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

Before Change

        g_t = torch.softmax(g_t, dim=-1) / sig_t + self.eps

        // each B x K x T_in
        g_t = g_t.unsqueeze(2).expand(g_t.size(0),
                                      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

After Change

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

        // apply masking
        if mask is not None:
            alpha_t.data.masked_fill_(~mask, self._mask_value)

Instances