def forward(self, x):
mean = self.out_bias(x, self.weight)
sigma = torch.exp(self.log_alpha) * self.weight * self.weight
std = torch.sqrt(1e-16 + self.out_nobias(x * x, sigma))
# if self.training:
#############################
# epsilon = std.data.new(std.size()).normal_()
#############################
# means = torch.zeros(std.size())
# epsilon = torch.normal(mean=means,std=1.0)
# print("std.size:", std.size())
epsilon = torch.randn(self.std_size)
# else:
# epsilon = 0.0
# Local reparameterization trick
out = mean + std * epsilon
if cfg.record_mean_var and cfg.record_now and self.training and self.name in cfg.record_layers:
utils.save_array_to_file(mean.cpu().detach().numpy(), self.mean_var_path, "mean")
utils.save_array_to_file(std.cpu().detach().numpy(), self.mean_var_path, "std")
return out
def forward(self, x):
mean = F.linear(x, self.W)
if self.bias is not None:
mean = mean + self.bias
sigma = torch.exp(self.log_alpha) * self.W * self.W
std = torch.sqrt(1e-16 + F.linear(x * x, sigma))
if self.training:
epsilon = std.data.new(std.size()).normal_()
else:
epsilon = 0.0
# Local reparameterization trick
out = mean + std * epsilon
if cfg.record_mean_var and cfg.record_now and self.training and self.name in cfg.record_layers:
utils.save_array_to_file(mean.cpu().detach().numpy(), self.mean_var_path, "mean")
utils.save_array_to_file(std.cpu().detach().numpy(), self.mean_var_path, "std")
return out