EquiFusion / src /core /network /layers /linear.py
ceccurr's picture
EquiFusion demo
e6d357a verified
Raw History Blame Contribute Delete
1.24 kB
from typing import Tuple, Optional, List, Union
import torch
import torch.nn as nn
import torch.nn.functional as F
import math
class MaskedLinear2(nn.Module):
def __init__(self, in_features, out_features, activation=None, bias=False):
super().__init__()
self.activation = activation
self.weight = torch.nn.Parameter(torch.ones(out_features, in_features))
self.bias = bias
nn.init.kaiming_uniform_(self.weight, a=math.sqrt(5))
if self.bias:
self.bias_w = torch.nn.Parameter(torch.zeros(out_features))
fan_in, _ = nn.init._calculate_fan_in_and_fan_out(self.weight)
bound = 1 / math.sqrt(fan_in) if fan_in > 0 else 0
nn.init.uniform_(self.bias_w, -bound, bound)
def forward(self, input, adj: torch.Tensor = None, mask: torch.Tensor = None):
# `input` is of shape (batch_size, n, n)
# `mask` is of shape (batch_size, n,n) as the weigths
w = self.weight
input = input@w.mT
if self.bias:
if mask is None:
mask = (adj.sum(-1) != 0).unsqueeze(-1)
input = (input + self.bias_w)* mask
return input if self.activation is None else self.activation(input)