84 lines
3.2 KiB
Python
84 lines
3.2 KiB
Python
|
import sys
|
||
|
sys.path.append("..")
|
||
|
|
||
|
import math
|
||
|
import torch
|
||
|
import torch.nn as nn
|
||
|
import torch.nn.functional as F
|
||
|
from torch.nn import Parameter
|
||
|
|
||
|
from metrics import calculate_kl as KL_DIV
|
||
|
from ..misc import ModuleWrapper
|
||
|
|
||
|
|
||
|
class BBBConv2d(ModuleWrapper):
|
||
|
def __init__(self, in_channels, out_channels, kernel_size,
|
||
|
stride=1, padding=0, dilation=1, bias=True, priors=None):
|
||
|
|
||
|
super(BBBConv2d, self).__init__()
|
||
|
self.in_channels = in_channels
|
||
|
self.out_channels = out_channels
|
||
|
self.kernel_size = kernel_size if isinstance(kernel_size, tuple) else (kernel_size, kernel_size)
|
||
|
self.stride = stride
|
||
|
self.padding = padding
|
||
|
self.dilation = dilation
|
||
|
self.groups = 1
|
||
|
self.use_bias = bias
|
||
|
self.device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
|
||
|
|
||
|
if priors is None:
|
||
|
priors = {
|
||
|
'prior_mu': 0,
|
||
|
'prior_sigma': 0.1,
|
||
|
'posterior_mu_initial': (0, 0.1),
|
||
|
'posterior_rho_initial': (-3, 0.1),
|
||
|
}
|
||
|
self.prior_mu = priors['prior_mu']
|
||
|
self.prior_sigma = priors['prior_sigma']
|
||
|
self.posterior_mu_initial = priors['posterior_mu_initial']
|
||
|
self.posterior_rho_initial = priors['posterior_rho_initial']
|
||
|
|
||
|
self.W_mu = Parameter(torch.empty((out_channels, in_channels, *self.kernel_size), device=self.device))
|
||
|
self.W_rho = Parameter(torch.empty((out_channels, in_channels, *self.kernel_size), device=self.device))
|
||
|
|
||
|
if self.use_bias:
|
||
|
self.bias_mu = Parameter(torch.empty((out_channels), device=self.device))
|
||
|
self.bias_rho = Parameter(torch.empty((out_channels), device=self.device))
|
||
|
else:
|
||
|
self.register_parameter('bias_mu', None)
|
||
|
self.register_parameter('bias_rho', None)
|
||
|
|
||
|
self.reset_parameters()
|
||
|
|
||
|
def reset_parameters(self):
|
||
|
self.W_mu.data.normal_(*self.posterior_mu_initial)
|
||
|
self.W_rho.data.normal_(*self.posterior_rho_initial)
|
||
|
|
||
|
if self.use_bias:
|
||
|
self.bias_mu.data.normal_(*self.posterior_mu_initial)
|
||
|
self.bias_rho.data.normal_(*self.posterior_rho_initial)
|
||
|
|
||
|
def forward(self, input, sample=True):
|
||
|
if self.training or sample:
|
||
|
W_eps = torch.empty(self.W_mu.size()).normal_(0, 1).to(self.device)
|
||
|
self.W_sigma = torch.log1p(torch.exp(self.W_rho))
|
||
|
weight = self.W_mu + W_eps * self.W_sigma
|
||
|
|
||
|
if self.use_bias:
|
||
|
bias_eps = torch.empty(self.bias_mu.size()).normal_(0, 1).to(self.device)
|
||
|
self.bias_sigma = torch.log1p(torch.exp(self.bias_rho))
|
||
|
bias = self.bias_mu + bias_eps * self.bias_sigma
|
||
|
else:
|
||
|
bias = None
|
||
|
else:
|
||
|
weight = self.W_mu
|
||
|
bias = self.bias_mu if self.use_bias else None
|
||
|
|
||
|
return F.conv2d(input, weight, bias, self.stride, self.padding, self.dilation, self.groups)
|
||
|
|
||
|
def kl_loss(self):
|
||
|
kl = KL_DIV(self.prior_mu, self.prior_sigma, self.W_mu, self.W_sigma)
|
||
|
if self.use_bias:
|
||
|
kl += KL_DIV(self.prior_mu, self.prior_sigma, self.bias_mu, self.bias_sigma)
|
||
|
return kl
|