siat-nlp/ddmn

Code and data for the paper "Dual Dynamic Memory Network for End-to-End Multi-turn Task-oriented Dialog Systems".

ddmn's Issues

bug question & reproduce

1. 请问 source/module/attention.py line 114 是否存在错误？final_memory = temp_memory + add_memory (add_memory 这里只是权重) 并没有加上更新量:add
2. 同时模型在给定的参数下 无法复现出在 MultiWoZ的结果
   复现: | 论文中：
   F1: 26.01 | 31.4
   BLEU: 13.97 | 12.4
   请问是 有另外的超参数配置吗？

Query about metrics computation

def update(self, metrics_list):
    for i, metrics in enumerate(metrics_list):
        num_samples = metrics.pop("num_samples", 1)
        self.num_samples += num_samples
        for key, val in metrics.items():
            if val is not None:
                key_turn = str(key) + "-turn-{}".format(str(i+1))
                self.metrics_val[key_turn] = val

                if isinstance(val, torch.Tensor):
                    val = val.item()
                    self.metrics_cum[key] += val * num_samples
                elif isinstance(val, tuple):
                    assert len(val) == 2
                    val, num_words = val[0].item(), val[1]
                    self.metrics_cum[key] += np.array([val * num_samples, num_words])
                else:
                    self.metrics_cum[key] += val * num_samples

Wanted to know the reason for scaling the metrics with the number of samples in the batch, and how can one compute the perplexity for the test set

source/model/seq2seq.py : 386 不应该从属于gpu判断分支下面

你好，seq2seq.py: 386 是否不应该从属于gpu判断分支下面？
当不使用gpu时，selector_target和pad_zeros就不会放在一起，会报错
而且判断逻辑上也不太适合根据是否使用gpu来判断是否pad吧