triangular_multiplicative_update.py 3.76 KB
Newer Older
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
1
2
# Copyright 2021 AlQuraishi Laboratory
# Copyright 2021 DeepMind Technologies Limited
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
3
#
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
4
5
6
7
8
9
10
11
12
13
14
15
16
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#      http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

from functools import partialmethod
17
18
from typing import Optional

Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
19
20
21
import torch
import torch.nn as nn

Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
22
23
from openfold.model.primitives import Linear
from openfold.utils.tensor_utils import permute_final_dims
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
24
25
26
27


class TriangleMultiplicativeUpdate(nn.Module):
    """
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
28
    Implements Algorithms 11 and 12.
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
29
30
31
    """
    def __init__(self, c_z, c_hidden, _outgoing=True):
        """
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
32
33
34
35
36
37
        Args:
            c_z:
                Input channel dimension
            c:
                Hidden channel dimension
        """
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
        super(TriangleMultiplicativeUpdate, self).__init__()
        self.c_z = c_z
        self.c_hidden = c_hidden
        self._outgoing = _outgoing

        self.linear_a_p = Linear(self.c_z, self.c_hidden)
        self.linear_a_g = Linear(self.c_z, self.c_hidden, init="gating")
        self.linear_b_p = Linear(self.c_z, self.c_hidden)
        self.linear_b_g = Linear(self.c_z, self.c_hidden, init="gating")
        self.linear_g = Linear(self.c_z, self.c_z, init="gating")
        self.linear_z = Linear(self.c_hidden, self.c_z, init="final")

        self.layer_norm_in = nn.LayerNorm(self.c_z)
        self.layer_norm_out = nn.LayerNorm(self.c_hidden)

        self.sigmoid = nn.Sigmoid()

55
56
57
58
59
    def _combine_projections(
        a: torch.Tensor,
        b: torch.Tensor,
    ) -> torch.Tensor:
        raise NotImplementedError("This method needs to be overridden")
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
60

61
62
63
64
    def forward(self, 
        z: torch.Tensor, 
        mask: Optional[torch.Tensor] = None
    ) -> torch.Tensor:
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
65
        """
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
66
67
68
69
70
71
72
        Args:
            x:
                [*, N_res, N_res, C_z] input tensor
            mask:
                [*, N_res, N_res] input mask
        Returns:
            [*, N_res, N_res, C_z] output tensor
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
73
        """
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
74
        if mask is None:
75
            mask = z.new_ones(z.shape[:-1])
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
76
77
78
79
80
81
82
83

        mask = mask.unsqueeze(-1)

        z = self.layer_norm_in(z)
        a = self.linear_a_p(z) * self.sigmoid(self.linear_a_g(z))
        a = a * mask
        b = self.linear_b_p(z) * self.sigmoid(self.linear_b_g(z))
        b = b * mask
84
        x = self._combine_projections(a, b)
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
85
86
87
88
89
90
91
92
93
94
        x = self.layer_norm_out(x)
        x = self.linear_z(x)
        g = self.sigmoid(self.linear_g(z))
        z = x * g

        return z


class TriangleMultiplicationOutgoing(TriangleMultiplicativeUpdate):
    """
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
95
    Implements Algorithm 11.
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
96
    """
97
98
99
100
101
102
103
104
105
106
    def _combine_projections(
        self,
        a: torch.Tensor,  # [*, N_i, N_k, C]
        b: torch.Tensor,  # [*, N_j, N_k, C]
    ):
        # [*, C, N_i, N_j]
        p = torch.matmul(
            permute_final_dims(a, (2, 0, 1)),
            permute_final_dims(b, (2, 1, 0)),
        )
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
107

108
109
        # [*, N_i, N_j, C]
        return permute_final_dims(p, (1, 2, 0))
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
110

Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
111
112
113

class TriangleMultiplicationIncoming(TriangleMultiplicativeUpdate):
    """
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
114
    Implements Algorithm 12.
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
115
    """
116
117
118
119
120
121
122
123
124
125
126
127
128
    def _combine_projections(
        self,
        a: torch.Tensor,  # [*, N_k, N_i, C]
        b: torch.Tensor,  # [*, N_k, N_j, C]
    ):
        # [*, C, N_i, N_j]
        p = torch.matmul(
            permute_final_dims(a, (2, 1, 0)),
            permute_final_dims(b, (2, 0, 1)),
        )

        # [*, N_i, N_j, C]
        return permute_final_dims(p, (1, 2, 0))
Gustaf Ahdritz's avatar
Gustaf Ahdritz committed
129