fe6cd9a95160fcdda74eec53b17a17800bf00f35,model.py,NoisyLinear,reset_batch_noise,#NoisyLinear#,46

Before Change

    self.bias_epsilon.copy_(self._scale_noise(self.out_features))

  def reset_batch_noise(self):
    for b in range(self.batch_size):
      epsilon_in = self._scale_noise(self.in_features)
      epsilon_out = self._scale_noise(self.out_features)
      self.weight_epsilon_batch[b].copy_(epsilon_out.ger(epsilon_in))
      self.bias_epsilon_batch[b].copy_(self._scale_noise(self.out_features))

  def forward(self, input):
    output = F.linear(input, self.weight_mu, self.bias_mu)
    if self.training:
      if input.size(0) > 1:

After Change

  def reset_batch_noise(self):
    epsilon_in = self._scale_noise(self.batch_size * self.in_features).view(self.batch_size, 1, self.in_features)
    epsilon_out = self._scale_noise(self.batch_size * self.out_features).view(self.batch_size, self.out_features, 1)
    self.weight_epsilon_batch.copy_(torch.bmm(epsilon_out, epsilon_in))
    self.bias_epsilon_batch.copy_(self._scale_noise(self.batch_size * self.out_features).view(self.batch_size, self.out_features))

  def forward(self, input):

Instances