use bayes for segementation

[hannah-saber]

Hi, afer reading your paper, I think bayesian can be used for segmentation, then I add bayes into CNN by resampling weight and bias in conv3d, but the result is bad. If it is possible, could you help me to see my code?

yours, Hannah

[hannah-saber]

`class bayesianConv(nn.Module): def init(self, in_channels, out_channels, kernel_size, stride=1, padding=1, dilation=1, bias=True, n=1024, kwargs): super(bayesianConv, self).init(kwargs) self.in_channels = in_channels self.out_channels = out_channels self.kernel_size = kernel_size self.stride = stride self.padding = padding self.dilation = dilation self.use_bias = bias self.n = n

    #初始化bias的mu和rho
    self.b_mu = torch.zeros(self.out_channels)
    self.b_rho = torch.zeros(self.out_channels)
    self.b = torch.zeros(self.out_channels)

    #初始化weight的mu,rho(3*3*3)
    self.w_mu = torch.zeros(size=(self.out_channels, self.in_channels, 3, 3, 3))
    self.w_rho = torch.zeros(size=(self.out_channels, self.in_channels, 3, 3, 3))
    self.w = torch.zeros(size=(self.out_channels, self.in_channels, 3, 3, 3))

def forward(self, input):
    #对bias采样，重参数
    b_epsilon = Normal(0, 1).sample(self.b_mu.shape)
    self.b = self.b_mu + torch.log(1+torch.exp(self.b_rho)) * b_epsilon

    #重新计算bias的均值以及方差，将train次数传进来n
    b_mu_1 = self.b_mu
    self.b_mu = (self.b_mu * (self.n-1) + self.b) / self.n
    self.b_rho = ((self.n-1) * self.b_rho + (self.n-1) * pow(self.b_mu - b_mu_1, 2) + pow(self.b - self.b_mu, 2)) / self.n

    # 将重参数得到的bias放进卷积层
    x = F.conv3d(input=input, weight=self.w, bias=self.b, stride=1, padding=1)
    loss = self.kl_loss()

    return x, loss

def kl_loss(self):
    loss = loss_kl()(self.w, self.w_mu, self.w_rho, self.b, self.b_mu, self.b_rho)
    return loss`