f98bd2ec9d4289939ff6661d5a9c43ee7e8996f7,models/shared_rnn.py,RNN,forward,#RNN#,195

Before Change

                norm = norm.unsqueeze(-1)
                detached_norm = torch.autograd.Variable(norm.data,
                                                        requires_grad=False)
                hidden[hidden_norms > max_norm] *= max_norm/detached_norm

            logits.append(logit)
            h1tohT.append(hidden)

After Change

                // This workaround for PyTorch v0.3.1 does everything in numpy,
                // because the PyTorch slicing and slice assignment is too
                // flaky.
                hidden_norms = hidden_norms.data.cpu().numpy()

                clipped_num += 1
                if hidden_norms.max() > max_clipped_norm:
                    max_clipped_norm = hidden_norms.max()

                clip_select = hidden_norms > max_norm
                clip_norms = hidden_norms[clip_select]

                mask = np.ones(hidden.size())
                normalizer = max_norm/clip_norms
                normalizer = normalizer[:, np.newaxis]

Instances