未验证 提交 2af08acc 编辑于 作者: dlagul's avatar dlagul 提交者： GitHub
浏览文件

Update tester.py

上级 edd2330d

sdfvae/tester.py

+6 −6
原始行号 差异行号 差异行
@@ -84,9 +84,9 @@ class Tester(object): 
        # See https://arxiv.org/pdf/1606.05908.pdf, Page 9, Section 2.2 for details.

        # The constant items in the loss function (not the coefficients) can be any number here, or even omitted        

        # due to they have no any impact on gradientis propagation during training. So do in testing.

        # log(N(x|mu,var))

        # = log{1/(sqrt(2*pi)*var)exp{-(x-mu)^2/(2*var^2)}} 

        # = -0.5*{log(2*pi)+2*log(var)+[(x-mu)/exp{log(var)}]^2}

        # log(N(x|mu,sigma^2))

        # = log{1/(sqrt(2*pi)*sigma)exp{-(x-mu)^2/(2*sigma^2)}} 

        # = -0.5*{log(2*pi)+2*log(sigma)+[(x-mu)/exp{log(sigma)}]^2}

        loglikelihood = -0.5 * torch.sum(torch.pow(((original_seq.float()-recon_seq_mu.float())/torch.exp(recon_seq_logvar.float())), 2) 

                                         + 2 * recon_seq_logvar.float() 

                                         + np.log(np.pi*2))
@@ -106,9 +106,9 @@ class Tester(object): 
        # The constant items in the loss function (not the coefficients) can be any number here, or even omitted        

        # due to they have no any impact on gradientis propagation during training. 

        # So do in testing, because they also have no any impact on the results (all anomaly scores increase or decrease to the same extent).

        # log(N(x|mu,var))

        # = log{1/(sqrt(2*pi)*var)exp{-(x-mu)^2/(2*var^2)}} 

        # = -0.5*{log(2*pi)+2*log(var)+[(x-mu)/exp{log(var)}]^2}

        # log(N(x|mu,sigma^2))

        # = log{1/(sqrt(2*pi)*sigma)exp{-(x-mu)^2/(2*sigma^2)}} 

        # = -0.5*{log(2*pi)+2*log(sigma)+[(x-mu)/exp{log(sigma)}]^2}

        llh = -0.5 * torch.sum(torch.pow(((x.float()-recon_x_mu.float())/torch.exp(recon_x_logvar.float())), 2) 

              + 2 * recon_x_logvar.float() 

              + np.log(np.pi*2))