tri_mul_out(pairwise_state, mask=tri_mask)
)
pairwise_state = pairwise_state + self.col_drop(
self.tri_mul_in(pairwise_state, mask=tri_mask)
)
pairwise_state = pairwise_state + self.row_drop(
self.tri_att_start(pairwise_state, mask=tri_mask, chunk_size=chunk_size)
)
pairwise_state = pairwise_state + self.col_drop(
self.tri_att_end(pairwise_state, mask=tri_mask, chunk_size=chunk_size)
)
# MLP over pairs.
pairwise_state = self.mlp_pair(pairwise_state)
return sequence_state, pairwise_state
|