5e148038be5971f2c7c811d46a1d7b28c759ecda,layers/common_layers.py,GravesAttention,forward,#GravesAttention#Any#Any#Any#Any#,144

Before Change

                                      g_t.size(1),
                                      inputs.size(1))
        sig_t = sig_t.unsqueeze(2).expand_as(g_t)
        mu_t_ = mu_t.unsqueeze(2).expand_as(g_t)
        j = self.J[:g_t.size(0), :, :inputs.size(1)]

        // attention weights
        phi_t = g_t * torch.exp(-0.5 * (mu_t_ - j)**2 / (sig_t**2))

After Change

        j = self.J[:inputs.size(1)+1]

        // attention weights
        phi_t = g_t.unsqueeze(-1) * torch.exp(-0.5 * (mu_t.unsqueeze(-1) - j)**2 / (sig_t.unsqueeze(-1)**2))

        // discritize attention weights
        alpha_t = self.COEF * torch.sum(phi_t, 1)

Instances