Some problem about the implement of Cascaded Model

[wizare]

I want to use your Cascaded Model. In your usage description, there are 5 steps in total. The first two steps I successfully run， but fail in third steps.

In the script "hard/train_critic.sh", it requires the folder "../douban_pretrain/checkpoint_epoch19.pkl.clean", while it's not generated before.

Can you tell me how to solve this problem?

[wizare]

In nmt/utils/Model.py, 230-238 rows: ref_contexts, ref_mask = [], [] for template_input, template_length in zip(template_inputs, template_lengths): emb_ref = self.dec_embedding(template_input) ref_context, _ = self.encoder_ref(emb_ref, template_length) ref_mask_ = sequence_mask(template_length) ref_contexts.append(ref_context) ref_mask.append(ref_mask_) ref_contexts = torch.cat(ref_contexts, 0) ref_mask = torch.cat(ref_mask, 1)

It encodes template_inputs separately, however the encode and mask operation is designed for a batch data. I think this code should be replaced as: emb_ref = self.dec_embedding(template_inputs) ref_contexts, _ = self.encoder_ref(emb_ref, template_lengths) ref_mask = sequence_mask(template_lengths)

[chijames]

Hi, May I know how do you do the preprocessing? Moreover, where did you get the stop word file? Thanks!