File size: 6,312 Bytes
827d9ec |
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 |
import importlib
import torch
from torch import Tensor, nn
from . import vb_layers_initialize as init
@torch.compiler.disable
def kernel_triangular_mult(
x,
direction,
mask,
norm_in_weight,
norm_in_bias,
p_in_weight,
g_in_weight,
norm_out_weight,
norm_out_bias,
p_out_weight,
g_out_weight,
eps,
):
triangle_module = importlib.import_module("cuequivariance_torch.primitives.triangle")
triangle_multiplicative_update = triangle_module.triangle_multiplicative_update
return triangle_multiplicative_update(
x,
direction=direction,
mask=mask,
norm_in_weight=norm_in_weight,
norm_in_bias=norm_in_bias,
p_in_weight=p_in_weight,
g_in_weight=g_in_weight,
norm_out_weight=norm_out_weight,
norm_out_bias=norm_out_bias,
p_out_weight=p_out_weight,
g_out_weight=g_out_weight,
eps=eps,
)
class TriangleMultiplicationOutgoing(nn.Module):
"""TriangleMultiplicationOutgoing."""
def __init__(self, dim: int = 128) -> None:
"""Initialize the TriangularUpdate module.
Parameters
----------
dim: int
The dimension of the input, default 128
"""
super().__init__()
self.norm_in = nn.LayerNorm(dim, eps=1e-5)
self.p_in = nn.Linear(dim, 2 * dim, bias=False)
self.g_in = nn.Linear(dim, 2 * dim, bias=False)
self.norm_out = nn.LayerNorm(dim)
self.p_out = nn.Linear(dim, dim, bias=False)
self.g_out = nn.Linear(dim, dim, bias=False)
init.bias_init_one_(self.norm_in.weight)
init.bias_init_zero_(self.norm_in.bias)
init.lecun_normal_init_(self.p_in.weight)
init.gating_init_(self.g_in.weight)
init.bias_init_one_(self.norm_out.weight)
init.bias_init_zero_(self.norm_out.bias)
init.final_init_(self.p_out.weight)
init.gating_init_(self.g_out.weight)
def forward(self, x: Tensor, mask: Tensor, use_kernels: bool = False) -> Tensor:
"""Perform a forward pass.
Parameters
----------
x: torch.Tensor
The input data of shape (B, N, N, D)
mask: torch.Tensor
The input mask of shape (B, N, N)
use_kernels: bool
Whether to use the kernel
Returns
-------
x: torch.Tensor
The output data of shape (B, N, N, D)
"""
if use_kernels:
return kernel_triangular_mult(
x,
direction="outgoing",
mask=mask,
norm_in_weight=self.norm_in.weight,
norm_in_bias=self.norm_in.bias,
p_in_weight=self.p_in.weight,
g_in_weight=self.g_in.weight,
norm_out_weight=self.norm_out.weight,
norm_out_bias=self.norm_out.bias,
p_out_weight=self.p_out.weight,
g_out_weight=self.g_out.weight,
eps=1e-5,
)
# Input gating: D -> D
x = self.norm_in(x)
x_in = x
x = self.p_in(x) * self.g_in(x).sigmoid()
# Apply mask
x = x * mask.unsqueeze(-1)
# Split input and cast to float
a, b = torch.chunk(x.float(), 2, dim=-1)
# Triangular projection
x = torch.einsum("bikd,bjkd->bijd", a, b)
# Output gating
x = self.p_out(self.norm_out(x)) * self.g_out(x_in).sigmoid()
return x
class TriangleMultiplicationIncoming(nn.Module):
"""TriangleMultiplicationIncoming."""
def __init__(self, dim: int = 128) -> None:
"""Initialize the TriangularUpdate module.
Parameters
----------
dim: int
The dimension of the input, default 128
"""
super().__init__()
self.norm_in = nn.LayerNorm(dim, eps=1e-5)
self.p_in = nn.Linear(dim, 2 * dim, bias=False)
self.g_in = nn.Linear(dim, 2 * dim, bias=False)
self.norm_out = nn.LayerNorm(dim)
self.p_out = nn.Linear(dim, dim, bias=False)
self.g_out = nn.Linear(dim, dim, bias=False)
init.bias_init_one_(self.norm_in.weight)
init.bias_init_zero_(self.norm_in.bias)
init.lecun_normal_init_(self.p_in.weight)
init.gating_init_(self.g_in.weight)
init.bias_init_one_(self.norm_out.weight)
init.bias_init_zero_(self.norm_out.bias)
init.final_init_(self.p_out.weight)
init.gating_init_(self.g_out.weight)
def forward(self, x: Tensor, mask: Tensor, use_kernels: bool = False) -> Tensor:
"""Perform a forward pass.
Parameters
----------
x: torch.Tensor
The input data of shape (B, N, N, D)
mask: torch.Tensor
The input mask of shape (B, N, N)
use_kernels: bool
Whether to use the kernel
Returns
-------
x: torch.Tensor
The output data of shape (B, N, N, D)
"""
if use_kernels:
return kernel_triangular_mult(
x,
direction="incoming",
mask=mask,
norm_in_weight=self.norm_in.weight,
norm_in_bias=self.norm_in.bias,
p_in_weight=self.p_in.weight,
g_in_weight=self.g_in.weight,
norm_out_weight=self.norm_out.weight,
norm_out_bias=self.norm_out.bias,
p_out_weight=self.p_out.weight,
g_out_weight=self.g_out.weight,
eps=1e-5,
)
# Input gating: D -> D
x = self.norm_in(x)
x_in = x
x = self.p_in(x) * self.g_in(x).sigmoid()
# Apply mask
x = x * mask.unsqueeze(-1)
# Split input and cast to float
a, b = torch.chunk(x.float(), 2, dim=-1)
# Triangular projection
x = torch.einsum("bkid,bkjd->bijd", a, b)
# Output gating
x = self.p_out(self.norm_out(x)) * self.g_out(x_in).sigmoid()
return x
|