Eliyar Eziz

> However, in the case where I have a sentence that has 26 different terms (most sentences avg. 5-6 terms), a model trained on only one term will require 5-6...

I think this issue may be related to the numeric embedding feature. Maybe we just add the numeric feature to WordEmbedding rather than embed the numeric feature then add WordEmbed...

这一块之前版本没有考虑，正在构思如何实现。 因为实际上我需要调用两次这个 Generator 第一次遍历构建词表和 Label 表，第二次遍历进行训练。虽然训练是用 fit_generator，但是构建词表目前还不是，导致必须全部加载。 有什么好的办法建议欢迎提出来~

在 kashgari2 分支提交了基于 TF2.0 实现，目前只实现了一个分类模型和 W2V embedding，可以测试看看。

1. 先按照正常流程构建 kashgari 模型，训练 2. kashgari 模型的 tf_model 属性就是对应的 tf 模型，使用这个模型的输入作为输入，目标输出层作为输出构建一个模型。 3. 使用预处理模块把文本转换为向量，然后输入到新构建的模型。

You can just add key value to the hyper params, which will pass to layers as **kwargs

2.0.2 版本应该修复了该问题，可以测试一下～

You can use a custom optimizer by using https://kashgari.readthedocs.io/en/latest/tutorial/text-labeling/#use-custom-optimizer.

根据 [Bert 论文](https://arxiv.org/pdf/1810.04805.pdf) 倒数第二段最后一句话，我们选择了最后四个层的输出合并起来作为后续层的输入。 > For the feature-based approach, we concatenate the last 4 layers of BERT as the features, which was shown to be the best approach in Section...

Wow. This is the first time I see some one used seq2seq on Uighur language generation. What kind tokenization process you have used? And please share the seq2seq definition code,...